ZHANGYUXUAN-zR commited on
Commit
53120f0
·
verified ·
1 Parent(s): f2d80e8

Add files using upload-large-folder tool

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. parse/train/4YlE2huxEsl/4YlE2huxEsl.md +411 -0
  2. parse/train/4YlE2huxEsl/4YlE2huxEsl_content_list.json +1650 -0
  3. parse/train/4YlE2huxEsl/4YlE2huxEsl_middle.json +0 -0
  4. parse/train/4YlE2huxEsl/4YlE2huxEsl_model.json +0 -0
  5. parse/train/8qDwejCuCN/8qDwejCuCN.md +314 -0
  6. parse/train/8qDwejCuCN/8qDwejCuCN_content_list.json +1735 -0
  7. parse/train/8qDwejCuCN/8qDwejCuCN_middle.json +0 -0
  8. parse/train/8qDwejCuCN/8qDwejCuCN_model.json +0 -0
  9. parse/train/B1gd0nEFwS/B1gd0nEFwS.md +433 -0
  10. parse/train/B1gd0nEFwS/B1gd0nEFwS_content_list.json +0 -0
  11. parse/train/B1gd0nEFwS/B1gd0nEFwS_middle.json +0 -0
  12. parse/train/B1gd0nEFwS/B1gd0nEFwS_model.json +0 -0
  13. parse/train/BJYwwY9ll/BJYwwY9ll.md +263 -0
  14. parse/train/BJYwwY9ll/BJYwwY9ll_content_list.json +1446 -0
  15. parse/train/BJYwwY9ll/BJYwwY9ll_middle.json +0 -0
  16. parse/train/BJYwwY9ll/BJYwwY9ll_model.json +0 -0
  17. parse/train/BJgr4kSFDS/BJgr4kSFDS.md +385 -0
  18. parse/train/BJgr4kSFDS/BJgr4kSFDS_content_list.json +0 -0
  19. parse/train/BJgr4kSFDS/BJgr4kSFDS_middle.json +0 -0
  20. parse/train/BJgr4kSFDS/BJgr4kSFDS_model.json +0 -0
  21. parse/train/By5ugjyCb/By5ugjyCb.md +354 -0
  22. parse/train/By5ugjyCb/By5ugjyCb_content_list.json +0 -0
  23. parse/train/By5ugjyCb/By5ugjyCb_middle.json +0 -0
  24. parse/train/By5ugjyCb/By5ugjyCb_model.json +0 -0
  25. parse/train/BylJUTEKvB/BylJUTEKvB.md +324 -0
  26. parse/train/BylJUTEKvB/BylJUTEKvB_content_list.json +1630 -0
  27. parse/train/BylJUTEKvB/BylJUTEKvB_middle.json +0 -0
  28. parse/train/BylJUTEKvB/BylJUTEKvB_model.json +0 -0
  29. parse/train/C0GmZH2RnVR/C0GmZH2RnVR.md +254 -0
  30. parse/train/C0GmZH2RnVR/C0GmZH2RnVR_content_list.json +1007 -0
  31. parse/train/C0GmZH2RnVR/C0GmZH2RnVR_middle.json +0 -0
  32. parse/train/C0GmZH2RnVR/C0GmZH2RnVR_model.json +0 -0
  33. parse/train/DfGu8WwT0d/DfGu8WwT0d.md +368 -0
  34. parse/train/DfGu8WwT0d/DfGu8WwT0d_content_list.json +0 -0
  35. parse/train/DfGu8WwT0d/DfGu8WwT0d_middle.json +0 -0
  36. parse/train/DfGu8WwT0d/DfGu8WwT0d_model.json +0 -0
  37. parse/train/Hk4_qw5xe/Hk4_qw5xe.md +505 -0
  38. parse/train/Hk4_qw5xe/Hk4_qw5xe_content_list.json +0 -0
  39. parse/train/Hk4_qw5xe/Hk4_qw5xe_middle.json +0 -0
  40. parse/train/Hk4_qw5xe/Hk4_qw5xe_model.json +0 -0
  41. parse/train/Hk9Xc_lR-/Hk9Xc_lR-.md +0 -0
  42. parse/train/Hk9Xc_lR-/Hk9Xc_lR-_content_list.json +0 -0
  43. parse/train/Hk9Xc_lR-/Hk9Xc_lR-_middle.json +0 -0
  44. parse/train/Hk9Xc_lR-/Hk9Xc_lR-_model.json +0 -0
  45. parse/train/HkeGhoA5FX/HkeGhoA5FX.md +352 -0
  46. parse/train/HkeGhoA5FX/HkeGhoA5FX_content_list.json +0 -0
  47. parse/train/HkeGhoA5FX/HkeGhoA5FX_middle.json +0 -0
  48. parse/train/HkeGhoA5FX/HkeGhoA5FX_model.json +0 -0
  49. parse/train/QYjO70ACDK/QYjO70ACDK_content_list.json +1515 -0
  50. parse/train/QYjO70ACDK/QYjO70ACDK_middle.json +0 -0
parse/train/4YlE2huxEsl/4YlE2huxEsl.md ADDED
@@ -0,0 +1,411 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Beltrami Flow and Neural Diffusion on Graphs
2
+
3
+ James Rowbottom∗ Twitter Inc.
4
+
5
+ Benjamin P. Chamberlain∗ Twitter Inc. bchamberlain@twitter.com
6
+
7
+ Davide Eynard Twitter Inc.
8
+
9
+ rancesco Di Giovann Twitter Inc.
10
+
11
+ Xiaowen Dong University of Oxford
12
+
13
+ Michael M. Bronstein Twitter Inc. and Imperial College London
14
+
15
+ # Abstract
16
+
17
+ We propose a novel class of graph neural networks based on the discretised Beltrami flow, a non-Euclidean diffusion PDE. In our model, node features are supplemented with positional encodings derived from the graph topology and jointly evolved by the Beltrami flow, producing simultaneously continuous feature learning and topology evolution. The resulting model generalises many popular graph neural networks and achieves state-of-the-art results on several benchmarks.
18
+
19
+ # 1 Introduction
20
+
21
+ The majority of graph neural networks (GNNs) are based on the message passing paradigm [30], wherein node features are learned by means of a non-linear propagation on the graph. Multiple recent works have pointed to the limitations of the message passing approach. These include; limited expressive power [80, 95, 9, 7], the related oversmoothing problem [60, 62] and the bottleneck phenomena [1, 93], which render such approaches inefficient, especially in deep GNNs. Multiple alternatives have been proposed, among which are higher-order methods [54, 7] and decoupling the propagation and input graphs by modifying the topology, often referred to as graph rewiring. Topological modifications can take different forms such as graph sampling [32], kNN [43], using the complete graph [86, 1], latent graph learning [89, 36], or multi-hop filters [92, 73]. However, there is no agreement in the literature on when and how to modify the graph, and a single principled framework for doing so.
22
+
23
+ A somewhat underappreciated fact is that GNNs are intimately related to diffusion equations [16], a connection that was exploited in the early work of Scarselli et al. [76]. Diffusion PDEs have been historically important in computer graphics [83, 11, 51, 64], computer vision [13, 18, 6], and image processing [65, 82, 90, 85, 26, 12], where they created an entire trend of variational and PDE-based approaches. In machine learning and data science, diffusion equations underpin such popular manifold learning methods as eigenmaps [5] and diffusion maps [20], as well as the family of PageRank algorithms [63, 14]. In deep learning, differential equations are used as models of neural networks [16, 19, 25, 94, 71, 98] and for physics-informed learning [72, 22, 75, 21, 81, 47].
24
+
25
+ Main contributions In this paper, we propose a novel class of GNNs based on the discretised non-Euclidean diffusion PDE in joint positional and feature space, inspired by the Beltrami flow [82] used two decades ago in the image processing literature for edge-preserving image denoising. We show that the discretisation of the spatial component of the Beltrami flow offers a principled view on positional encoding and graph rewiring, whereas the discretisation of the temporal component can replace GNN layers with more flexible adaptive numerical schemes. Based on this model, we introduce Beltrami Neural Diffusion (BLEND) that generalises a broad range of GNN architectures and shows state-of-the-art performance on many popular benchmarks. In a broader perspective, our approach explores new tools from PDEs and differential geometry that are less well known in the graph ML community.
26
+
27
+ # 2 Background
28
+
29
+ Beltrami flow Kimmel et al. [40, 82, 39] considered images as 2-manifolds (parametric surfaces) $( \Sigma , g )$ embedded in some larger ambient space as $\mathbf { z } ( \mathbf { u } ) \overset { } { = } ( \mathbf { u } , \alpha \mathbf { x } ( \mathbf { u } ) ) \ \subseteq \ \mathbb { R } ^ { { \hat { d } } + 2 }$ where $\alpha \geq 0$ is a scaling factor, $\mathbf { u } = ( u _ { 1 } , u _ { 2 } )$ are the 2D positional coordinates of the pixels, and $\mathbf { x }$ are the $d$ - dimensional colour or feature coordinates (with $d = 1$ or 3 for grayscale or RGB images, or $d = k ^ { 2 }$ when using $k \times k$ patches as features [12]). In these works, the image is evolved along the gradient flow of a functional $S [ \mathbf { z } , g ]$ called the Polyakov action $I 6 8 J$ , which roughly measures the smoothness of the embedding1. For images embedded in Euclidean space with the functional $S$ minimised with respect to both the embedding $\mathbf { z }$ and the metric $g$ , one obtains the following PDE:
30
+
31
+ $$
32
+ \frac { \partial \mathbf { z } ( \mathbf { u } , t ) } { \partial t } = \Delta _ { \mathbf { G } } \mathbf { z } ( \mathbf { u } , t ) , \qquad \mathbf { z } ( \mathbf { u } , 0 ) = \mathbf { z } ( \mathbf { u } ) , \quad t \geq 0 ,
33
+ $$
34
+
35
+ and boundary conditions as appropriate. Here $\Delta _ { \mathbf { G } }$ is the Laplace-Beltrami operator, the Laplacian operator induced on $\Sigma$ by the Euclidean space we embed the image into. Namely, the embedding of the manifold allows us to pull-back the Euclidean distance structure on the image: the distance between two nearby points $\mathbf { u }$ and $\mathbf { u } + \mathrm { d } \mathbf { u }$ is given by
36
+
37
+ $$
38
+ \mathrm { d } \ell ^ { 2 } = \mathrm { d } \mathbf { u } ^ { \top } \mathbf { G } ( \mathbf { u } ) \mathrm { d } \mathbf { u } = \mathrm { d } u _ { 1 } ^ { 2 } + \mathrm { d } u _ { 2 } ^ { 2 } + \alpha ^ { 2 } \sum _ { i = 1 } ^ { d } \mathrm { d } x _ { i } ^ { 2 } ,
39
+ $$
40
+
41
+ where $\mathbf { G } = \mathbf { I } + \alpha ^ { 2 } ( \nabla _ { \mathbf { u } } \mathbf { x } ( \mathbf { u } ) ) ^ { \top } \nabla _ { \mathbf { u } } \mathbf { x } ( \mathbf { u } )$ is a $2 \times 2$ matrix called the Riemannian metric. The fact that the distance is a combination of the positional component (distance between pixels in the plane, $\| \mathbf { u } - \mathbf { u } ^ { \prime } \| )$ and the colour component (distance between the colours of the pixels, $\| \mathbf { \bar { x } } ( \mathbf { u } ) - \mathbf { x } ( \mathbf { u } ^ { \bar { \prime } } ) \| )$ is crucial as it allows edge-preserving image diffusion.
42
+
43
+ When dealing with images, the evolution of the first two components of $( z _ { 1 } , z _ { 2 } ) = \mathbf { u }$ is a nuisance amounting to the reparametrisation of the manifold and can be ignored. For grayscale images (the case when $d = 1$ and ${ \bf z } = ( u _ { 1 } , u _ { 2 } , x ) )$ , this is done by projection along the dimension $z _ { 3 }$ , in which case the Beltrami flow takes the form of an inhomogeneous diffusion equation of $x$ ,
44
+
45
+ $$
46
+ \frac { \partial x ( \mathbf { u } , t ) } { \partial t } = \frac { 1 } { \sqrt { \operatorname* { d e t } \mathbf { G } ( \mathbf { u } , t ) } } \mathrm { d i v } \left( \frac { \nabla x ( \mathbf { u } , t ) } { \sqrt { \operatorname* { d e t } \mathbf { G } ( \mathbf { u } , t ) } } \right) \qquad t \geq 0 .
47
+ $$
48
+
49
+ The diffusivity
50
+
51
+ $$
52
+ a = \frac { 1 } { \sqrt { \mathrm { d e t } \mathbf { G } } } = \frac { 1 } { \sqrt { 1 + \alpha ^ { 2 } \| \nabla x \| ^ { 2 } } }
53
+ $$
54
+
55
+ determining the speed of diffusion at each point, can be interpreted as an edge indicator: diffusion is weak across edges where $\| \nabla x \| \gg 1$ . The result is an adaptive diffusion [65] popular in image processing due to its ability to denoise images while preserving their edges. For cases with $d > 1$ (multiple colour channels), equation (3) is applied to each channel separately; however, the metric $\mathbf { G }$ couples the channels, which results in their gradients becoming aligned [38].
56
+
57
+ Special cases In the limit case $\alpha = 0$ , equation (3) becomes the simple homogeneous isotropic diffusion ∂∂t x = div(∇x) = ∆x, where ∆ = $\begin{array} { r } { \Delta = \frac { \partial ^ { 2 } } { \partial u _ { 1 } ^ { 2 } } + \frac { \partial ^ { 2 } } { \partial u _ { 2 } ^ { 2 } } } \end{array}$ is the standard Euclidean Laplacian operator. The solution is given in closed form as the convolution of the initial image and a Gaussian kernel with time-dependent variance,
58
+
59
+ ![](images/19cfd037de15d7a638286db4726e77dd6428436abe263995cbbcf0882de82ac1.jpg)
60
+ Figure 1: Two interpretations of the Beltrami flow: position-dependent bilateral kernel (top) and a Gaussian passed on the manifold (bottom).
61
+
62
+ $$
63
+ x ( { \mathbf { u } } , t ) = x ( { \mathbf { u } } , 0 ) \star \frac { 1 } { ( 4 \pi t ) ^ { d / 2 } } e ^ { - \| { \mathbf { u } } \| ^ { 2 } / 4 t }
64
+ $$
65
+
66
+ and can be considered a simple linear low-pass filtering. In the limit $t \to \infty$ , the image becomes constant and equal to the average colour.2
67
+
68
+ Another interpretation of the Beltrami flow is passing a Gaussian on the manifold (see Figure 1, bottom), which can locally be expressed as non-linear filtering with the bilateral kernel [85] dependent on the joint positional and colour distance (Figure 1, top),
69
+
70
+ $$
71
+ x ( \mathbf { u } , t ) = \frac { 1 } { ( 4 \pi t ) ^ { d / 2 } } \int _ { \mathbb { R } ^ { 2 } } x ( \mathbf { v } , 0 ) e ^ { - \| \mathbf { u } - \mathbf { v } \| ^ { 2 } / 4 t } e ^ { - \alpha ^ { 2 } \| \mathbf { x } ( \mathbf { u } , \mathbf { 0 } ) - \mathbf { x } ( \mathbf { v } , \mathbf { 0 } ) \| ^ { 2 } / 4 t } \mathrm { d } \mathbf { v } .
72
+ $$
73
+
74
+ For $\alpha = 0$ , the bilateral filter (6) reduces to a simple convolution with a time-dependent Gaussian.
75
+
76
+ # 3 Discrete Beltrami flow on graphs
77
+
78
+ We now develop the analogy of Beltrami flow for graphs. We consider a graph to be a discretisation of a continuous structure (manifold), and show that the evolution of the feature coordinates in time amounts to message passing layers in GNNs, whereas the evolution of the positional coordinates amounts to graph rewiring, which is used in some GNN architectures.
79
+
80
+ # 3.1 Graph Beltrami flow
81
+
82
+ Let ${ \mathcal { G } } = ( \gamma = \{ 1 , \ldots , n \} , \mathcal { E } )$ be an undirected graph, where $\nu$ and $\mathcal { E }$ denote node and edge sets, respectively. We further assume node-wise $d .$ -dimensional features $\mathbf { x } _ { i } \in \mathbb { R } ^ { d }$ for $i = 1 , \ldots , n$ . Denote by $\mathbf { z } _ { i } = ( \mathbf { u } _ { i } , \alpha \mathbf { x } _ { i } )$ the embedding of the graph in a joint space $\mathcal { C } \times \mathbb { R } ^ { d }$ , where $\mathcal { C }$ is a $d ^ { \prime }$ -dimensional space with a metric $d _ { \mathcal { C } }$ representing the node coordinates (for simplicity, we will assume $\mathcal { C } = \mathbb { R } ^ { d ^ { \prime } }$ unless otherwise stated). We refer to $\mathbf { u } _ { i } \mathbf { x } _ { i }$ and $\mathbf { z } _ { i }$ as the positional, feature and joint coordinates of node $i$ , respectively, and arrange them into the matrices U, $\mathbf { X }$ , and $\mathbf { Z }$ , of sizes $n \times d ^ { \prime } , n \times d$ , and $n \times ( d ^ { \prime } + d )$ .
83
+
84
+ For images, Beltrami flow amounts to evolving the embedding $\mathbf { z }$ along $\mathrm { d i v } ( a ( \mathbf { z } ) \nabla \mathbf { z } )$ , with $a$ a diffusivity map.3 Thus, we consider the graph Beltrami flow to be the discrete diffusion equation
85
+
86
+ $$
87
+ \frac { \partial \mathbf z _ { i } ( t ) } { \partial t } = \sum _ { j : ( i , j ) \in \mathcal E ^ { \prime } } a ( \mathbf z _ { i } ( t ) , \mathbf z _ { j } ( t ) ) ( \mathbf z _ { j } ( t ) - \mathbf z _ { i } ( t ) ) \qquad \mathbf z _ { i } ( 0 ) = \mathbf z _ { i } ; \quad i = 1 , \ldots , n ; \quad t \geq 0 .
88
+ $$
89
+
90
+ We motivate our definition as follows: $\mathbf { g } _ { i j } = \mathbf { z } _ { j } - \mathbf { z } _ { i }$ and $\begin{array} { r } { \mathbf { d } _ { i } = \sum _ { j : ( i , j ) \in \mathcal { E } } \mathbf { g } _ { i j } } \end{array}$ are the discrete analogies of the gradient $\nabla \mathbf { z }$ and divergence $\mathrm { d i v } ( \mathbf { g } )$ , both with respect to a graph $( \mathcal { V } , \mathcal { E } ^ { \prime } )$ that can be interpreted as the numerical stencil for the discretisation of the continuous Laplace-Beltrami operator in (3). Note that $\mathcal { E } ^ { \prime }$ can be different from the input $\mathcal { E }$ (referred to as ‘rewiring’). As discussed in Section 3.3, most GNNs use $\mathcal { E } ^ { \prime } = \mathcal { E }$ (input graph is used for diffusion, no rewiring). Alternatively, the positional coordinates of the nodes can be used to define a new graph topology either with $\mathcal { E } ( \bar { \mathbf { U } } ) = \{ ( i , j ) : d _ { \mathcal { C } } ( \mathbf { u } _ { i } , \mathbf { u } _ { j } ) < r \}$ , for some radius $r > 0$ , or using $k$ nearest neighbours. This new rewiring is precomputed using the input positional coordinates (i.e., $\mathcal { E } ^ { \prime } = \mathcal { E } ( \mathbf { U } ( 0 ) ) )$ or updated throughout the diffusion (i.e., $\mathcal { E } ^ { \prime } ( t \bar { ) } = \mathcal { E } ( \bar { \mathbf { U } } ( t \bar { ) } ) )$ . Therefore, (7) can be compactly rewritten as
91
+
92
+ $$
93
+ \frac { \partial { \bf z } _ { i } ( t ) } { \partial t } = \mathrm { d i v } \left( a ( { \bf z } ( t ) ) \nabla { \bf z } _ { i } ( t ) \right) .
94
+ $$
95
+
96
+ The function $a$ is the diffusivity controlling the diffusion strength between nodes $i$ and $j$ and is assumed to be normalised: $\begin{array} { r } { \sum _ { j : ( i , j ) \in \mathcal { E } ^ { \prime } } a ( \mathbf { z } _ { i } , \mathbf { z } _ { j } ) \ : = \ : 1 } \end{array}$ . The dependence of the diffusivity on the
97
+
98
+ embedding $\mathbf { z }$ matches the smooth PDE analysed in e.g. [82, Section 4.2] and is consistent with the form of attention mechanism used in e.g. [88, 86]. In matrix-form, we can also rewrite (7) as
99
+
100
+ $$
101
+ \begin{array} { r l } & { \left( \frac { \partial } { \partial t } \mathbf { U } ( t ) , \frac { \partial } { \partial t } \mathbf { X } ( t ) \right) = \left( \mathbf { A } ( \mathbf { U } ( t ) , \mathbf { X } ( t ) ) - \mathbf { I } \right) \left( \mathbf { U } ( t ) , \mathbf { X } ( t ) \right) } \\ & { \qquad \mathbf { U } ( 0 ) = \mathbf { U } ; \ \mathbf { X } ( 0 ) = \alpha \mathbf { X } ; \ t \geq 0 , } \end{array}
102
+ $$
103
+
104
+ where we emphasise the evolution of both the positional and feature components, coupled through the matrix-valued function A
105
+
106
+ $$
107
+ a _ { i j } ( t ) = \left\{ \begin{array} { l l } { a ( ( \mathbf { u } _ { i } ( t ) , \mathbf { x } _ { i } ( t ) ) , ( \mathbf { u } _ { j } ( t ) , \mathbf { x } _ { j } ( t ) ) ) } & { ( i , j ) \in \mathcal { E } ( \mathbf { U } ( t ) ) } \\ { 0 } & { \mathrm { e l s e } . } \end{array} \right.
108
+ $$
109
+
110
+ representing the diffusivity. The graph Beltrami flow produces an evolution of the joint positional and feature coordinates, ${ \bf Z } ( t ) = ( { \bf U } ( t ) , { \bf X } ( t ) )$ . In Section 3.3 we will show how the evolution of the feature coordinates $\mathbf X ( t )$ results in feature diffusion or message passing on the graph, the core of GNNs. As noted in Section 2, in the smooth case the Beltrami flow is obtained as gradient flow of an energy functional when minimised with respect to both the embedding and the metric on the surface (an image). When the embedding takes values in the Euclidean space, this leads to equations of the form (3) with no channel-mixing and an exact form of the diffusivity determined by the pull-back $\mathbf { G }$ of the Euclidean metric. To further motivate our approach, it is tempting to investigate whether a similar conclusion can be attained here. Although in the discrete case the operation of pull-back is not well-defined, we are able to derive that the gradient flow of a modified graph Dirichlet energy gives rise to an equation of the form (7). We note though that the gradient flow does not recover the exact form of the diffusivity implemented in this paper. This is not a limitation of the theory and should be expected: by requiring the gradient flow to avoid channel-mixing and imitate the image analogy in [82] and by inducing a discrete pull-back condition, we are imposing constraints on the problem. We leave the theoretical implications for future work and refer to the Supplementary Materials for a more thorough discussion, including definitions and proofs.
111
+
112
+ Theorem 1. Under structural assumptions on the diffusivity, graph Beltrami flow (7) is the gradient flow of the discrete Polyakov functional.
113
+
114
+ # 3.2 Numerical solvers
115
+
116
+ Explicit vs implicit schemes Equation (7) is solved numerically, which in the simplest case is done by replacing the continuous time derivative $\frac { \partial } { \partial t }$ with forward time difference:
117
+
118
+ $$
119
+ \frac { \mathbf { z } _ { i } ^ { ( k + 1 ) } - \mathbf { z } _ { i } ^ { ( k ) } } { \tau } = \sum _ { j : ( i , j ) \in \mathcal { E } ( \mathbf { U } ^ { ( k ) } ) } a \left( \mathbf { z } _ { i } ^ { ( k ) } , \mathbf { z } _ { j } ^ { ( k ) } \right) ( \mathbf { z } _ { j } ^ { ( k ) } - \mathbf { z } _ { i } ^ { ( k ) } ) .
120
+ $$
121
+
122
+ Here $k$ denotes the discrete time index (iteration) and $\tau$ is the time step (discretisation parameter). Rewriting (9) compactly in matrix-vector form with $\tau = 1$ leads to the explicit Euler scheme:
123
+
124
+ $$
125
+ \begin{array} { r } { { \bf Z } ^ { ( k + 1 ) } = ( { \bf A } ^ { ( k ) } - { \bf I } ) { \bf Z } ^ { ( k ) } = { \bf Q } ^ { ( k ) } { \bf Z } ^ { ( k ) } , } \end{array}
126
+ $$
127
+
128
+ where a(k)ij $a _ { i j } ^ { ( k ) } = a ( \mathbf { z } _ { i } ^ { ( k ) } , \mathbf { z } _ { j } ^ { ( k ) } )$ and the matrix $\mathbf { Q } ^ { ( k ) }$ (diffusion operator) is given by
129
+
130
+ $$
131
+ q _ { i j } ^ { ( k ) } = \left\{ \begin{array} { l l } { 1 - \tau \sum _ { l : ( i , l ) \in \mathcal { E } } a _ { i l } ^ { ( k ) } } & { i = j } \\ { \tau a _ { i j } ^ { ( k ) } } & { ( i , j ) \in \mathcal { E } ( \mathbf { U } ^ { ( k ) } ) } \\ { 0 } & { \mathrm { o t h e r w i s e } } \end{array} \right.
132
+ $$
133
+
134
+ The solution to the diffusion equation is computed by applying scheme (10) multiple times in sequence, starting from some initial $\mathbf { Z } ^ { ( 0 ) }$ . It is ‘explicit’ because the update $\mathbf { Z } ^ { ( k + 1 ) }$ is done directly by the application of the diffusion operator $\mathbf { Q } ^ { ( k ) }$ on $\mathbf { Z } ^ { ( k ) }$ (as opposed to implicit schemes of the form $\dot { \mathbf { Z } } ^ { ( k ) } = \dot { \mathbf { Q } } ^ { ( k ) } \mathbf { Z } ^ { ( k + 1 ) }$ arising from backward time differences that require inversion of the diffusion operator [91]).
135
+
136
+ Multi-step and adaptive schemes Higher-order approximation of temporal derivatives amount to using intermediate fractional steps, which are then linearly combined. Runge-Kutta (RK) [74, 44], ubiquitously used in numerical analysis, is a classical family of explicit numerical schemes, including Euler as a particular case. The Dormand-Prince (DOPRI) [24] is an RK method based on fifth and fourth-order approximations, the difference between which is used as an error estimate guiding the time step size [78].
137
+
138
+ Adaptive spatial discretisation and rewiring Many numerical PDE solvers also employ adaptive spatial discretisation. The choice of the stencil (mesh) for spatial derivatives is done based on the character of the solution at these points; in the simulation of phenomena such as shock waves it is often desired to use denser sampling in the respective regions of the domain, which can change in time. A class of techniques for adaptive rewiring of the spatial derivatives are known as Moving Mesh (MM) methods [33]. Interpreting the graph $\mathcal { E } ^ { \prime }$ in (7) as the numerical stencil for the discretisation of the continuous Laplace-Beltrami operator in (3), we can regard rewiring as a form of MM.
139
+
140
+ # 3.3 Relation to graph neural networks
141
+
142
+ Equation (9) has the structure of many GNN architectures of the ‘attentional’ type [10], where the discrete time index $k$ corresponds to a (convolutional or attentional) layer of the GNN and multiple diffusion iterations amount to a deep GNN. In the diffusion formalism, the time parameter $t$ acts as a continuous analogy of the layers, in the spirit of neural differential equations [19]. Typical GNNs amount to explicit single-step (Euler) discretisation schemes, whereas our continuous interpretation can exploit more efficient numerical schemes.
143
+
144
+ GNNs as instances of graph Beltrami flow The graph Beltrami framework leads to a family of graph neural networks that generalise many popular architectures (see Table 1). For example, GAT [88] can be obtained as a particular setting of our framework where the input graph is fixed $\mathbf { \mathcal { E } ^ { \prime } } = \mathbf { \mathcal { E } }$ ) and only the feature coordinates $\mathbf { X }$ are evolved. Equation (10) in this case becomes
145
+
146
+ $$
147
+ \mathbf { x } _ { i } ^ { ( k + 1 ) } = \mathbf { x } _ { i } ^ { ( k ) } + \tau \sum _ { j : ( i , j ) \in \mathcal { E } } a \left( \mathbf { x } _ { i } ^ { ( k ) } , \mathbf { x } _ { j } ^ { ( k ) } \right) ( \mathbf { x } _ { j } ^ { ( k ) } - \mathbf { x } _ { i } ^ { ( k ) } )
148
+ $$
149
+
150
+ and corresponds to the update formula of GAT with a residual connection and the assumption of no non-linearity between the layers. The role of the diffusivity is played by a learnable parametric attention function, which is generally time-dependent: $a ( \mathbf { z } _ { i } ^ { ( \bar { k } ) } , \bar { \mathbf { z } _ { j } ^ { ( k ) } } , k )$ . This results in separate attention parameters per layer $k$ , which can be learned independently. Our intentionally simplistic choice of a time-independent attention function amounts to parameter sharing across layers. We will show in Section 5.1 that this leads to a smaller model that is less likely to overfit.
151
+
152
+ Another popular architecture MoNet [57] uses linear diffusion of the features with weights dependent on the structure of the graph expressed as ‘pseudo-coordinates’, which can be cast as attention of the form $a ( \mathbf { u } _ { i } , \mathbf { u } _ { j } )$ . Transformers [87] can be interpreted as feature diffusion on a fixed complete graph with $\mathcal { E } ^ { \prime } = \mathcal { V } \times \mathcal { V }$ [10]. Positional encoding (used in Transformers as well as in several recent GNN architectures [9, 27]) amounts to attention dependent on both $\mathbf { X }$ and U, which allows the diffusion to adapt to the local structure of the graph; importantly, the positional coordinates U are precomputed. Similarly, DeepSets [97] and PointNet [70] architectures can be interpreted as GNNs applied on a graph with no edges ${ \mathcal { E } } ^ { \prime } = \emptyset$ ), where each node is treated independently of the others [10]. DIGL [43] performs graph rewiring as a pre-processing step using personalised page rank as positional coordinates, which are then fixed and not evolved. In the point cloud methods, DGCNN [89] and DGM [36], the graph is constructed based on the feature coordinates X and rewired adaptively (in our notation, $\bar { \mathcal { E } ^ { \prime } = \mathcal { E } ( \mathbf { X } ( t ) ) } )$ ).
153
+
154
+ <table><tr><td>Method</td><td>Evolution</td><td>Diffusivity</td><td>Graph (V,ε&#x27;)</td><td>Discretisation</td></tr><tr><td>ChebNet</td><td>Features X</td><td>Fixed aij</td><td>Fixed ε</td><td>Explicit fixed step</td></tr><tr><td>GAT</td><td>Features X</td><td>a(xi,Xj)</td><td>Fixedε</td><td>Explicit fixed step</td></tr><tr><td>MoNet</td><td>Features X</td><td>a(ui,uj)</td><td>Fixed ε</td><td>Explicit fixed step</td></tr><tr><td>Transformer</td><td>Features X</td><td>a(u,xi),(uj,Xj))</td><td>Fixedε=V×V</td><td>Explicit fixed step</td></tr><tr><td>DeepSet/PointNet</td><td>Features X</td><td>a(xi)</td><td>Fixedε=0</td><td>Explicit fixed step</td></tr><tr><td>DIGL*</td><td>Features X</td><td>a(xi,xj)</td><td>Fixed ε(U)</td><td>Explicit fixed step</td></tr><tr><td>DGCNN/DGM*</td><td>Features X</td><td>a(xi,xj)</td><td>Adaptive ε(X)</td><td>Explicit fixed step</td></tr><tr><td>Beltrami</td><td>Positions U +Features X</td><td>a((u,xi),(uj,xj))</td><td>Adaptive ε(U)</td><td>Explicit adaptive step / Implicit</td></tr></table>
155
+
156
+ Table 1: GNN architectures interpreted as particular instances of our framework. ∗Attentional variant.
157
+
158
+ Graph rewiring Multiple authors have recently argued in favor of decoupling the input graph from the graph used for diffusion. Such rewiring can take the form of graph sampling [32] to address scalability issues, data denoising [43], removal of information bottlenecks [1], or larger multi-hop filters [73]. The graph construction can also be made differentiable and a task-specific rewiring can be learned [89, 36]. The statement of Klicpera et al. [43] that ‘diffusion improves graph learning’, leading to the eponymous paradigm (DIGL), can be understood as a form of diffusion on the graph connectivity independent of the features. Specifically, the authors used as node positional encoding the Personalised PageRank (PPR), which can be interpreted as the steady-state of a diffusion process
159
+
160
+ $$
161
+ \mathbf { U } _ { \mathrm { P P R } } = \sum _ { k \geq 0 } ( 1 - \beta ) \beta ^ { k } \underline { { \mathbf { A } } } _ { \mathrm { R W } } ^ { k } = ( 1 - \beta ) ( \mathbf { I } - \beta \underline { { \mathbf { A } } } _ { \mathrm { R W } } ) ^ { - 1 } , \quad 0 < \beta < 1 ,
162
+ $$
163
+
164
+ where $\Delta _ { \mathrm { R W } }$ is the random walk graph Laplacian and $\beta \in ( 0 , 1 )$ is a parameter such that $1 - \beta$ represents the restart probability. The resulting positional encoding of dimension $d = n$ can be used to rewire the graph by $k \mathbf { N N }$ sampling, which corresponds to using $\mathcal { E } ^ { \prime } = \mathcal { E } ( \mathbf { U } _ { \mathrm { P P R } } )$ in our framework.
165
+
166
+ Numerical schemes All the aforementioned GNN architectures can be seen as an explicit discretisation of equation (7) with fixed step size. On the other hand, our continuous diffusion framework offers an additional advantage of employing more efficient numerical schemes with adaptive step size. Graph rewiring of the form $\mathcal { E } ^ { \prime } \overset { = } { = } \mathcal { \bar { E } } ( \mathbf { U } ( t ) )$ can be interpreted as adaptive spatial discretisation (moving mesh method).
167
+
168
+ # 3.4 Extensions
169
+
170
+ Non-Euclidean geometry There are multiple theoretical and empirical arguments [52, 17] in favor of using hyperbolic spaces to represent real-life ‘small-world’ graphs (in particular, scale-free networks can be obtained as $k \mathbf { N N }$ graphs in such spaces [8]). Our framework allows using a non-Euclidean metric $d _ { \mathcal { C } }$ for the positional coordinates U (Figure 2). In Section 5 we show that hyperbolic positional encodings allow for a significant reduction in model size with only a marginal degradation of performance.
171
+
172
+ Time-dependent diffusivity The diffusivity function $a$ which we assumed time-independent and which lead to parameter sharing across layers (i.e., updates of the form $\mathbf { \tilde { Z } } ^ { ( k + 1 ) } = \mathbf { Q } ( \mathbf { Z } ^ { ( k ) } , \pmb { \theta } ) \mathbf { Z } ^ { ( k ) } )$ can be made time-dependent of the form $\mathbf { Z } ^ { ( k + 1 ) } = \mathbf { Q } ( \mathbf { Z } ^ { ( k ) } , \pmb { \theta } ^ { ( k ) } ) \mathbf { Z } ^ { ( k ) }$ , where $\pmb { \theta }$ and $\pmb \theta ^ { ( k ) }$ denote shared and layer-dependent parameters, respectively.
173
+
174
+ ![](images/787c5eb76ed5b96795bf9345210b888e2cd6b6e8f5b8ec1263b5999a6dbc23df.jpg)
175
+ Figure 2: Graph Beltrami flow with hyperbolic positional coordinates.
176
+
177
+ Onsager diffusion As we noted, the Beltrami flow diffuses each channel separately. A more general variant of diffusion allowing for feature mixing is the Onsager diffusion [61] of the form $\begin{array} { r } { \dot { \frac { \partial } { \partial t } } \mathbf { Z } ( t ) = \mathbf { Q } ( \mathbf { Z } ( t ) ) \mathbf { Z } ( t ) \mathbf { W } ( t ) } \end{array}$ , where the matrix-valued function W acts across the channels. GCN [42] can be regarded a particular setting thereof, with update of the form .
178
+
179
+ MPNNs Finally, we note that the Beltrami flow amounts to linear aggregation with non-linear coefficients, or the ‘attentional’ flavor of GNNs [10]. The more general message passing flavor [30] is possible using a generic non-linear equation of the form $\begin{array} { r } { \frac { \partial } { \partial t } \mathbf { Z } ( t ) = \Psi ( \mathbf { Z } ( t ) ) } \end{array}$ .
180
+
181
+ # 4 BLEND: Beltrami Neural Diffusion
182
+
183
+ Beltrami Neural Diffusion (BLEND) is a novel class of graph neural network architectures derived from the graph Beltrami framework. We assume an input graph $\mathcal { G } = ( \nu , \mathcal { E } )$ with $n$ nodes and $d$ -dimensional node-wise features represented as a matrix $\mathbf { X } _ { \mathrm { i n } }$ . We further assume a $d ^ { \prime }$ -dimensional positional encoding $\mathbf { U } _ { \mathrm { i n } }$ of the graph nodes. BLEND architectures implement a learnable joint diffusion process of $\mathbf { U }$ and $\mathbf { X }$ and runs it for time $T$ , to produce an output node embeddings $\mathbf { Y }$ ,
184
+
185
+ $$
186
+ \mathbf { Z } ( 0 ) = \left( \phi ( \mathbf { U } _ { \mathrm { i n } } ) , \psi ( \mathbf { X } _ { \mathrm { i n } } ) \right) \qquad \mathbf { Z } ( T ) = \mathbf { Z } ( 0 ) + \int _ { 0 } ^ { T } \frac { \partial \mathbf { Z } ( t ) } { \partial t } \mathrm { d } t \qquad \mathbf { Y } = \xi ( \mathbf { Z } ( T ) ) ,
187
+ $$
188
+
189
+ where $\phi , \psi$ are learnable positional and feature encoders and $\xi$ is a learnable decoder (possibly changing the output dimensions). Here the $\alpha$ in Equations (2) and (8) is absorbed by $\psi$ and made learnable. $\frac { \partial \mathbf { Z } ( t ) } { \partial t }$ is given by the graph Beltrami flow equation (8), where the diffusivity function (attention) $a$ is also learnable. The choice of attention function depends on the geometry of the positional encoding and for Euclidean encodings we find the scaled dot product attention [86] performs well, in which case
190
+
191
+ $$
192
+ a ( \mathbf { z } _ { i } , \mathbf { z } _ { j } ) = \mathrm { s o f t m a x } \left( \frac { ( \mathbf { W } _ { K } \mathbf { z } _ { i } ) ^ { \top } \mathbf { W } _ { Q } \mathbf { z } _ { j } } { d _ { k } } \right)
193
+ $$
194
+
195
+ where $\mathbf { W } _ { K }$ and $\mathbf { W } _ { Q }$ are learned matrices, and $d _ { k }$ is a hyperparameter.
196
+
197
+ # 5 Experimental results
198
+
199
+ In this section, we compare the proposed Beltrami framework to popular GNN architectures on standard node classification benchmarks and provide a detailed study of the choice of the positional encoding space. Additional experiments and implementation details, including runtimes and hyperparameter tuning are given in the Supplementary Materials. The code is available at https://github.com/twitter-research/graph-neural-pde.
200
+
201
+ Datasets In our experiments, we use the following datasets: Cora [56], Citeseer [77], Pubmed [58], CoauthorCS [79], Amazon, Computer, and Photo [55], and OGB-arxiv [35]. Since many works using the first three datasets rely on the Planetoid splits [96], we included them Table 2, together with a more robust evaluation on 100 random splits with 20 random initialisations [79].
202
+
203
+ Baselines We compare to the following GNN architectures: GCN [42], GAT [88], MoNet [57] and GraphSAGE [32], and recent ODE-based GNN models: CGNN [94], GDE [67], GODE [98], and two versions of LanczosNet [49]. We use two variants of our method: using fixed input graph (BLEND) and using $k \mathbf { N N }$ graph in the positional coordinates (BLEND-knn).
204
+
205
+ # 5.1 Node Classification
206
+
207
+ In these experiments, we followed the methodology of [79] using 20 random weight initialisations for datasets with fixed Planetoid splits and 100 random splits otherwise. Where available, results from [79] are reported. Hyperparameters with the highest validation accuracy were chosen and results are reported on a test set that is used only once. Hyperparameter search used Ray Tune [50] with a thousand random trials using an asynchronous hyperband scheduler with a grace period and half life of ten epochs. The code to reproduce our results is included with the submission and will be released publicly fol
208
+
209
+ Table 2: Performance (test accuracy $\pm$ std) of different GNN models using Planetoid splits.
210
+
211
+ <table><tr><td>Method</td><td>CORA</td><td>CiteSeer</td><td>PubMed</td></tr><tr><td>GCN</td><td>81.9±0.8</td><td>69.5±0.9</td><td>79.0±0.5</td></tr><tr><td>GAT</td><td>82.8±0.5</td><td>71.0±0.6</td><td>77.0±1.3</td></tr><tr><td>MoNet</td><td>82.2±0.7</td><td>70.0±0.6</td><td>77.7±0.6</td></tr><tr><td>GS-maxpool</td><td>77.4±1.0</td><td>67.0±1.0</td><td>76.6±0.8</td></tr><tr><td>Lanczos</td><td>79.5±1.8</td><td>66.2±1.9</td><td>78.3±0.3</td></tr><tr><td>AdaLanczos</td><td>80.4±1.1</td><td>68.7±1.0</td><td>78.1±0.4</td></tr><tr><td>CGNN</td><td>81.7±0.7</td><td>68.1±1.2</td><td>80.2±0.3</td></tr><tr><td>GDE</td><td>83.8±0.5</td><td>72.5±0.5</td><td>79.9±0.3</td></tr><tr><td>GODE</td><td>83.3±0.3</td><td>72.4±0.6</td><td>80.1±0.3</td></tr><tr><td>BLEND</td><td>84.2±0.6</td><td>74.4±0.7</td><td>80.7± 0.7</td></tr><tr><td>BLEND-kNN</td><td>83.1±0.8</td><td>73.3±0.9</td><td>81.5±0.5</td></tr></table>
212
+
213
+ lowing the review process. Experiments ran on AWS p2.8xlarge machines, each with 8 Tesla V100-SXM2 GPUs.
214
+
215
+ Implementation details For all datasets excepting ogb-arxiv, adaptive explicit Dormand-Prince scheme was used as the numerical solver; for ogb-arxiv, we used the Runge-Kutta method. For the two smallest datasets (Cora and Citeseer) we performed direct backpropagation through each step of the numerical integrator. For the larger datasets, to reduce memory complexity, we use Pontryagin’s maximum principle to propagate gradients backwards in time [69]. For the larger datasets, kinetic energy and Jacobian regularisation [28, 37] was employed. The regularisation ensures the learned dynamics is well-conditioned and easily solvable by a numeric solver, which reduced training time. We use constant initialisation for the attention weights, ${ \mathbf W } _ { K } , { \mathbf W } _ { Q }$ , so training starts from a well-conditioned system that induces small regularisation penalty terms [28].
216
+
217
+ The space complexity of BLEND is dominated by evaluating attention (13) over edges and is $\mathcal { O } ( \vert \mathcal { E } ^ { \prime } \vert ( d + d ^ { \prime } ) )$ where $\mathcal { E } ^ { \prime }$ is the edge set following rewiring and $d$ is dimension of features and $d ^ { \prime }$ is the dimension of positional encoding. The runtime complexity is $\mathcal { O } ( | \mathcal { E } ^ { \prime } | ( d + d ^ { \prime } ) ) ( E _ { b } + E _ { f } )$ , split
218
+
219
+ <table><tr><td>Method</td><td>CORA</td><td>CiteSeer</td><td>PubMed</td><td>Coauthor CS</td><td>Computer</td><td>Photo</td><td>ogb-arxiv</td></tr><tr><td>GCN</td><td>81.5±1.3</td><td>71.9 ±1.9</td><td>77.8±2.9</td><td>91.1±0.5</td><td>82.6±2.4</td><td>91.2±1.2</td><td>71.74±0.29</td></tr><tr><td>GAT</td><td>81.8±1.3</td><td>71.4±1.9</td><td>78.7±2.3</td><td>90.5±0.6</td><td>78.0±19</td><td>85.7±20</td><td>73.01±0.19*</td></tr><tr><td>GAT-ppr</td><td>81.6±0.3</td><td>68.5±0.2</td><td>76.7±0.3</td><td>91.3±0.1</td><td>85.4±0.3</td><td>90.9±0.3</td><td>一</td></tr><tr><td>MoNet</td><td>81.3±1.3</td><td>71.2±2.0</td><td>78.6±2.3</td><td>90.8±0.6</td><td>83.5±2.2</td><td>91.2±2.3</td><td></td></tr><tr><td>GS-mean</td><td>79.2±7.7</td><td>71.6±1.9</td><td>77.4±2.2</td><td>91.3±2.8</td><td>82.4±1.8</td><td>91.4±1.3</td><td>71.49±0.27</td></tr><tr><td>GS-maxpool</td><td>76.6±1.9</td><td>67.5±2.3</td><td>76.1±2.3</td><td>85.0±1.1</td><td></td><td>90.4±1.3</td><td></td></tr><tr><td>CGNN</td><td>81.4±1.6</td><td>66.9±1.8</td><td>66.6±4.4</td><td>92.3±0.2</td><td>80.3±2.0</td><td>91.4±1.5</td><td>58.70±2.50</td></tr><tr><td>GDE</td><td>78.7±2.2</td><td>71.8±1.1</td><td>73.9±3.7</td><td>91.6±0.1</td><td>82.9±0.6</td><td>92.4±2.0</td><td>56.66±10.9</td></tr><tr><td>BLEND</td><td>84.8±0.9</td><td>75.9±1.3</td><td>79.5±1.4</td><td>92.9±0.2</td><td>86.9±0.6</td><td>92.9±0.6</td><td>72.56±0.1 +</td></tr><tr><td>BLEND-kNN</td><td>82.5±0.9</td><td>73.4±0.5</td><td>80.9±0.7</td><td>92.3±0.1</td><td>86.7±0.6</td><td>93.5±0.3</td><td></td></tr></table>
220
+
221
+ Table 3: Performance (test accuracy±std) of different GNN models using random splits. $^ { * } \mathrm { O G B }$ GAT reference has 1.5M parameters vs ours $7 0 \mathrm { K . \dagger }$ BLEND-kNN pre-processes the graph using the DIGL methodology [43] (Section 3.3), which constructs an n-dimensional representation of each node (an n-by-n matrix), then sparsifies into a kNN graph. The ogb-arxiv dataset has ${ > } 1 5 0 \mathrm { K }$ nodes and goes OOM. This is not a limitation of BLEND, but that of DIGL. Other forms of initial rewiring are possible, but we chose to compare with DIGL (arguably the most popular graph rewiring) and so this result is missing
222
+
223
+ between the forward and backward pass and can be dominated by either depending on the number of function evaluations $( E _ { b } , E _ { f } )$ .
224
+
225
+ Tables 2–3 summarise the results of our experiments. BLEND outperforms other GNNs in most of the experiments, showing state-of-the-art results on some datasets. Another important point to note is that compared GNNs use different sets of parameters per layer, whereas in BLEND, due to our choice of a time-independent attention, parameters are shared. This results in significantly fewer parameters: for comparison, the OGB versions of GCN, SAGE and GAT used in the ogb-arxiv experiment require 143K, 219K and 1.63M parameters respectively, compared to only 70K in BLEND.
226
+
227
+ # 5.2 Positional encoding
228
+
229
+ In the second experiment, we investigated the impact of the positional encodings and vary the dimensionality and underlying geometry, in order to showcase the flexibility of our framework.
230
+
231
+ Figure 3 shows that for all datasets BLEND is superior to a Euclidean model where positional encodings are not used, which corresponds to $Z \ = \ X$ (BLEND w/o positional in Figure 3) and a version of GAT where attention is over a concatenation of the same positional encodings used in BLEND and the features. The only exception is
232
+
233
+ ![](images/ea9c6d93db31838eeb5eb73be60167114b95bbc38d5f0785125854d6adc2db46.jpg)
234
+ Figure 3: An ablation study showing BLEND with and without positional encodings as well as GAT, the most similar conventional GNN with positional encodings added
235
+
236
+ CoathorCS, where the performance without positional encodings is unchanged.
237
+
238
+ We experimented with three forms of positional encoding: DIGL PPR embeddings of dimension $n$ [43], DeepWalk embeddings [66] of dimensions 16–256, and hyperbolic embeddings in the Poincare ball [15, 59] of dimension 2–16. Positional encodings are calculated as a preprocessing step and input as the U coordinates to BLEND. We calculated DeepWalk node embeddings using PyTorch Geometric’s Node2Vec implementation with parameters $p = 1$ , $q = 1$ , context size 20, and 16 walks per node. We trained with walk length ranging between 40 and 120 and took the embeddings with the highest accuracy for the link prediction task. Shallow hyperbolic embeddings were generated using the HGCN [17] implementation with the default parameters provided by the authors.
239
+
240
+ Figure 4 (left) compares the performance of the best model using DIGL $n$ -dimensional positional encodings against the best performing $d$ -dimensional hyperbolic positional encodings with $d$ tuned over the range 2-16. The average performance with DIGL positional encodings is 85.48, compared to 85.28 for hyperbolic encodings. Only in one case the DIGL encodings outperform the best hyperbolic encodings. Figure 4 (right) show the change in performance with the dimension $d ^ { \prime }$ of the positional encoding using a fixed hyperparameter configuration. As expected, we observe monotonic increase in performance as function of $d ^ { \prime }$ . Importantly most of the performance is captured by either 16 dimensional hyperbolic or positional encodings, with a small additional margin gained for going up to $d ^ { \prime } = n$ , which is impractical for larger graphs.
241
+
242
+ ![](images/cf2f3c6184ee4def22e15e97841984600ca4613036d2ee6fa001f1eb560430f0.jpg)
243
+ Figure 4: Left: performance comparison between Euclidean and hyperbolic positional embeddings. Right: results of positional embeddings ablation. Hyperbolic or Euclidean embeddings with $d ^ { \prime } = \bar { 1 } 6$ allow to obtain performances comparable to euclidean embeddings with $d ^ { \prime } = n > > 1 6$ .
244
+
245
+ # 5.3 Additional ablations
246
+
247
+ In addition to studying the affect of positional encodings , we performed ablation studies on the step size used in the integrator as well as different forms of attention.
248
+
249
+ In Figure 5 we studied the affect of changing the step size of the integrator using the explicit Euler method with a fixed terminal time set to be the optimal terminal time. The left hand side of the figure shows the performance using the adaptive stepsize Dopri5 for comparison. Dopri5 gives the most consistent performance and is best if three of the six datasets. Details of additional attention functions and their relative performance can be found in the Supplementary Material.
250
+
251
+ ![](images/7e73b7317e03a716d1f4115508a1d2d2936e06fc5f2ae39ad94121efc2b6334d.jpg)
252
+ Figure 5: step size against accurcy for explicit Euler compared to the adaptive dopri5.
253
+
254
+ # 6 Related work
255
+
256
+ Image processing, computer vision, and graphics. Following the Perona-Malik scheme [65], PDE-based approaches flourished in the 1990s with multiple versions of anisotropic [90] and nonEuclidean [82] diffusion used primarily for image denoising. The realisation of these ideas in the form of nonlinear filters [85, 12] was adopted in the industry. PDE-based methods were also used for low-level vision tasks including inpainting [6] and image segmentation [13, 18]. In computer graphics, fundamental solutions (‘heat kernels’) of diffusion equations were used as shape descriptors [83]. The closed-form expression of such solutions using the Laplace-Beltrami operator served as inspiration for some of the early approaches for GNNs [34, 23, 42, 45].
257
+
258
+ Neural differential equations The interpretation of neural networks as discretised differential equations (‘neural ODEs’) [19] was an influential recent result with multiple follow-up works [25, 28, 53, 46]. In deep learning on graphs, this mindset has been applied to GNN architectures [2, 67] and continuous message passing [94]. Continuous GNNs were also explored in [31] who, similarly to [76], addressed the solutions of fixed point equations. Ordinary Differential Equations
259
+
260
+ on Graph Networks (GODE)[98] approach the problem using the technique of invertible ResNets.
261
+ Finally, [75] used graph-based ODEs to generate physics simulations.
262
+
263
+ Physics-inspired learning Solving PDEs with deep learning has been explored by [72]. Neural networks appeared in [47] to accelerate PDE solvers with applications in the physical sciences. These have been applied to problems where the PDE can be described on a graph [48]. [4] consider the problem of predicting fluid flow and use a PDE inside a GNN. These approaches differ from ours in that they solve a given PDE, whereas we use the notion of discretising PDEs as a principle to understand and design GNNs.
264
+
265
+ Neural ODEs on graphs The most similar work to this is GRAND [16] of which BLEND can be considered a non-Euclidean extension. In addition there are several other works that apply the neural ODE framework to graphs. In GDE [67], GODE [98] and CGNN [94], the goal is to adapt neural ordinary differential equations to graphs. In contrast, we consider non-Euclidean partial differential equations and regard GNNs as particular instances of their discretisation (both spatial and temporal). We can naturally use spaces with any metric, in particular, extending recent works on hyperbolic graph embeddings. None of the previous techniques explore the link to differential geometry. More specifically, GDE, GODE, and CGNN consider Neural ODEs of the canonical form $\begin{array} { r } { \frac { \partial x } { \partial t } = f ( x , t , \theta ) } \end{array}$ where $f$ is a graph neural network (GODE), the message passing component of a GNN (CGNN), or restricting $f$ to be layers of bijective functions on graphs (GDE). Furthermore, in CGNN only ODEs with closed form solutions are considered. [75] on the other hand is quite distinct as they are not concerned with GNN design and instead use graph-based ODEs to generate physics simulations.
266
+
267
+ # 7 Conclusion
268
+
269
+ We developed a new class of graph neural networks based on the discretisation of a non-Euclidean diffusion PDE called Beltrami flow. We represent the graph structure and node features as a manifold in a joint continuous space, whose evolution by a parametric diffusion PDE (driven by the downstream learning task) gives rise to feature learning, positional encoding, and possibly also graph rewiring. Our experimental results show very good performance on popular benchmarks with a small fraction of parameters used by other GNN models. Perhaps most importantly, our framework establishes important links between GNNs, differential geometry, and PDEs – fields with a rich history and profound results that in our opinion are still insufficiently explored in the graph ML community.
270
+
271
+ Future directions While we show that our framework generalises many popular GNN architectures we seek to use the graph as a numerical representation of an underlying continuous space. This view of the graph as an approximation of a continuous latent structure is a common paradigm of manifold learning [84, 5, 20] and network geometry [8]. If adopted in graph ML, this mindset offers a rigorous mathematical framework for formalising and generalising some of the recent trends in the field, including the departure from the input graph as the basis for message passing [1], latent graph inference [41, 89, 29, 36] higher-order [7, 54] and directional [3] message passing (which can be expressed as anisotropic diffusion arising from additional structure of the underlying continuous space and different discretisation of the PDEs e.g. based on finite elements), and exploiting efficient numerical PDE solvers [19]. We leave these exciting directions for future research.
272
+
273
+ Societal impact GNNs have recently become increasingly utilized in industrial applications e.g. in recommender systems and social networks, and hence could potentially lead to a negative societal impact if used improperly. We would like to emphasize that our paper does not study such potential negative applications and the mathematical framework we develop could help to interpret and understand existing GNN models. We believe that better understanding of ML models is key to managing their potential societal implications and preventing their negative impact.
274
+
275
+ Limitations The assumption that the graph can be modelled as a discretisation of some continuous space makes our framework applicable only to cases where the edge and node features are continuous in nature. Applications e.g. to knowledge graphs with categorical attributes could only be addressed by first embedding such attributes in a continuous space. Finally, the structural result presented in Theorem 1 that link our model to the Polyakov action have not been implemented. This will be addressed in future works.
276
+
277
+ # 8 Acknowledgements
278
+
279
+ We thank Nils Hammerla and Gabriele Corso for feedback on early version of this manuscript. MB and JR are supported in part by ERC Consolidator grant no 724228 (LEMAN).
280
+
281
+ # References
282
+
283
+ [1] Uri Alon and Eran Yahav. On the bottleneck of graph neural networks and its practical implications. In ICLR, 2021.
284
+ [2] Pedro H. C. Avelar, Anderson R. Tavares, Marco Gori, and Luis C. Lamb. Discrete and continuous deep residual learning over graphs. arXiv:1911.09554, 2019. [3] Dominique Beaini, Saro Passaro, Vincent Létourneau, William L Hamilton, Gabriele Corso, and Pietro Liò. Directional graph networks. arXiv:2010.02863, 2020. [4] Filipe de Avila Belbute-Peres, Thomas Economon, and Zico Kolter. Combining differentiable pde solvers and graph neural networks for fluid flow prediction. In ICML, 2020. [5] Mikhail Belkin and Partha Niyogi. Laplacian eigenmaps for dimensionality reduction and data representation. Neural Computation, 15(6):1373–1396, 2003.
285
+ [6] Marcelo Bertalmio, Guillermo Sapiro, Vincent Caselles, and Coloma Ballester. Image inpainting. In PACMCGIT, 2000. [7] Cristian Bodnar, Fabrizio Frasca, Yu Guang Wang, Nina Otter, Guido Montúfar, Pietro Liò, and Michael Bronstein. Weisfeiler and lehman go topological: Message passing simplicial networks. arXiv:2103.03212, 2021.
286
+ [8] Marian Boguna, Ivan Bonamassa, Manlio De Domenico, Shlomo Havlin, Dmitri Krioukov, and M Ángeles Serrano. Network geometry. Nature Reviews Physics, pages 1–22, 2021.
287
+ [9] Giorgos Bouritsas, Fabrizio Frasca, Stefanos Zafeiriou, and Michael M Bronstein. Improving graph neural network expressivity via subgraph isomorphism counting. arXiv:2006.09252, 2020.
288
+ [10] Michael M Bronstein, Joan Bruna, Taco Cohen, and Petar Velickovi ˇ c. Geometric deep learning: ´ Grids, groups, graphs, geodesics, and gauges. arXiv:2104.13478, 2021.
289
+ [11] Michael M Bronstein and Iasonas Kokkinos. Scale-invariant heat kernel signatures for non-rigid shape recognition. In CVPR, 2010.
290
+ [12] Antoni Buades, Bartomeu Coll, and J-M Morel. A non-local algorithm for image denoising. In ICCV, 2005.
291
+ [13] Vicent Caselles, Ron Kimmel, and Guillermo Sapiro. Geodesic active contours. IJCV, 22(1):61–79, 1997.
292
+ [14] Soumen Chakrabarti. Dynamic personalized pagerank in entity-relation graphs. In WWW, 2007.
293
+ [15] Benjamin P. Chamberlain, James Clough, and Marc Peter Deisenroth. Neural embeddings of graphs in hyperbolic space. arXiv:1705.10359, 2017.
294
+ [16] Benjamin P. Chamberlain, James Rowbottom, Maria Gorinova, Michael M. Bronstein, Stefan Webb, and Emanuele Rossi. GRAND: graph neural diffusion. In ICML 2021, pages 1407–1418, 2021.
295
+ [17] Ines Chami, Rex Ying, Christopher Ré, and Jure Leskovec. Hyperbolic graph convolutional neural networks. In NeurIPS, 2019.
296
+ [18] Tony F Chan and Luminita A Vese. Active contours without edges. IEEE Trans. Image Processing, 10(2):266–277, 2001.
297
+ [19] Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David Duvenaud. Neural ordinary differential equations. In NeurIPS, pages 6571–6583, 2018.
298
+ [20] Ronald R Coifman and Stéphane Lafon. Diffusion maps. Applied and computational harmonic analysis, 21(1):5–30, 2006.
299
+ [21] Miles Cranmer, Sam Greydanus, Stephan Hoyer, Peter Battaglia, David Spergel, and Shirley Ho. Lagrangian neural networks. 2020.
300
+ [22] Miles Cranmer, Rui Xu, Peter Battaglia, and Shirley Ho. Learning symbolic physics with graph networks. arXiv:1909.05862, 2019.
301
+ [23] Michaël Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. In NeurIPS, 2016.
302
+ [24] John R Dormand and Peter J Prince. A family of embedded Runge-Kutta formulae. Journal of Computational and Applied Mathematics, 6(1):19–26, 1980.
303
+ [25] Emilien Dupont, Arnaud Doucet, and Yee Whye Teh. Augmented neural ODEs. In NeurIPS, pages 3134–3144, 2019.
304
+ [26] Frédo Durand and Julie Dorsey. Fast bilateral filtering for the display of high-dynamic-range images. In Proc Computer Graphics and Interactive Techniques, 2002.
305
+ [27] Vijay Prakash Dwivedi and Xavier Bresson. A generalization of transformer networks to graphs. arXiv:2012.09699, 2020.
306
+ [28] Chris Finlay, Jörn-Henrik Jacobsen, Levon Nurbekyan, and Adam M Oberman. How to train your neural ode: The world of Jacobian and kinetic regularization. In ICML, 2020.
307
+ [29] Luca Franceschi, Mathias Niepert, Massimiliano Pontil, and Xiao He. Learning discrete structures for graph neural networks. In ICML, 2019.
308
+ [30] Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E. Dahl. Neural message passing for quantum chemistry. In ICML, 2017.
309
+ [31] Fangda Gu, Heng Chang, Wenwu Zhu, Somayeh Sojoudi, and Laurent El Ghaoui. Implicit graph neural networks. In NeurIPS, 2020.
310
+ [32] William L Hamilton, Rex Ying, and Jure Leskovec. Inductive representation learning on large graphs. In NeurIPS, 2017.
311
+ [33] DF Hawken, JJ Gottlieb, and JS Hansen. Review of some adaptive node-movement techniques in finite-element and finite-difference solutions of partial differential equations. Journal of Computational Physics, 95(2):254–302, 1991.
312
+ [34] Mikael Henaff, Joan Bruna, and Yann LeCun. Deep convolutional networks on graph-structured data. arXiv:1506.05163, 2015.
313
+ [35] Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and Jure Leskovec. Open graph benchmark: Datasets for machine learning on graphs. arXiv:2005.00687, 2020.
314
+ [36] Anees Kazi, Luca Cosmo, Nassir Navab, and Michael Bronstein. Differentiable graph module (DGN) graph convolutional networks. arXiv:2002.04999, 2020.
315
+ [37] Jacob Kelly, Jesse Bettencourt, Matthew James Johnson, and David Duvenaud. Learning differential equations that are easy to solve. In NeurIPS, 2020.
316
+ [38] Ron Kimmel. Numerical geometry of images: Theory, algorithms, and applications. Springer, 2012.
317
+ [39] Ron Kimmel, Ravi Malladi, and Nir Sochen. Images as embedded maps and minimal surfaces: movies, color, texture, and volumetric medical images. IJCV, 39(2):111–129, 2000.
318
+ [40] Ron Kimmel, Nir Sochen, and Ravi Malladi. From high energy physics to low level vision. pages 236–247. Springer, 1997.
319
+ [41] Thomas Kipf, Ethan Fetaya, Kuan-Chieh Wang, Max Welling, and Richard Zemel. Neural relational inference for interacting systems. In ICML, 2018.
320
+ [42] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In ICLR, 2017.
321
+ [43] Johannes Klicpera, Stefan Weißenberger, and Stephan Günnemann. Diffusion improves graph learning. In NeurIPS, volume 32, 2019.
322
+ [44] Wilhelm Kutta. Beitrag zur naherungsweisen integration totaler differentialgleichungen. Z. Math. Phys., 46:435–453, 1901.
323
+ [45] Ron Levie, Federico Monti, Xavier Bresson, and Michael M Bronstein. Cayleynets: Graph convolutional neural networks with complex rational spectral filters. IEEE Trans. Signal Processing, 2017.
324
+ [46] Xuechen Li, Ricky Tian Qi Chen, Ting-Kam Leonard Wong, and David Duvenaud. Scalable gradients for stochastic differential equations. In Artificial Intelligence and Statistics, 2020.
325
+ [47] Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. 2020.
326
+ [48] Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Multipole graph neural operator for parametric partial differential equations. In NeurIPS, 2020.
327
+ [49] Renjie Liao, Zhizhen Zhao, Raquel Urtasun, and Richard S Zemel. Lanczosnet: Multi-scale deep graph convo-lutional networks. In 7th International Conference on Learning Representations, ICLR 2019, 2019.
328
+ [50] Richard Liaw, Eric Liang, Robert Nishihara, Philipp Moritz, Joseph E. Gonzalez, and Ion Stoica. Tune: A research platform for distributed model selection and training. In arXiv: 1807.05118., 2018.
329
+ [51] Roee Litman and Alexander M Bronstein. Learning spectral descriptors for deformable shape correspondence. PAMI, 36(1):171–180, 2013.
330
+ [52] Qi Liu, Maximilian Nickel, and Douwe Kiela. Hyperbolic graph neural networks. arXiv preprint arXiv:1910.12892, 2019.
331
+ [53] Xuanqing Liu, Tesi Xiao, Si Si, Qin Cao, Sanjiv Kumar, and Cho-Jui Hsieh. Neural SDE: Stabilizing neural ODE networks with stochastic noise. (2), 2019.
332
+ [54] Haggai Maron, Heli Ben-Hamu, Hadar Serviansky, and Yaron Lipman. Provably powerful graph networks. In NeurIPS, pages 2153–2164, 2019.
333
+ [55] Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton Van Den Hengel. Image-based recommendations on styles and substitutes. In Proceedings Information Retrieval, 2015.
334
+ [56] Andrew Kachites McCallum, Kamal Nigam, Jason Rennie, and Kristie Seymore. Automating the construction of internet portals with machine learning. Information Retrieval, 3(2):127–163, 2000.
335
+ [57] Federico Monti, Davide Boscaini, Jonathan Masci, Emanuele Rodolà, Jan Svoboda, and Michael M. Bronstein. Geometric deep learning on graphs and manifolds using mixture model CNNs. In CVPR, 2017.
336
+ [58] Galileo Namata, Ben London, Lise Getoor, Bert Huang, and UMD EDU. Query-driven active surveying for collective classification. In Proceedings Mining and Learning with Graphs, 2012.
337
+ [59] Maximillian Nickel and Douwe Kiela. Learning continuous hierarchies in the lorentz model of hyperbolic geometry. In International Conference on Machine Learning, pages 3779–3788. PMLR, 2018.
338
+ [60] Hoang NT and Takanori Maehara. Revisiting graph neural networks: All we have is low-pass filters. 2019.
339
+ [61] Lars Onsager. Reciprocal relations in irreversible processes. i. Physical Review, 37(4):405, 1931.
340
+ [62] Kenta Oono and Taiji Suzuki. Graph neural networks exponentially lose expressive power for node classification. In ICLR, 2020.
341
+ [63] Lawrence Page, Sergey Brin, Rajeev Motwani, and Terry Winograd. The pagerank citation ranking: Bringing order to the web. Technical report, 1999.
342
+ [64] Giuseppe Patané. Laplacian spectral kernels and distances for geometry processing and shape analysis. Computer Graphics Forum, 35(2):599–624, 2016.
343
+ [65] Pietro Perona and Jitendra Malik. Scale-space and edge detection using anisotropic diffusion. PAMI, 12(7):629–639, 1990.
344
+ [66] Bryan Perozzi, Rami Al-Rfou, and Steven Skiena. Deepwalk: Online learning of social representations. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 701–710, 2014.
345
+ [67] Michael Poli, Stefano Massaroli, Junyoung Park, Atsushi Yamashita, Hajime Asama, and Jinkyoo Park. Graph neural ordinary differential equations. pages 6571–6583, 2019.
346
+ [68] Alexander M Polyakov. Quantum geometry of bosonic strings. Physics Letters B, 103(3):207–210, 1981.
347
+ [69] Lev Semenovich Pontryagin. Mathematical theory of optimal processes. Routledge, 2018.
348
+ [70] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017.
349
+ [71] Alejandro F. Queiruga, N. Benjamin Erichson, Dane Taylor, and Michael W. Mahoney. Continuous-in-depth neural networks. CoRR, 2020.
350
+ [72] Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Physics informed deep learning (Part II): Data-driven discovery of nonlinear partial differential equations. CoRR, (Part II), 2017.
351
+ [73] Emanuele Rossi, Fabrizio Frasca, Ben Chamberlain, Davide Eynard, Michael Bronstein, and Federico Monti. SIGN: Scalable inception graph neural networks. arXiv:2004.11198, 2020.
352
+ [74] Carl Runge. Über die numerische auflösung von differentialgleichungen. Mathematische Annalen, 46(2):167–178, 1895.
353
+ [75] Alvaro Sanchez-Gonzalez, Victor Bapst, Kyle Cranmer, and Peter Battaglia. Hamiltonian graph networks with ODE integrators. 2019.
354
+ [76] Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE Trans. Neural Networks, 27(8):61–80, 2009.
355
+ [77] Prithviraj Sen, Galileo Namata, Mustafa Bilgic, Lise Getoor, Brian Galligher, and Tina EliassiRad. Collective classification in network data. AI Magazine, 29(3):93–93, 2008.
356
+ [78] Lawrence F Shampine. Some practical Runge-Kutta formulas. Mathematics of Computation, 46(173):135–150, 1986.
357
+ [79] Oleksandr Shchur, Maximilian Mumme, Aleksandar Bojchevski, and Stephan Günnemann. Pitfalls of graph neural network evaluation. arXiv:1811.05868, 2018.
358
+ [80] Nino Shervashidze, Pascal Schweitzer, Erik Jan Van Leeuwen, Kurt Mehlhorn, and Karsten M Borgwardt. Weisfeiler-lehman graph kernels. JMLR, 12:2539–2561, 2011.
359
+ [81] Jonathan Shlomi, Peter Battaglia, and Jean-Roch Vlimant. Graph neural networks in particle physics. Machine Learning: Science and Technology, 2(2):021001, 2020.
360
+ [82] Nir Sochen, Ron Kimmel, and Ravi Malladi. A general framework for low level vision. IEEE Trans. Image Processing, 7(3):310–318, 1998.
361
+ [83] Jian Sun, Maks Ovsjanikov, and Leonidas Guibas. A concise and provably informative multiscale signature based on heat diffusion. Computer Graphics Forum, 28(5):1383–1392, 2009.
362
+ [84] Joshua B Tenenbaum, Vin De Silva, and John C Langford. A global geometric framework for nonlinear dimensionality reduction. Dcience, 290(5500):2319–2323, 2000.
363
+ [85] Carlo Tomasi and Roberto Manduchi. Bilateral filtering for gray and color images. In ICCV, 1998.
364
+ [86] Ashish Vaswani, Noam Shazeer, Niki Parmar, Akob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998–6008, 2017.
365
+ [87] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. 2017.
366
+ [88] Petar Velickovi ˇ c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua ´ Bengio. Graph attention networks. In ICLR, 2018.
367
+ [89] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph CNN for learning on point clouds. ACM Trans. Graphics, 38(5):1–12, 2019.
368
+ [90] Joachim Weickert. Anisotropic diffusion in image processing. Teubner Stuttgart, 1998.
369
+ [91] Joachim Weickert, Bart M. Ter Haar Romeny, and Max A. Viergever. Efficient and reliable schemes for nonlinear diffusion filtering. IEEE Trans. Image Processing, 7(3):398–410, 1998.
370
+ [92] Felix Wu, Amauri Souza, Tianyi Zhang, Christopher Fifty, Tao Yu, and Kilian Weinberger. Simplifying graph convolutional networks. In ICML. PMLR, 2019.
371
+ [93] Tailin Wu, Hongyu Ren, Pan Li, and Jure Leskovec. Graph information bottleneck. NeurIPS, 2020.
372
+ [94] Louis-pascal A C Xhonneux, Meng Qu, and Jian Tang. Continuous graph neural networks. In ICML, 2020.
373
+ [95] Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? In ICLR, 2019.
374
+ [96] Zhilin Yang, William Cohen, and Ruslan Salakhudinov. Revisiting semi-supervised learning with graph embeddings. In International conference on machine learning, pages 40–48. PMLR, 2016.
375
+ [97] Manzil Zaheer, Satwik Kottur, Siamak Ravanbakhsh, Barnabas Poczos, Russ R Salakhutdinov, and Alexander J Smola. Deep sets. In NeurIPS, 2017.
376
+ [98] Juntang Zhuang, Nicha Dvornek, Xiaoxiao Li, and James S. Duncan. Ordinary differential equations on graph networks. Technical Report 1, 2020.
377
+
378
+ # Checklist
379
+
380
+ 1. For all authors...
381
+
382
+ (a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes]
383
+ (b) Did you describe the limitations of your work? [Yes] In limitations (S6) and an extensions (S3.4)
384
+ (c) Did you discuss any potential negative societal impacts of your work? [Yes] Section 6
385
+ (d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes]
386
+
387
+ 2. If you are including theoretical results...
388
+
389
+ (a) Did you state the full set of assumptions of all theoretical results? Theorem 1 is dealt with in the supplementary material [Yes] Supplementary Material
390
+ (b) Did you include complete proofs of all theoretical results? [Yes] Supplementary Material
391
+
392
+ 3. If you ran experiments...
393
+
394
+ (a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] In supplementary material
395
+ (b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] hyperparameters are included in Supplementary Material and splits are addressed in section 5
396
+ (c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [Yes] Results in Section 5 include error bars
397
+ (d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] AWS, details at the end of 5.1 and supplementary material
398
+
399
+ 4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...
400
+
401
+ (a) If your work uses existing assets, did you cite the creators? [Yes] papers have been cited for the most relevant OS libraries
402
+ (b) Did you mention the license of the assets? [No]
403
+ (c) Did you include any new assets either in the supplemental material or as a URL? [N/A]
404
+ (d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [N/A]
405
+ (e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [N/A]
406
+
407
+ 5. If you used crowdsourcing or conducted research with human subjects...
408
+
409
+ (a) Did you include the full text of instructions given to participants and screenshots, if applicable? [N/A]
410
+ (b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A]
411
+ (c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [N/A]
parse/train/4YlE2huxEsl/4YlE2huxEsl_content_list.json ADDED
@@ -0,0 +1,1650 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "type": "text",
4
+ "text": "Beltrami Flow and Neural Diffusion on Graphs ",
5
+ "text_level": 1,
6
+ "bbox": [
7
+ 210,
8
+ 122,
9
+ 789,
10
+ 147
11
+ ],
12
+ "page_idx": 0
13
+ },
14
+ {
15
+ "type": "text",
16
+ "text": "James Rowbottom∗ Twitter Inc. ",
17
+ "bbox": [
18
+ 470,
19
+ 202,
20
+ 607,
21
+ 228
22
+ ],
23
+ "page_idx": 0
24
+ },
25
+ {
26
+ "type": "text",
27
+ "text": "Benjamin P. Chamberlain∗ Twitter Inc. bchamberlain@twitter.com ",
28
+ "bbox": [
29
+ 223,
30
+ 200,
31
+ 411,
32
+ 243
33
+ ],
34
+ "page_idx": 0
35
+ },
36
+ {
37
+ "type": "text",
38
+ "text": "Davide Eynard Twitter Inc. ",
39
+ "bbox": [
40
+ 665,
41
+ 202,
42
+ 774,
43
+ 228
44
+ ],
45
+ "page_idx": 0
46
+ },
47
+ {
48
+ "type": "text",
49
+ "text": "rancesco Di Giovann Twitter Inc. ",
50
+ "bbox": [
51
+ 199,
52
+ 263,
53
+ 343,
54
+ 291
55
+ ],
56
+ "page_idx": 0
57
+ },
58
+ {
59
+ "type": "text",
60
+ "text": "Xiaowen Dong University of Oxford ",
61
+ "bbox": [
62
+ 370,
63
+ 265,
64
+ 511,
65
+ 291
66
+ ],
67
+ "page_idx": 0
68
+ },
69
+ {
70
+ "type": "text",
71
+ "text": "Michael M. Bronstein Twitter Inc. and Imperial College London ",
72
+ "bbox": [
73
+ 534,
74
+ 263,
75
+ 808,
76
+ 292
77
+ ],
78
+ "page_idx": 0
79
+ },
80
+ {
81
+ "type": "text",
82
+ "text": "Abstract ",
83
+ "text_level": 1,
84
+ "bbox": [
85
+ 462,
86
+ 328,
87
+ 535,
88
+ 344
89
+ ],
90
+ "page_idx": 0
91
+ },
92
+ {
93
+ "type": "text",
94
+ "text": "We propose a novel class of graph neural networks based on the discretised Beltrami flow, a non-Euclidean diffusion PDE. In our model, node features are supplemented with positional encodings derived from the graph topology and jointly evolved by the Beltrami flow, producing simultaneously continuous feature learning and topology evolution. The resulting model generalises many popular graph neural networks and achieves state-of-the-art results on several benchmarks. ",
95
+ "bbox": [
96
+ 232,
97
+ 361,
98
+ 766,
99
+ 443
100
+ ],
101
+ "page_idx": 0
102
+ },
103
+ {
104
+ "type": "text",
105
+ "text": "1 Introduction ",
106
+ "text_level": 1,
107
+ "bbox": [
108
+ 173,
109
+ 470,
110
+ 310,
111
+ 488
112
+ ],
113
+ "page_idx": 0
114
+ },
115
+ {
116
+ "type": "text",
117
+ "text": "The majority of graph neural networks (GNNs) are based on the message passing paradigm [30], wherein node features are learned by means of a non-linear propagation on the graph. Multiple recent works have pointed to the limitations of the message passing approach. These include; limited expressive power [80, 95, 9, 7], the related oversmoothing problem [60, 62] and the bottleneck phenomena [1, 93], which render such approaches inefficient, especially in deep GNNs. Multiple alternatives have been proposed, among which are higher-order methods [54, 7] and decoupling the propagation and input graphs by modifying the topology, often referred to as graph rewiring. Topological modifications can take different forms such as graph sampling [32], kNN [43], using the complete graph [86, 1], latent graph learning [89, 36], or multi-hop filters [92, 73]. However, there is no agreement in the literature on when and how to modify the graph, and a single principled framework for doing so. ",
118
+ "bbox": [
119
+ 174,
120
+ 503,
121
+ 825,
122
+ 655
123
+ ],
124
+ "page_idx": 0
125
+ },
126
+ {
127
+ "type": "text",
128
+ "text": "A somewhat underappreciated fact is that GNNs are intimately related to diffusion equations [16], a connection that was exploited in the early work of Scarselli et al. [76]. Diffusion PDEs have been historically important in computer graphics [83, 11, 51, 64], computer vision [13, 18, 6], and image processing [65, 82, 90, 85, 26, 12], where they created an entire trend of variational and PDE-based approaches. In machine learning and data science, diffusion equations underpin such popular manifold learning methods as eigenmaps [5] and diffusion maps [20], as well as the family of PageRank algorithms [63, 14]. In deep learning, differential equations are used as models of neural networks [16, 19, 25, 94, 71, 98] and for physics-informed learning [72, 22, 75, 21, 81, 47]. ",
129
+ "bbox": [
130
+ 174,
131
+ 661,
132
+ 825,
133
+ 773
134
+ ],
135
+ "page_idx": 0
136
+ },
137
+ {
138
+ "type": "text",
139
+ "text": "Main contributions In this paper, we propose a novel class of GNNs based on the discretised non-Euclidean diffusion PDE in joint positional and feature space, inspired by the Beltrami flow [82] used two decades ago in the image processing literature for edge-preserving image denoising. We show that the discretisation of the spatial component of the Beltrami flow offers a principled view on positional encoding and graph rewiring, whereas the discretisation of the temporal component can replace GNN layers with more flexible adaptive numerical schemes. Based on this model, we introduce Beltrami Neural Diffusion (BLEND) that generalises a broad range of GNN architectures and shows state-of-the-art performance on many popular benchmarks. In a broader perspective, our approach explores new tools from PDEs and differential geometry that are less well known in the graph ML community. ",
140
+ "bbox": [
141
+ 174,
142
+ 787,
143
+ 825,
144
+ 900
145
+ ],
146
+ "page_idx": 0
147
+ },
148
+ {
149
+ "type": "text",
150
+ "text": "",
151
+ "bbox": [
152
+ 173,
153
+ 90,
154
+ 825,
155
+ 119
156
+ ],
157
+ "page_idx": 1
158
+ },
159
+ {
160
+ "type": "text",
161
+ "text": "2 Background ",
162
+ "text_level": 1,
163
+ "bbox": [
164
+ 174,
165
+ 140,
166
+ 308,
167
+ 156
168
+ ],
169
+ "page_idx": 1
170
+ },
171
+ {
172
+ "type": "text",
173
+ "text": "Beltrami flow Kimmel et al. [40, 82, 39] considered images as 2-manifolds (parametric surfaces) $( \\Sigma , g )$ embedded in some larger ambient space as $\\mathbf { z } ( \\mathbf { u } ) \\overset { } { = } ( \\mathbf { u } , \\alpha \\mathbf { x } ( \\mathbf { u } ) ) \\ \\subseteq \\ \\mathbb { R } ^ { { \\hat { d } } + 2 }$ where $\\alpha \\geq 0$ is a scaling factor, $\\mathbf { u } = ( u _ { 1 } , u _ { 2 } )$ are the 2D positional coordinates of the pixels, and $\\mathbf { x }$ are the $d$ - dimensional colour or feature coordinates (with $d = 1$ or 3 for grayscale or RGB images, or $d = k ^ { 2 }$ when using $k \\times k$ patches as features [12]). In these works, the image is evolved along the gradient flow of a functional $S [ \\mathbf { z } , g ]$ called the Polyakov action $I 6 8 J$ , which roughly measures the smoothness of the embedding1. For images embedded in Euclidean space with the functional $S$ minimised with respect to both the embedding $\\mathbf { z }$ and the metric $g$ , one obtains the following PDE: ",
174
+ "bbox": [
175
+ 173,
176
+ 170,
177
+ 826,
178
+ 284
179
+ ],
180
+ "page_idx": 1
181
+ },
182
+ {
183
+ "type": "equation",
184
+ "img_path": "images/d9fda124b36f34e112063273f64df9ffe953b4efa65f18f341c29b3fd7067973.jpg",
185
+ "text": "$$\n\\frac { \\partial \\mathbf { z } ( \\mathbf { u } , t ) } { \\partial t } = \\Delta _ { \\mathbf { G } } \\mathbf { z } ( \\mathbf { u } , t ) , \\qquad \\mathbf { z } ( \\mathbf { u } , 0 ) = \\mathbf { z } ( \\mathbf { u } ) , \\quad t \\geq 0 ,\n$$",
186
+ "text_format": "latex",
187
+ "bbox": [
188
+ 320,
189
+ 287,
190
+ 676,
191
+ 320
192
+ ],
193
+ "page_idx": 1
194
+ },
195
+ {
196
+ "type": "text",
197
+ "text": "and boundary conditions as appropriate. Here $\\Delta _ { \\mathbf { G } }$ is the Laplace-Beltrami operator, the Laplacian operator induced on $\\Sigma$ by the Euclidean space we embed the image into. Namely, the embedding of the manifold allows us to pull-back the Euclidean distance structure on the image: the distance between two nearby points $\\mathbf { u }$ and $\\mathbf { u } + \\mathrm { d } \\mathbf { u }$ is given by ",
198
+ "bbox": [
199
+ 173,
200
+ 325,
201
+ 825,
202
+ 382
203
+ ],
204
+ "page_idx": 1
205
+ },
206
+ {
207
+ "type": "equation",
208
+ "img_path": "images/3e9e10b86a9cc60b366565b7bf7d426791b2c8bb1fa2552cd9f68d8e5986a0e2.jpg",
209
+ "text": "$$\n\\mathrm { d } \\ell ^ { 2 } = \\mathrm { d } \\mathbf { u } ^ { \\top } \\mathbf { G } ( \\mathbf { u } ) \\mathrm { d } \\mathbf { u } = \\mathrm { d } u _ { 1 } ^ { 2 } + \\mathrm { d } u _ { 2 } ^ { 2 } + \\alpha ^ { 2 } \\sum _ { i = 1 } ^ { d } \\mathrm { d } x _ { i } ^ { 2 } ,\n$$",
210
+ "text_format": "latex",
211
+ "bbox": [
212
+ 334,
213
+ 390,
214
+ 661,
215
+ 434
216
+ ],
217
+ "page_idx": 1
218
+ },
219
+ {
220
+ "type": "text",
221
+ "text": "where $\\mathbf { G } = \\mathbf { I } + \\alpha ^ { 2 } ( \\nabla _ { \\mathbf { u } } \\mathbf { x } ( \\mathbf { u } ) ) ^ { \\top } \\nabla _ { \\mathbf { u } } \\mathbf { x } ( \\mathbf { u } )$ is a $2 \\times 2$ matrix called the Riemannian metric. The fact that the distance is a combination of the positional component (distance between pixels in the plane, $\\| \\mathbf { u } - \\mathbf { u } ^ { \\prime } \\| )$ and the colour component (distance between the colours of the pixels, $\\| \\mathbf { \\bar { x } } ( \\mathbf { u } ) - \\mathbf { x } ( \\mathbf { u } ^ { \\bar { \\prime } } ) \\| )$ is crucial as it allows edge-preserving image diffusion. ",
222
+ "bbox": [
223
+ 173,
224
+ 440,
225
+ 825,
226
+ 498
227
+ ],
228
+ "page_idx": 1
229
+ },
230
+ {
231
+ "type": "text",
232
+ "text": "When dealing with images, the evolution of the first two components of $( z _ { 1 } , z _ { 2 } ) = \\mathbf { u }$ is a nuisance amounting to the reparametrisation of the manifold and can be ignored. For grayscale images (the case when $d = 1$ and ${ \\bf z } = ( u _ { 1 } , u _ { 2 } , x ) )$ , this is done by projection along the dimension $z _ { 3 }$ , in which case the Beltrami flow takes the form of an inhomogeneous diffusion equation of $x$ , ",
233
+ "bbox": [
234
+ 174,
235
+ 503,
236
+ 825,
237
+ 560
238
+ ],
239
+ "page_idx": 1
240
+ },
241
+ {
242
+ "type": "equation",
243
+ "img_path": "images/02086e7f80ade8e0c541eb11851810da0c88a58ab2088a46e3f8cb3dbe9b7415.jpg",
244
+ "text": "$$\n\\frac { \\partial x ( \\mathbf { u } , t ) } { \\partial t } = \\frac { 1 } { \\sqrt { \\operatorname* { d e t } \\mathbf { G } ( \\mathbf { u } , t ) } } \\mathrm { d i v } \\left( \\frac { \\nabla x ( \\mathbf { u } , t ) } { \\sqrt { \\operatorname* { d e t } \\mathbf { G } ( \\mathbf { u } , t ) } } \\right) \\qquad t \\geq 0 .\n$$",
245
+ "text_format": "latex",
246
+ "bbox": [
247
+ 299,
248
+ 568,
249
+ 699,
250
+ 609
251
+ ],
252
+ "page_idx": 1
253
+ },
254
+ {
255
+ "type": "text",
256
+ "text": "The diffusivity ",
257
+ "bbox": [
258
+ 174,
259
+ 613,
260
+ 269,
261
+ 627
262
+ ],
263
+ "page_idx": 1
264
+ },
265
+ {
266
+ "type": "equation",
267
+ "img_path": "images/d483613ce9e523513cd981769acff8e06daa741e580093d5f31f8c81ecd0091b.jpg",
268
+ "text": "$$\na = \\frac { 1 } { \\sqrt { \\mathrm { d e t } \\mathbf { G } } } = \\frac { 1 } { \\sqrt { 1 + \\alpha ^ { 2 } \\| \\nabla x \\| ^ { 2 } } }\n$$",
269
+ "text_format": "latex",
270
+ "bbox": [
271
+ 243,
272
+ 627,
273
+ 467,
274
+ 664
275
+ ],
276
+ "page_idx": 1
277
+ },
278
+ {
279
+ "type": "text",
280
+ "text": "determining the speed of diffusion at each point, can be interpreted as an edge indicator: diffusion is weak across edges where $\\| \\nabla x \\| \\gg 1$ . The result is an adaptive diffusion [65] popular in image processing due to its ability to denoise images while preserving their edges. For cases with $d > 1$ (multiple colour channels), equation (3) is applied to each channel separately; however, the metric $\\mathbf { G }$ couples the channels, which results in their gradients becoming aligned [38]. ",
281
+ "bbox": [
282
+ 173,
283
+ 669,
284
+ 537,
285
+ 795
286
+ ],
287
+ "page_idx": 1
288
+ },
289
+ {
290
+ "type": "text",
291
+ "text": "Special cases In the limit case $\\alpha = 0$ , equation (3) becomes the simple homogeneous isotropic diffusion ∂∂t x = div(∇x) = ∆x, where ∆ = $\\begin{array} { r } { \\Delta = \\frac { \\partial ^ { 2 } } { \\partial u _ { 1 } ^ { 2 } } + \\frac { \\partial ^ { 2 } } { \\partial u _ { 2 } ^ { 2 } } } \\end{array}$ is the standard Euclidean Laplacian operator. The solution is given in closed form as the convolution of the initial image and a Gaussian kernel with time-dependent variance, ",
292
+ "bbox": [
293
+ 173,
294
+ 810,
295
+ 537,
296
+ 873
297
+ ],
298
+ "page_idx": 1
299
+ },
300
+ {
301
+ "type": "image",
302
+ "img_path": "images/19cfd037de15d7a638286db4726e77dd6428436abe263995cbbcf0882de82ac1.jpg",
303
+ "image_caption": [
304
+ "Figure 1: Two interpretations of the Beltrami flow: position-dependent bilateral kernel (top) and a Gaussian passed on the manifold (bottom). "
305
+ ],
306
+ "image_footnote": [],
307
+ "bbox": [
308
+ 570,
309
+ 633,
310
+ 799,
311
+ 819
312
+ ],
313
+ "page_idx": 1
314
+ },
315
+ {
316
+ "type": "text",
317
+ "text": "",
318
+ "bbox": [
319
+ 173,
320
+ 90,
321
+ 825,
322
+ 121
323
+ ],
324
+ "page_idx": 2
325
+ },
326
+ {
327
+ "type": "equation",
328
+ "img_path": "images/64b155556902cc740dac7e45c9289660848a6892046e9aa5231fcec263542fda.jpg",
329
+ "text": "$$\nx ( { \\mathbf { u } } , t ) = x ( { \\mathbf { u } } , 0 ) \\star \\frac { 1 } { ( 4 \\pi t ) ^ { d / 2 } } e ^ { - \\| { \\mathbf { u } } \\| ^ { 2 } / 4 t }\n$$",
330
+ "text_format": "latex",
331
+ "bbox": [
332
+ 367,
333
+ 126,
334
+ 629,
335
+ 160
336
+ ],
337
+ "page_idx": 2
338
+ },
339
+ {
340
+ "type": "text",
341
+ "text": "and can be considered a simple linear low-pass filtering. In the limit $t \\to \\infty$ , the image becomes constant and equal to the average colour.2 ",
342
+ "bbox": [
343
+ 171,
344
+ 166,
345
+ 825,
346
+ 195
347
+ ],
348
+ "page_idx": 2
349
+ },
350
+ {
351
+ "type": "text",
352
+ "text": "Another interpretation of the Beltrami flow is passing a Gaussian on the manifold (see Figure 1, bottom), which can locally be expressed as non-linear filtering with the bilateral kernel [85] dependent on the joint positional and colour distance (Figure 1, top), ",
353
+ "bbox": [
354
+ 174,
355
+ 200,
356
+ 826,
357
+ 243
358
+ ],
359
+ "page_idx": 2
360
+ },
361
+ {
362
+ "type": "equation",
363
+ "img_path": "images/8791cc2195cd0c34672f15c2ca3052cd7ef484dbcad31bd37d5784303c458f16.jpg",
364
+ "text": "$$\nx ( \\mathbf { u } , t ) = \\frac { 1 } { ( 4 \\pi t ) ^ { d / 2 } } \\int _ { \\mathbb { R } ^ { 2 } } x ( \\mathbf { v } , 0 ) e ^ { - \\| \\mathbf { u } - \\mathbf { v } \\| ^ { 2 } / 4 t } e ^ { - \\alpha ^ { 2 } \\| \\mathbf { x } ( \\mathbf { u } , \\mathbf { 0 } ) - \\mathbf { x } ( \\mathbf { v } , \\mathbf { 0 } ) \\| ^ { 2 } / 4 t } \\mathrm { d } \\mathbf { v } .\n$$",
365
+ "text_format": "latex",
366
+ "bbox": [
367
+ 263,
368
+ 251,
369
+ 736,
370
+ 285
371
+ ],
372
+ "page_idx": 2
373
+ },
374
+ {
375
+ "type": "text",
376
+ "text": "For $\\alpha = 0$ , the bilateral filter (6) reduces to a simple convolution with a time-dependent Gaussian. ",
377
+ "bbox": [
378
+ 171,
379
+ 291,
380
+ 813,
381
+ 306
382
+ ],
383
+ "page_idx": 2
384
+ },
385
+ {
386
+ "type": "text",
387
+ "text": "3 Discrete Beltrami flow on graphs ",
388
+ "text_level": 1,
389
+ "bbox": [
390
+ 174,
391
+ 327,
392
+ 478,
393
+ 344
394
+ ],
395
+ "page_idx": 2
396
+ },
397
+ {
398
+ "type": "text",
399
+ "text": "We now develop the analogy of Beltrami flow for graphs. We consider a graph to be a discretisation of a continuous structure (manifold), and show that the evolution of the feature coordinates in time amounts to message passing layers in GNNs, whereas the evolution of the positional coordinates amounts to graph rewiring, which is used in some GNN architectures. ",
400
+ "bbox": [
401
+ 173,
402
+ 358,
403
+ 825,
404
+ 415
405
+ ],
406
+ "page_idx": 2
407
+ },
408
+ {
409
+ "type": "text",
410
+ "text": "3.1 Graph Beltrami flow ",
411
+ "text_level": 1,
412
+ "bbox": [
413
+ 174,
414
+ 433,
415
+ 357,
416
+ 448
417
+ ],
418
+ "page_idx": 2
419
+ },
420
+ {
421
+ "type": "text",
422
+ "text": "Let ${ \\mathcal { G } } = ( \\gamma = \\{ 1 , \\ldots , n \\} , \\mathcal { E } )$ be an undirected graph, where $\\nu$ and $\\mathcal { E }$ denote node and edge sets, respectively. We further assume node-wise $d .$ -dimensional features $\\mathbf { x } _ { i } \\in \\mathbb { R } ^ { d }$ for $i = 1 , \\ldots , n$ . Denote by $\\mathbf { z } _ { i } = ( \\mathbf { u } _ { i } , \\alpha \\mathbf { x } _ { i } )$ the embedding of the graph in a joint space $\\mathcal { C } \\times \\mathbb { R } ^ { d }$ , where $\\mathcal { C }$ is a $d ^ { \\prime }$ -dimensional space with a metric $d _ { \\mathcal { C } }$ representing the node coordinates (for simplicity, we will assume $\\mathcal { C } = \\mathbb { R } ^ { d ^ { \\prime } }$ unless otherwise stated). We refer to $\\mathbf { u } _ { i } \\mathbf { x } _ { i }$ and $\\mathbf { z } _ { i }$ as the positional, feature and joint coordinates of node $i$ , respectively, and arrange them into the matrices U, $\\mathbf { X }$ , and $\\mathbf { Z }$ , of sizes $n \\times d ^ { \\prime } , n \\times d$ , and $n \\times ( d ^ { \\prime } + d )$ . ",
423
+ "bbox": [
424
+ 173,
425
+ 457,
426
+ 826,
427
+ 560
428
+ ],
429
+ "page_idx": 2
430
+ },
431
+ {
432
+ "type": "text",
433
+ "text": "For images, Beltrami flow amounts to evolving the embedding $\\mathbf { z }$ along $\\mathrm { d i v } ( a ( \\mathbf { z } ) \\nabla \\mathbf { z } )$ , with $a$ a diffusivity map.3 Thus, we consider the graph Beltrami flow to be the discrete diffusion equation ",
434
+ "bbox": [
435
+ 169,
436
+ 565,
437
+ 823,
438
+ 594
439
+ ],
440
+ "page_idx": 2
441
+ },
442
+ {
443
+ "type": "equation",
444
+ "img_path": "images/d4f810cb147f3afd7d0d30f0edc1754ee4a2f674b411f479ded607f93c72947f.jpg",
445
+ "text": "$$\n\\frac { \\partial \\mathbf z _ { i } ( t ) } { \\partial t } = \\sum _ { j : ( i , j ) \\in \\mathcal E ^ { \\prime } } a ( \\mathbf z _ { i } ( t ) , \\mathbf z _ { j } ( t ) ) ( \\mathbf z _ { j } ( t ) - \\mathbf z _ { i } ( t ) ) \\qquad \\mathbf z _ { i } ( 0 ) = \\mathbf z _ { i } ; \\quad i = 1 , \\ldots , n ; \\quad t \\geq 0 .\n$$",
446
+ "text_format": "latex",
447
+ "bbox": [
448
+ 220,
449
+ 601,
450
+ 779,
451
+ 642
452
+ ],
453
+ "page_idx": 2
454
+ },
455
+ {
456
+ "type": "text",
457
+ "text": "We motivate our definition as follows: $\\mathbf { g } _ { i j } = \\mathbf { z } _ { j } - \\mathbf { z } _ { i }$ and $\\begin{array} { r } { \\mathbf { d } _ { i } = \\sum _ { j : ( i , j ) \\in \\mathcal { E } } \\mathbf { g } _ { i j } } \\end{array}$ are the discrete analogies of the gradient $\\nabla \\mathbf { z }$ and divergence $\\mathrm { d i v } ( \\mathbf { g } )$ , both with respect to a graph $( \\mathcal { V } , \\mathcal { E } ^ { \\prime } )$ that can be interpreted as the numerical stencil for the discretisation of the continuous Laplace-Beltrami operator in (3). Note that $\\mathcal { E } ^ { \\prime }$ can be different from the input $\\mathcal { E }$ (referred to as ‘rewiring’). As discussed in Section 3.3, most GNNs use $\\mathcal { E } ^ { \\prime } = \\mathcal { E }$ (input graph is used for diffusion, no rewiring). Alternatively, the positional coordinates of the nodes can be used to define a new graph topology either with $\\mathcal { E } ( \\bar { \\mathbf { U } } ) = \\{ ( i , j ) : d _ { \\mathcal { C } } ( \\mathbf { u } _ { i } , \\mathbf { u } _ { j } ) < r \\}$ , for some radius $r > 0$ , or using $k$ nearest neighbours. This new rewiring is precomputed using the input positional coordinates (i.e., $\\mathcal { E } ^ { \\prime } = \\mathcal { E } ( \\mathbf { U } ( 0 ) ) )$ or updated throughout the diffusion (i.e., $\\mathcal { E } ^ { \\prime } ( t \\bar { ) } = \\mathcal { E } ( \\bar { \\mathbf { U } } ( t \\bar { ) } ) )$ . Therefore, (7) can be compactly rewritten as ",
458
+ "bbox": [
459
+ 173,
460
+ 650,
461
+ 825,
462
+ 780
463
+ ],
464
+ "page_idx": 2
465
+ },
466
+ {
467
+ "type": "equation",
468
+ "img_path": "images/929a4a36e683353f2d413823de3a8b969d32a591c8616c6a58c2b34f3d56e72d.jpg",
469
+ "text": "$$\n\\frac { \\partial { \\bf z } _ { i } ( t ) } { \\partial t } = \\mathrm { d i v } \\left( a ( { \\bf z } ( t ) ) \\nabla { \\bf z } _ { i } ( t ) \\right) .\n$$",
470
+ "text_format": "latex",
471
+ "bbox": [
472
+ 395,
473
+ 786,
474
+ 602,
475
+ 819
476
+ ],
477
+ "page_idx": 2
478
+ },
479
+ {
480
+ "type": "text",
481
+ "text": "The function $a$ is the diffusivity controlling the diffusion strength between nodes $i$ and $j$ and is assumed to be normalised: $\\begin{array} { r } { \\sum _ { j : ( i , j ) \\in \\mathcal { E } ^ { \\prime } } a ( \\mathbf { z } _ { i } , \\mathbf { z } _ { j } ) \\ : = \\ : 1 } \\end{array}$ . The dependence of the diffusivity on the ",
482
+ "bbox": [
483
+ 173,
484
+ 824,
485
+ 825,
486
+ 857
487
+ ],
488
+ "page_idx": 2
489
+ },
490
+ {
491
+ "type": "text",
492
+ "text": "embedding $\\mathbf { z }$ matches the smooth PDE analysed in e.g. [82, Section 4.2] and is consistent with the form of attention mechanism used in e.g. [88, 86]. In matrix-form, we can also rewrite (7) as ",
493
+ "bbox": [
494
+ 173,
495
+ 90,
496
+ 823,
497
+ 119
498
+ ],
499
+ "page_idx": 3
500
+ },
501
+ {
502
+ "type": "equation",
503
+ "img_path": "images/28a3fc892d8afdb52be8f61892e2df431697edc499dccd42436db1a0caa5a7d9.jpg",
504
+ "text": "$$\n\\begin{array} { r l } & { \\left( \\frac { \\partial } { \\partial t } \\mathbf { U } ( t ) , \\frac { \\partial } { \\partial t } \\mathbf { X } ( t ) \\right) = \\left( \\mathbf { A } ( \\mathbf { U } ( t ) , \\mathbf { X } ( t ) ) - \\mathbf { I } \\right) \\left( \\mathbf { U } ( t ) , \\mathbf { X } ( t ) \\right) } \\\\ & { \\qquad \\mathbf { U } ( 0 ) = \\mathbf { U } ; \\ \\mathbf { X } ( 0 ) = \\alpha \\mathbf { X } ; \\ t \\geq 0 , } \\end{array}\n$$",
505
+ "text_format": "latex",
506
+ "bbox": [
507
+ 312,
508
+ 119,
509
+ 687,
510
+ 160
511
+ ],
512
+ "page_idx": 3
513
+ },
514
+ {
515
+ "type": "text",
516
+ "text": "where we emphasise the evolution of both the positional and feature components, coupled through the matrix-valued function A ",
517
+ "bbox": [
518
+ 171,
519
+ 160,
520
+ 823,
521
+ 188
522
+ ],
523
+ "page_idx": 3
524
+ },
525
+ {
526
+ "type": "equation",
527
+ "img_path": "images/9bacb15855978e384beb929cb11de46c50ab0bee5e653b9a14290578dda839ba.jpg",
528
+ "text": "$$\na _ { i j } ( t ) = \\left\\{ \\begin{array} { l l } { a ( ( \\mathbf { u } _ { i } ( t ) , \\mathbf { x } _ { i } ( t ) ) , ( \\mathbf { u } _ { j } ( t ) , \\mathbf { x } _ { j } ( t ) ) ) } & { ( i , j ) \\in \\mathcal { E } ( \\mathbf { U } ( t ) ) } \\\\ { 0 } & { \\mathrm { e l s e } . } \\end{array} \\right.\n$$",
529
+ "text_format": "latex",
530
+ "bbox": [
531
+ 284,
532
+ 188,
533
+ 702,
534
+ 222
535
+ ],
536
+ "page_idx": 3
537
+ },
538
+ {
539
+ "type": "text",
540
+ "text": "representing the diffusivity. The graph Beltrami flow produces an evolution of the joint positional and feature coordinates, ${ \\bf Z } ( t ) = ( { \\bf U } ( t ) , { \\bf X } ( t ) )$ . In Section 3.3 we will show how the evolution of the feature coordinates $\\mathbf X ( t )$ results in feature diffusion or message passing on the graph, the core of GNNs. As noted in Section 2, in the smooth case the Beltrami flow is obtained as gradient flow of an energy functional when minimised with respect to both the embedding and the metric on the surface (an image). When the embedding takes values in the Euclidean space, this leads to equations of the form (3) with no channel-mixing and an exact form of the diffusivity determined by the pull-back $\\mathbf { G }$ of the Euclidean metric. To further motivate our approach, it is tempting to investigate whether a similar conclusion can be attained here. Although in the discrete case the operation of pull-back is not well-defined, we are able to derive that the gradient flow of a modified graph Dirichlet energy gives rise to an equation of the form (7). We note though that the gradient flow does not recover the exact form of the diffusivity implemented in this paper. This is not a limitation of the theory and should be expected: by requiring the gradient flow to avoid channel-mixing and imitate the image analogy in [82] and by inducing a discrete pull-back condition, we are imposing constraints on the problem. We leave the theoretical implications for future work and refer to the Supplementary Materials for a more thorough discussion, including definitions and proofs. ",
541
+ "bbox": [
542
+ 173,
543
+ 222,
544
+ 825,
545
+ 444
546
+ ],
547
+ "page_idx": 3
548
+ },
549
+ {
550
+ "type": "text",
551
+ "text": "Theorem 1. Under structural assumptions on the diffusivity, graph Beltrami flow (7) is the gradient flow of the discrete Polyakov functional. ",
552
+ "bbox": [
553
+ 173,
554
+ 446,
555
+ 823,
556
+ 474
557
+ ],
558
+ "page_idx": 3
559
+ },
560
+ {
561
+ "type": "text",
562
+ "text": "3.2 Numerical solvers ",
563
+ "text_level": 1,
564
+ "bbox": [
565
+ 174,
566
+ 491,
567
+ 339,
568
+ 506
569
+ ],
570
+ "page_idx": 3
571
+ },
572
+ {
573
+ "type": "text",
574
+ "text": "Explicit vs implicit schemes Equation (7) is solved numerically, which in the simplest case is done by replacing the continuous time derivative $\\frac { \\partial } { \\partial t }$ with forward time difference: ",
575
+ "bbox": [
576
+ 173,
577
+ 515,
578
+ 821,
579
+ 546
580
+ ],
581
+ "page_idx": 3
582
+ },
583
+ {
584
+ "type": "equation",
585
+ "img_path": "images/4e119414b1122da66a190c1a893a9f33f1077aa991613b4dc5ceb300da9efc37.jpg",
586
+ "text": "$$\n\\frac { \\mathbf { z } _ { i } ^ { ( k + 1 ) } - \\mathbf { z } _ { i } ^ { ( k ) } } { \\tau } = \\sum _ { j : ( i , j ) \\in \\mathcal { E } ( \\mathbf { U } ^ { ( k ) } ) } a \\left( \\mathbf { z } _ { i } ^ { ( k ) } , \\mathbf { z } _ { j } ^ { ( k ) } \\right) ( \\mathbf { z } _ { j } ^ { ( k ) } - \\mathbf { z } _ { i } ^ { ( k ) } ) .\n$$",
587
+ "text_format": "latex",
588
+ "bbox": [
589
+ 315,
590
+ 549,
591
+ 681,
592
+ 594
593
+ ],
594
+ "page_idx": 3
595
+ },
596
+ {
597
+ "type": "text",
598
+ "text": "Here $k$ denotes the discrete time index (iteration) and $\\tau$ is the time step (discretisation parameter). Rewriting (9) compactly in matrix-vector form with $\\tau = 1$ leads to the explicit Euler scheme: ",
599
+ "bbox": [
600
+ 173,
601
+ 597,
602
+ 826,
603
+ 625
604
+ ],
605
+ "page_idx": 3
606
+ },
607
+ {
608
+ "type": "equation",
609
+ "img_path": "images/a28db2d33a6729262ce8cb62b13d1aacdefca94257ad675d8a972f0971cbab58.jpg",
610
+ "text": "$$\n\\begin{array} { r } { { \\bf Z } ^ { ( k + 1 ) } = ( { \\bf A } ^ { ( k ) } - { \\bf I } ) { \\bf Z } ^ { ( k ) } = { \\bf Q } ^ { ( k ) } { \\bf Z } ^ { ( k ) } , } \\end{array}\n$$",
611
+ "text_format": "latex",
612
+ "bbox": [
613
+ 367,
614
+ 626,
615
+ 629,
616
+ 646
617
+ ],
618
+ "page_idx": 3
619
+ },
620
+ {
621
+ "type": "text",
622
+ "text": "where a(k)ij $a _ { i j } ^ { ( k ) } = a ( \\mathbf { z } _ { i } ^ { ( k ) } , \\mathbf { z } _ { j } ^ { ( k ) } )$ and the matrix $\\mathbf { Q } ^ { ( k ) }$ (diffusion operator) is given by ",
623
+ "bbox": [
624
+ 174,
625
+ 648,
626
+ 692,
627
+ 667
628
+ ],
629
+ "page_idx": 3
630
+ },
631
+ {
632
+ "type": "equation",
633
+ "img_path": "images/c534a44b4749b71a7751a83df3018d50b78115ef205d12f175e5c6e8c9d237b3.jpg",
634
+ "text": "$$\nq _ { i j } ^ { ( k ) } = \\left\\{ \\begin{array} { l l } { 1 - \\tau \\sum _ { l : ( i , l ) \\in \\mathcal { E } } a _ { i l } ^ { ( k ) } } & { i = j } \\\\ { \\tau a _ { i j } ^ { ( k ) } } & { ( i , j ) \\in \\mathcal { E } ( \\mathbf { U } ^ { ( k ) } ) } \\\\ { 0 } & { \\mathrm { o t h e r w i s e } } \\end{array} \\right.\n$$",
635
+ "text_format": "latex",
636
+ "bbox": [
637
+ 343,
638
+ 670,
639
+ 645,
640
+ 738
641
+ ],
642
+ "page_idx": 3
643
+ },
644
+ {
645
+ "type": "text",
646
+ "text": "The solution to the diffusion equation is computed by applying scheme (10) multiple times in sequence, starting from some initial $\\mathbf { Z } ^ { ( 0 ) }$ . It is ‘explicit’ because the update $\\mathbf { Z } ^ { ( k + 1 ) }$ is done directly by the application of the diffusion operator $\\mathbf { Q } ^ { ( k ) }$ on $\\mathbf { Z } ^ { ( k ) }$ (as opposed to implicit schemes of the form $\\dot { \\mathbf { Z } } ^ { ( k ) } = \\dot { \\mathbf { Q } } ^ { ( k ) } \\mathbf { Z } ^ { ( k + 1 ) }$ arising from backward time differences that require inversion of the diffusion operator [91]). ",
647
+ "bbox": [
648
+ 173,
649
+ 739,
650
+ 825,
651
+ 814
652
+ ],
653
+ "page_idx": 3
654
+ },
655
+ {
656
+ "type": "text",
657
+ "text": "Multi-step and adaptive schemes Higher-order approximation of temporal derivatives amount to using intermediate fractional steps, which are then linearly combined. Runge-Kutta (RK) [74, 44], ubiquitously used in numerical analysis, is a classical family of explicit numerical schemes, including Euler as a particular case. The Dormand-Prince (DOPRI) [24] is an RK method based on fifth and fourth-order approximations, the difference between which is used as an error estimate guiding the time step size [78]. ",
658
+ "bbox": [
659
+ 174,
660
+ 827,
661
+ 825,
662
+ 912
663
+ ],
664
+ "page_idx": 3
665
+ },
666
+ {
667
+ "type": "text",
668
+ "text": "Adaptive spatial discretisation and rewiring Many numerical PDE solvers also employ adaptive spatial discretisation. The choice of the stencil (mesh) for spatial derivatives is done based on the character of the solution at these points; in the simulation of phenomena such as shock waves it is often desired to use denser sampling in the respective regions of the domain, which can change in time. A class of techniques for adaptive rewiring of the spatial derivatives are known as Moving Mesh (MM) methods [33]. Interpreting the graph $\\mathcal { E } ^ { \\prime }$ in (7) as the numerical stencil for the discretisation of the continuous Laplace-Beltrami operator in (3), we can regard rewiring as a form of MM. ",
669
+ "bbox": [
670
+ 173,
671
+ 90,
672
+ 825,
673
+ 189
674
+ ],
675
+ "page_idx": 4
676
+ },
677
+ {
678
+ "type": "text",
679
+ "text": "3.3 Relation to graph neural networks ",
680
+ "text_level": 1,
681
+ "bbox": [
682
+ 176,
683
+ 204,
684
+ 452,
685
+ 219
686
+ ],
687
+ "page_idx": 4
688
+ },
689
+ {
690
+ "type": "text",
691
+ "text": "Equation (9) has the structure of many GNN architectures of the ‘attentional’ type [10], where the discrete time index $k$ corresponds to a (convolutional or attentional) layer of the GNN and multiple diffusion iterations amount to a deep GNN. In the diffusion formalism, the time parameter $t$ acts as a continuous analogy of the layers, in the spirit of neural differential equations [19]. Typical GNNs amount to explicit single-step (Euler) discretisation schemes, whereas our continuous interpretation can exploit more efficient numerical schemes. ",
692
+ "bbox": [
693
+ 173,
694
+ 228,
695
+ 825,
696
+ 314
697
+ ],
698
+ "page_idx": 4
699
+ },
700
+ {
701
+ "type": "text",
702
+ "text": "GNNs as instances of graph Beltrami flow The graph Beltrami framework leads to a family of graph neural networks that generalise many popular architectures (see Table 1). For example, GAT [88] can be obtained as a particular setting of our framework where the input graph is fixed $\\mathbf { \\mathcal { E } ^ { \\prime } } = \\mathbf { \\mathcal { E } }$ ) and only the feature coordinates $\\mathbf { X }$ are evolved. Equation (10) in this case becomes ",
703
+ "bbox": [
704
+ 174,
705
+ 327,
706
+ 825,
707
+ 383
708
+ ],
709
+ "page_idx": 4
710
+ },
711
+ {
712
+ "type": "equation",
713
+ "img_path": "images/34f8a81dfe8d9197f87f13bf206ff2150630a7fae5bb3038c783b2d39f17f580.jpg",
714
+ "text": "$$\n\\mathbf { x } _ { i } ^ { ( k + 1 ) } = \\mathbf { x } _ { i } ^ { ( k ) } + \\tau \\sum _ { j : ( i , j ) \\in \\mathcal { E } } a \\left( \\mathbf { x } _ { i } ^ { ( k ) } , \\mathbf { x } _ { j } ^ { ( k ) } \\right) ( \\mathbf { x } _ { j } ^ { ( k ) } - \\mathbf { x } _ { i } ^ { ( k ) } )\n$$",
715
+ "text_format": "latex",
716
+ "bbox": [
717
+ 326,
718
+ 387,
719
+ 669,
720
+ 425
721
+ ],
722
+ "page_idx": 4
723
+ },
724
+ {
725
+ "type": "text",
726
+ "text": "and corresponds to the update formula of GAT with a residual connection and the assumption of no non-linearity between the layers. The role of the diffusivity is played by a learnable parametric attention function, which is generally time-dependent: $a ( \\mathbf { z } _ { i } ^ { ( \\bar { k } ) } , \\bar { \\mathbf { z } _ { j } ^ { ( k ) } } , k )$ . This results in separate attention parameters per layer $k$ , which can be learned independently. Our intentionally simplistic choice of a time-independent attention function amounts to parameter sharing across layers. We will show in Section 5.1 that this leads to a smaller model that is less likely to overfit. ",
727
+ "bbox": [
728
+ 174,
729
+ 428,
730
+ 825,
731
+ 518
732
+ ],
733
+ "page_idx": 4
734
+ },
735
+ {
736
+ "type": "text",
737
+ "text": "Another popular architecture MoNet [57] uses linear diffusion of the features with weights dependent on the structure of the graph expressed as ‘pseudo-coordinates’, which can be cast as attention of the form $a ( \\mathbf { u } _ { i } , \\mathbf { u } _ { j } )$ . Transformers [87] can be interpreted as feature diffusion on a fixed complete graph with $\\mathcal { E } ^ { \\prime } = \\mathcal { V } \\times \\mathcal { V }$ [10]. Positional encoding (used in Transformers as well as in several recent GNN architectures [9, 27]) amounts to attention dependent on both $\\mathbf { X }$ and U, which allows the diffusion to adapt to the local structure of the graph; importantly, the positional coordinates U are precomputed. Similarly, DeepSets [97] and PointNet [70] architectures can be interpreted as GNNs applied on a graph with no edges ${ \\mathcal { E } } ^ { \\prime } = \\emptyset$ ), where each node is treated independently of the others [10]. DIGL [43] performs graph rewiring as a pre-processing step using personalised page rank as positional coordinates, which are then fixed and not evolved. In the point cloud methods, DGCNN [89] and DGM [36], the graph is constructed based on the feature coordinates X and rewired adaptively (in our notation, $\\bar { \\mathcal { E } ^ { \\prime } = \\mathcal { E } ( \\mathbf { X } ( t ) ) } )$ ). ",
738
+ "bbox": [
739
+ 173,
740
+ 522,
741
+ 825,
742
+ 690
743
+ ],
744
+ "page_idx": 4
745
+ },
746
+ {
747
+ "type": "table",
748
+ "img_path": "images/b860055c9fbc29c19c29ba0aad8842fb0604d754f946961463e1469fbd941dff.jpg",
749
+ "table_caption": [],
750
+ "table_footnote": [
751
+ "Table 1: GNN architectures interpreted as particular instances of our framework. ∗Attentional variant. "
752
+ ],
753
+ "table_body": "<table><tr><td>Method</td><td>Evolution</td><td>Diffusivity</td><td>Graph (V,ε&#x27;)</td><td>Discretisation</td></tr><tr><td>ChebNet</td><td>Features X</td><td>Fixed aij</td><td>Fixed ε</td><td>Explicit fixed step</td></tr><tr><td>GAT</td><td>Features X</td><td>a(xi,Xj)</td><td>Fixedε</td><td>Explicit fixed step</td></tr><tr><td>MoNet</td><td>Features X</td><td>a(ui,uj)</td><td>Fixed ε</td><td>Explicit fixed step</td></tr><tr><td>Transformer</td><td>Features X</td><td>a(u,xi),(uj,Xj))</td><td>Fixedε=V×V</td><td>Explicit fixed step</td></tr><tr><td>DeepSet/PointNet</td><td>Features X</td><td>a(xi)</td><td>Fixedε=0</td><td>Explicit fixed step</td></tr><tr><td>DIGL*</td><td>Features X</td><td>a(xi,xj)</td><td>Fixed ε(U)</td><td>Explicit fixed step</td></tr><tr><td>DGCNN/DGM*</td><td>Features X</td><td>a(xi,xj)</td><td>Adaptive ε(X)</td><td>Explicit fixed step</td></tr><tr><td>Beltrami</td><td>Positions U +Features X</td><td>a((u,xi),(uj,xj))</td><td>Adaptive ε(U)</td><td>Explicit adaptive step / Implicit</td></tr></table>",
754
+ "bbox": [
755
+ 174,
756
+ 700,
757
+ 789,
758
+ 827
759
+ ],
760
+ "page_idx": 4
761
+ },
762
+ {
763
+ "type": "text",
764
+ "text": "Graph rewiring Multiple authors have recently argued in favor of decoupling the input graph from the graph used for diffusion. Such rewiring can take the form of graph sampling [32] to address scalability issues, data denoising [43], removal of information bottlenecks [1], or larger multi-hop filters [73]. The graph construction can also be made differentiable and a task-specific rewiring can be learned [89, 36]. The statement of Klicpera et al. [43] that ‘diffusion improves graph learning’, leading to the eponymous paradigm (DIGL), can be understood as a form of diffusion on the graph connectivity independent of the features. Specifically, the authors used as node positional encoding the Personalised PageRank (PPR), which can be interpreted as the steady-state of a diffusion process ",
765
+ "bbox": [
766
+ 174,
767
+ 856,
768
+ 825,
769
+ 911
770
+ ],
771
+ "page_idx": 4
772
+ },
773
+ {
774
+ "type": "text",
775
+ "text": "",
776
+ "bbox": [
777
+ 173,
778
+ 90,
779
+ 825,
780
+ 147
781
+ ],
782
+ "page_idx": 5
783
+ },
784
+ {
785
+ "type": "equation",
786
+ "img_path": "images/3dcf3ea781967fad9ea69fdf9aead9ffcc80a61ceb94f5dcd66c8a2433ba1837.jpg",
787
+ "text": "$$\n\\mathbf { U } _ { \\mathrm { P P R } } = \\sum _ { k \\geq 0 } ( 1 - \\beta ) \\beta ^ { k } \\underline { { \\mathbf { A } } } _ { \\mathrm { R W } } ^ { k } = ( 1 - \\beta ) ( \\mathbf { I } - \\beta \\underline { { \\mathbf { A } } } _ { \\mathrm { R W } } ) ^ { - 1 } , \\quad 0 < \\beta < 1 ,\n$$",
788
+ "text_format": "latex",
789
+ "bbox": [
790
+ 261,
791
+ 150,
792
+ 735,
793
+ 185
794
+ ],
795
+ "page_idx": 5
796
+ },
797
+ {
798
+ "type": "text",
799
+ "text": "where $\\Delta _ { \\mathrm { R W } }$ is the random walk graph Laplacian and $\\beta \\in ( 0 , 1 )$ is a parameter such that $1 - \\beta$ represents the restart probability. The resulting positional encoding of dimension $d = n$ can be used to rewire the graph by $k \\mathbf { N N }$ sampling, which corresponds to using $\\mathcal { E } ^ { \\prime } = \\mathcal { E } ( \\mathbf { U } _ { \\mathrm { P P R } } )$ in our framework. ",
800
+ "bbox": [
801
+ 176,
802
+ 189,
803
+ 826,
804
+ 233
805
+ ],
806
+ "page_idx": 5
807
+ },
808
+ {
809
+ "type": "text",
810
+ "text": "Numerical schemes All the aforementioned GNN architectures can be seen as an explicit discretisation of equation (7) with fixed step size. On the other hand, our continuous diffusion framework offers an additional advantage of employing more efficient numerical schemes with adaptive step size. Graph rewiring of the form $\\mathcal { E } ^ { \\prime } \\overset { = } { = } \\mathcal { \\bar { E } } ( \\mathbf { U } ( t ) )$ can be interpreted as adaptive spatial discretisation (moving mesh method). ",
811
+ "bbox": [
812
+ 173,
813
+ 244,
814
+ 826,
815
+ 315
816
+ ],
817
+ "page_idx": 5
818
+ },
819
+ {
820
+ "type": "text",
821
+ "text": "3.4 Extensions ",
822
+ "text_level": 1,
823
+ "bbox": [
824
+ 174,
825
+ 330,
826
+ 289,
827
+ 344
828
+ ],
829
+ "page_idx": 5
830
+ },
831
+ {
832
+ "type": "text",
833
+ "text": "Non-Euclidean geometry There are multiple theoretical and empirical arguments [52, 17] in favor of using hyperbolic spaces to represent real-life ‘small-world’ graphs (in particular, scale-free networks can be obtained as $k \\mathbf { N N }$ graphs in such spaces [8]). Our framework allows using a non-Euclidean metric $d _ { \\mathcal { C } }$ for the positional coordinates U (Figure 2). In Section 5 we show that hyperbolic positional encodings allow for a significant reduction in model size with only a marginal degradation of performance. ",
834
+ "bbox": [
835
+ 173,
836
+ 356,
837
+ 550,
838
+ 481
839
+ ],
840
+ "page_idx": 5
841
+ },
842
+ {
843
+ "type": "text",
844
+ "text": "Time-dependent diffusivity The diffusivity function $a$ which we assumed time-independent and which lead to parameter sharing across layers (i.e., updates of the form $\\mathbf { \\tilde { Z } } ^ { ( k + 1 ) } = \\mathbf { Q } ( \\mathbf { Z } ^ { ( k ) } , \\pmb { \\theta } ) \\mathbf { Z } ^ { ( k ) } )$ can be made time-dependent of the form $\\mathbf { Z } ^ { ( k + 1 ) } = \\mathbf { Q } ( \\mathbf { Z } ^ { ( k ) } , \\pmb { \\theta } ^ { ( k ) } ) \\mathbf { Z } ^ { ( k ) }$ , where $\\pmb { \\theta }$ and $\\pmb \\theta ^ { ( k ) }$ denote shared and layer-dependent parameters, respectively. ",
845
+ "bbox": [
846
+ 174,
847
+ 494,
848
+ 549,
849
+ 573
850
+ ],
851
+ "page_idx": 5
852
+ },
853
+ {
854
+ "type": "image",
855
+ "img_path": "images/787c5eb76ed5b96795bf9345210b888e2cd6b6e8f5b8ec1263b5999a6dbc23df.jpg",
856
+ "image_caption": [
857
+ "Figure 2: Graph Beltrami flow with hyperbolic positional coordinates. "
858
+ ],
859
+ "image_footnote": [],
860
+ "bbox": [
861
+ 566,
862
+ 313,
863
+ 787,
864
+ 516
865
+ ],
866
+ "page_idx": 5
867
+ },
868
+ {
869
+ "type": "text",
870
+ "text": "",
871
+ "bbox": [
872
+ 174,
873
+ 569,
874
+ 598,
875
+ 583
876
+ ],
877
+ "page_idx": 5
878
+ },
879
+ {
880
+ "type": "text",
881
+ "text": "Onsager diffusion As we noted, the Beltrami flow diffuses each channel separately. A more general variant of diffusion allowing for feature mixing is the Onsager diffusion [61] of the form $\\begin{array} { r } { \\dot { \\frac { \\partial } { \\partial t } } \\mathbf { Z } ( t ) = \\mathbf { Q } ( \\mathbf { Z } ( t ) ) \\mathbf { Z } ( t ) \\mathbf { W } ( t ) } \\end{array}$ , where the matrix-valued function W acts across the channels. GCN [42] can be regarded a particular setting thereof, with update of the form . ",
882
+ "bbox": [
883
+ 173,
884
+ 597,
885
+ 825,
886
+ 659
887
+ ],
888
+ "page_idx": 5
889
+ },
890
+ {
891
+ "type": "text",
892
+ "text": "MPNNs Finally, we note that the Beltrami flow amounts to linear aggregation with non-linear coefficients, or the ‘attentional’ flavor of GNNs [10]. The more general message passing flavor [30] is possible using a generic non-linear equation of the form $\\begin{array} { r } { \\frac { \\partial } { \\partial t } \\mathbf { Z } ( t ) = \\Psi ( \\mathbf { Z } ( t ) ) } \\end{array}$ . ",
893
+ "bbox": [
894
+ 173,
895
+ 671,
896
+ 825,
897
+ 717
898
+ ],
899
+ "page_idx": 5
900
+ },
901
+ {
902
+ "type": "text",
903
+ "text": "4 BLEND: Beltrami Neural Diffusion ",
904
+ "text_level": 1,
905
+ "bbox": [
906
+ 173,
907
+ 733,
908
+ 503,
909
+ 751
910
+ ],
911
+ "page_idx": 5
912
+ },
913
+ {
914
+ "type": "text",
915
+ "text": "Beltrami Neural Diffusion (BLEND) is a novel class of graph neural network architectures derived from the graph Beltrami framework. We assume an input graph $\\mathcal { G } = ( \\nu , \\mathcal { E } )$ with $n$ nodes and $d$ -dimensional node-wise features represented as a matrix $\\mathbf { X } _ { \\mathrm { i n } }$ . We further assume a $d ^ { \\prime }$ -dimensional positional encoding $\\mathbf { U } _ { \\mathrm { i n } }$ of the graph nodes. BLEND architectures implement a learnable joint diffusion process of $\\mathbf { U }$ and $\\mathbf { X }$ and runs it for time $T$ , to produce an output node embeddings $\\mathbf { Y }$ , ",
916
+ "bbox": [
917
+ 173,
918
+ 763,
919
+ 825,
920
+ 834
921
+ ],
922
+ "page_idx": 5
923
+ },
924
+ {
925
+ "type": "equation",
926
+ "img_path": "images/f23e2871b4e55bd9f8e1821733163e0933524b4da2eccdb7acefd9db3f39bd6c.jpg",
927
+ "text": "$$\n\\mathbf { Z } ( 0 ) = \\left( \\phi ( \\mathbf { U } _ { \\mathrm { i n } } ) , \\psi ( \\mathbf { X } _ { \\mathrm { i n } } ) \\right) \\qquad \\mathbf { Z } ( T ) = \\mathbf { Z } ( 0 ) + \\int _ { 0 } ^ { T } \\frac { \\partial \\mathbf { Z } ( t ) } { \\partial t } \\mathrm { d } t \\qquad \\mathbf { Y } = \\xi ( \\mathbf { Z } ( T ) ) ,\n$$",
928
+ "text_format": "latex",
929
+ "bbox": [
930
+ 230,
931
+ 837,
932
+ 766,
933
+ 873
934
+ ],
935
+ "page_idx": 5
936
+ },
937
+ {
938
+ "type": "text",
939
+ "text": "where $\\phi , \\psi$ are learnable positional and feature encoders and $\\xi$ is a learnable decoder (possibly changing the output dimensions). Here the $\\alpha$ in Equations (2) and (8) is absorbed by $\\psi$ and made learnable. $\\frac { \\partial \\mathbf { Z } ( t ) } { \\partial t }$ is given by the graph Beltrami flow equation (8), where the diffusivity function (attention) $a$ is also learnable. The choice of attention function depends on the geometry of the positional encoding and for Euclidean encodings we find the scaled dot product attention [86] performs well, in which case ",
940
+ "bbox": [
941
+ 173,
942
+ 882,
943
+ 821,
944
+ 912
945
+ ],
946
+ "page_idx": 5
947
+ },
948
+ {
949
+ "type": "text",
950
+ "text": "",
951
+ "bbox": [
952
+ 173,
953
+ 87,
954
+ 825,
955
+ 147
956
+ ],
957
+ "page_idx": 6
958
+ },
959
+ {
960
+ "type": "equation",
961
+ "img_path": "images/06867c90d4c6ee4d1d4e7d78b2b7f59a2f78c3f2e2f48f42070b56fbdce94598.jpg",
962
+ "text": "$$\na ( \\mathbf { z } _ { i } , \\mathbf { z } _ { j } ) = \\mathrm { s o f t m a x } \\left( \\frac { ( \\mathbf { W } _ { K } \\mathbf { z } _ { i } ) ^ { \\top } \\mathbf { W } _ { Q } \\mathbf { z } _ { j } } { d _ { k } } \\right)\n$$",
963
+ "text_format": "latex",
964
+ "bbox": [
965
+ 356,
966
+ 154,
967
+ 640,
968
+ 190
969
+ ],
970
+ "page_idx": 6
971
+ },
972
+ {
973
+ "type": "text",
974
+ "text": "where $\\mathbf { W } _ { K }$ and $\\mathbf { W } _ { Q }$ are learned matrices, and $d _ { k }$ is a hyperparameter. ",
975
+ "bbox": [
976
+ 173,
977
+ 195,
978
+ 635,
979
+ 212
980
+ ],
981
+ "page_idx": 6
982
+ },
983
+ {
984
+ "type": "text",
985
+ "text": "5 Experimental results ",
986
+ "text_level": 1,
987
+ "bbox": [
988
+ 174,
989
+ 231,
990
+ 380,
991
+ 248
992
+ ],
993
+ "page_idx": 6
994
+ },
995
+ {
996
+ "type": "text",
997
+ "text": "In this section, we compare the proposed Beltrami framework to popular GNN architectures on standard node classification benchmarks and provide a detailed study of the choice of the positional encoding space. Additional experiments and implementation details, including runtimes and hyperparameter tuning are given in the Supplementary Materials. The code is available at https://github.com/twitter-research/graph-neural-pde. ",
998
+ "bbox": [
999
+ 174,
1000
+ 261,
1001
+ 825,
1002
+ 332
1003
+ ],
1004
+ "page_idx": 6
1005
+ },
1006
+ {
1007
+ "type": "text",
1008
+ "text": "Datasets In our experiments, we use the following datasets: Cora [56], Citeseer [77], Pubmed [58], CoauthorCS [79], Amazon, Computer, and Photo [55], and OGB-arxiv [35]. Since many works using the first three datasets rely on the Planetoid splits [96], we included them Table 2, together with a more robust evaluation on 100 random splits with 20 random initialisations [79]. ",
1009
+ "bbox": [
1010
+ 173,
1011
+ 347,
1012
+ 825,
1013
+ 402
1014
+ ],
1015
+ "page_idx": 6
1016
+ },
1017
+ {
1018
+ "type": "text",
1019
+ "text": "Baselines We compare to the following GNN architectures: GCN [42], GAT [88], MoNet [57] and GraphSAGE [32], and recent ODE-based GNN models: CGNN [94], GDE [67], GODE [98], and two versions of LanczosNet [49]. We use two variants of our method: using fixed input graph (BLEND) and using $k \\mathbf { N N }$ graph in the positional coordinates (BLEND-knn). ",
1020
+ "bbox": [
1021
+ 173,
1022
+ 417,
1023
+ 825,
1024
+ 473
1025
+ ],
1026
+ "page_idx": 6
1027
+ },
1028
+ {
1029
+ "type": "text",
1030
+ "text": "5.1 Node Classification ",
1031
+ "text_level": 1,
1032
+ "bbox": [
1033
+ 174,
1034
+ 489,
1035
+ 346,
1036
+ 505
1037
+ ],
1038
+ "page_idx": 6
1039
+ },
1040
+ {
1041
+ "type": "text",
1042
+ "text": "In these experiments, we followed the methodology of [79] using 20 random weight initialisations for datasets with fixed Planetoid splits and 100 random splits otherwise. Where available, results from [79] are reported. Hyperparameters with the highest validation accuracy were chosen and results are reported on a test set that is used only once. Hyperparameter search used Ray Tune [50] with a thousand random trials using an asynchronous hyperband scheduler with a grace period and half life of ten epochs. The code to reproduce our results is included with the submission and will be released publicly fol",
1043
+ "bbox": [
1044
+ 174,
1045
+ 516,
1046
+ 485,
1047
+ 695
1048
+ ],
1049
+ "page_idx": 6
1050
+ },
1051
+ {
1052
+ "type": "table",
1053
+ "img_path": "images/1e0ab59a03c430d638cefe2fa8ba2ce9bc36c24c08d933f5f6432ae55652d257.jpg",
1054
+ "table_caption": [
1055
+ "Table 2: Performance (test accuracy $\\pm$ std) of different GNN models using Planetoid splits. "
1056
+ ],
1057
+ "table_footnote": [],
1058
+ "table_body": "<table><tr><td>Method</td><td>CORA</td><td>CiteSeer</td><td>PubMed</td></tr><tr><td>GCN</td><td>81.9±0.8</td><td>69.5±0.9</td><td>79.0±0.5</td></tr><tr><td>GAT</td><td>82.8±0.5</td><td>71.0±0.6</td><td>77.0±1.3</td></tr><tr><td>MoNet</td><td>82.2±0.7</td><td>70.0±0.6</td><td>77.7±0.6</td></tr><tr><td>GS-maxpool</td><td>77.4±1.0</td><td>67.0±1.0</td><td>76.6±0.8</td></tr><tr><td>Lanczos</td><td>79.5±1.8</td><td>66.2±1.9</td><td>78.3±0.3</td></tr><tr><td>AdaLanczos</td><td>80.4±1.1</td><td>68.7±1.0</td><td>78.1±0.4</td></tr><tr><td>CGNN</td><td>81.7±0.7</td><td>68.1±1.2</td><td>80.2±0.3</td></tr><tr><td>GDE</td><td>83.8±0.5</td><td>72.5±0.5</td><td>79.9±0.3</td></tr><tr><td>GODE</td><td>83.3±0.3</td><td>72.4±0.6</td><td>80.1±0.3</td></tr><tr><td>BLEND</td><td>84.2±0.6</td><td>74.4±0.7</td><td>80.7± 0.7</td></tr><tr><td>BLEND-kNN</td><td>83.1±0.8</td><td>73.3±0.9</td><td>81.5±0.5</td></tr></table>",
1059
+ "bbox": [
1060
+ 500,
1061
+ 496,
1062
+ 820,
1063
+ 642
1064
+ ],
1065
+ "page_idx": 6
1066
+ },
1067
+ {
1068
+ "type": "text",
1069
+ "text": "lowing the review process. Experiments ran on AWS p2.8xlarge machines, each with 8 Tesla V100-SXM2 GPUs. ",
1070
+ "bbox": [
1071
+ 176,
1072
+ 695,
1073
+ 823,
1074
+ 722
1075
+ ],
1076
+ "page_idx": 6
1077
+ },
1078
+ {
1079
+ "type": "text",
1080
+ "text": "Implementation details For all datasets excepting ogb-arxiv, adaptive explicit Dormand-Prince scheme was used as the numerical solver; for ogb-arxiv, we used the Runge-Kutta method. For the two smallest datasets (Cora and Citeseer) we performed direct backpropagation through each step of the numerical integrator. For the larger datasets, to reduce memory complexity, we use Pontryagin’s maximum principle to propagate gradients backwards in time [69]. For the larger datasets, kinetic energy and Jacobian regularisation [28, 37] was employed. The regularisation ensures the learned dynamics is well-conditioned and easily solvable by a numeric solver, which reduced training time. We use constant initialisation for the attention weights, ${ \\mathbf W } _ { K } , { \\mathbf W } _ { Q }$ , so training starts from a well-conditioned system that induces small regularisation penalty terms [28]. ",
1081
+ "bbox": [
1082
+ 174,
1083
+ 738,
1084
+ 825,
1085
+ 863
1086
+ ],
1087
+ "page_idx": 6
1088
+ },
1089
+ {
1090
+ "type": "text",
1091
+ "text": "The space complexity of BLEND is dominated by evaluating attention (13) over edges and is $\\mathcal { O } ( \\vert \\mathcal { E } ^ { \\prime } \\vert ( d + d ^ { \\prime } ) )$ where $\\mathcal { E } ^ { \\prime }$ is the edge set following rewiring and $d$ is dimension of features and $d ^ { \\prime }$ is the dimension of positional encoding. The runtime complexity is $\\mathcal { O } ( | \\mathcal { E } ^ { \\prime } | ( d + d ^ { \\prime } ) ) ( E _ { b } + E _ { f } )$ , split ",
1092
+ "bbox": [
1093
+ 176,
1094
+ 869,
1095
+ 825,
1096
+ 911
1097
+ ],
1098
+ "page_idx": 6
1099
+ },
1100
+ {
1101
+ "type": "table",
1102
+ "img_path": "images/43b8099e3c86c11458bbdc55c61dfa1133088786ecb3732bd591befb7f2d5d9a.jpg",
1103
+ "table_caption": [],
1104
+ "table_footnote": [],
1105
+ "table_body": "<table><tr><td>Method</td><td>CORA</td><td>CiteSeer</td><td>PubMed</td><td>Coauthor CS</td><td>Computer</td><td>Photo</td><td>ogb-arxiv</td></tr><tr><td>GCN</td><td>81.5±1.3</td><td>71.9 ±1.9</td><td>77.8±2.9</td><td>91.1±0.5</td><td>82.6±2.4</td><td>91.2±1.2</td><td>71.74±0.29</td></tr><tr><td>GAT</td><td>81.8±1.3</td><td>71.4±1.9</td><td>78.7±2.3</td><td>90.5±0.6</td><td>78.0±19</td><td>85.7±20</td><td>73.01±0.19*</td></tr><tr><td>GAT-ppr</td><td>81.6±0.3</td><td>68.5±0.2</td><td>76.7±0.3</td><td>91.3±0.1</td><td>85.4±0.3</td><td>90.9±0.3</td><td>一</td></tr><tr><td>MoNet</td><td>81.3±1.3</td><td>71.2±2.0</td><td>78.6±2.3</td><td>90.8±0.6</td><td>83.5±2.2</td><td>91.2±2.3</td><td></td></tr><tr><td>GS-mean</td><td>79.2±7.7</td><td>71.6±1.9</td><td>77.4±2.2</td><td>91.3±2.8</td><td>82.4±1.8</td><td>91.4±1.3</td><td>71.49±0.27</td></tr><tr><td>GS-maxpool</td><td>76.6±1.9</td><td>67.5±2.3</td><td>76.1±2.3</td><td>85.0±1.1</td><td></td><td>90.4±1.3</td><td></td></tr><tr><td>CGNN</td><td>81.4±1.6</td><td>66.9±1.8</td><td>66.6±4.4</td><td>92.3±0.2</td><td>80.3±2.0</td><td>91.4±1.5</td><td>58.70±2.50</td></tr><tr><td>GDE</td><td>78.7±2.2</td><td>71.8±1.1</td><td>73.9±3.7</td><td>91.6±0.1</td><td>82.9±0.6</td><td>92.4±2.0</td><td>56.66±10.9</td></tr><tr><td>BLEND</td><td>84.8±0.9</td><td>75.9±1.3</td><td>79.5±1.4</td><td>92.9±0.2</td><td>86.9±0.6</td><td>92.9±0.6</td><td>72.56±0.1 +</td></tr><tr><td>BLEND-kNN</td><td>82.5±0.9</td><td>73.4±0.5</td><td>80.9±0.7</td><td>92.3±0.1</td><td>86.7±0.6</td><td>93.5±0.3</td><td></td></tr></table>",
1106
+ "bbox": [
1107
+ 178,
1108
+ 87,
1109
+ 818,
1110
+ 222
1111
+ ],
1112
+ "page_idx": 7
1113
+ },
1114
+ {
1115
+ "type": "text",
1116
+ "text": "Table 3: Performance (test accuracy±std) of different GNN models using random splits. $^ { * } \\mathrm { O G B }$ GAT reference has 1.5M parameters vs ours $7 0 \\mathrm { K . \\dagger }$ BLEND-kNN pre-processes the graph using the DIGL methodology [43] (Section 3.3), which constructs an n-dimensional representation of each node (an n-by-n matrix), then sparsifies into a kNN graph. The ogb-arxiv dataset has ${ > } 1 5 0 \\mathrm { K }$ nodes and goes OOM. This is not a limitation of BLEND, but that of DIGL. Other forms of initial rewiring are possible, but we chose to compare with DIGL (arguably the most popular graph rewiring) and so this result is missing ",
1117
+ "bbox": [
1118
+ 173,
1119
+ 228,
1120
+ 825,
1121
+ 306
1122
+ ],
1123
+ "page_idx": 7
1124
+ },
1125
+ {
1126
+ "type": "text",
1127
+ "text": "between the forward and backward pass and can be dominated by either depending on the number of function evaluations $( E _ { b } , E _ { f } )$ . ",
1128
+ "bbox": [
1129
+ 173,
1130
+ 315,
1131
+ 823,
1132
+ 345
1133
+ ],
1134
+ "page_idx": 7
1135
+ },
1136
+ {
1137
+ "type": "text",
1138
+ "text": "Tables 2–3 summarise the results of our experiments. BLEND outperforms other GNNs in most of the experiments, showing state-of-the-art results on some datasets. Another important point to note is that compared GNNs use different sets of parameters per layer, whereas in BLEND, due to our choice of a time-independent attention, parameters are shared. This results in significantly fewer parameters: for comparison, the OGB versions of GCN, SAGE and GAT used in the ogb-arxiv experiment require 143K, 219K and 1.63M parameters respectively, compared to only 70K in BLEND. ",
1139
+ "bbox": [
1140
+ 173,
1141
+ 349,
1142
+ 825,
1143
+ 434
1144
+ ],
1145
+ "page_idx": 7
1146
+ },
1147
+ {
1148
+ "type": "text",
1149
+ "text": "5.2 Positional encoding ",
1150
+ "text_level": 1,
1151
+ "bbox": [
1152
+ 174,
1153
+ 450,
1154
+ 348,
1155
+ 465
1156
+ ],
1157
+ "page_idx": 7
1158
+ },
1159
+ {
1160
+ "type": "text",
1161
+ "text": "In the second experiment, we investigated the impact of the positional encodings and vary the dimensionality and underlying geometry, in order to showcase the flexibility of our framework. ",
1162
+ "bbox": [
1163
+ 173,
1164
+ 477,
1165
+ 388,
1166
+ 559
1167
+ ],
1168
+ "page_idx": 7
1169
+ },
1170
+ {
1171
+ "type": "text",
1172
+ "text": "Figure 3 shows that for all datasets BLEND is superior to a Euclidean model where positional encodings are not used, which corresponds to $Z \\ = \\ X$ (BLEND w/o positional in Figure 3) and a version of GAT where attention is over a concatenation of the same positional encodings used in BLEND and the features. The only exception is ",
1173
+ "bbox": [
1174
+ 173,
1175
+ 566,
1176
+ 390,
1177
+ 718
1178
+ ],
1179
+ "page_idx": 7
1180
+ },
1181
+ {
1182
+ "type": "image",
1183
+ "img_path": "images/ea9c6d93db31838eeb5eb73be60167114b95bbc38d5f0785125854d6adc2db46.jpg",
1184
+ "image_caption": [
1185
+ "Figure 3: An ablation study showing BLEND with and without positional encodings as well as GAT, the most similar conventional GNN with positional encodings added "
1186
+ ],
1187
+ "image_footnote": [],
1188
+ "bbox": [
1189
+ 413,
1190
+ 443,
1191
+ 813,
1192
+ 667
1193
+ ],
1194
+ "page_idx": 7
1195
+ },
1196
+ {
1197
+ "type": "text",
1198
+ "text": "CoathorCS, where the performance without positional encodings is unchanged. ",
1199
+ "bbox": [
1200
+ 174,
1201
+ 718,
1202
+ 694,
1203
+ 733
1204
+ ],
1205
+ "page_idx": 7
1206
+ },
1207
+ {
1208
+ "type": "text",
1209
+ "text": "We experimented with three forms of positional encoding: DIGL PPR embeddings of dimension $n$ [43], DeepWalk embeddings [66] of dimensions 16–256, and hyperbolic embeddings in the Poincare ball [15, 59] of dimension 2–16. Positional encodings are calculated as a preprocessing step and input as the U coordinates to BLEND. We calculated DeepWalk node embeddings using PyTorch Geometric’s Node2Vec implementation with parameters $p = 1$ , $q = 1$ , context size 20, and 16 walks per node. We trained with walk length ranging between 40 and 120 and took the embeddings with the highest accuracy for the link prediction task. Shallow hyperbolic embeddings were generated using the HGCN [17] implementation with the default parameters provided by the authors. ",
1210
+ "bbox": [
1211
+ 173,
1212
+ 738,
1213
+ 825,
1214
+ 849
1215
+ ],
1216
+ "page_idx": 7
1217
+ },
1218
+ {
1219
+ "type": "text",
1220
+ "text": "Figure 4 (left) compares the performance of the best model using DIGL $n$ -dimensional positional encodings against the best performing $d$ -dimensional hyperbolic positional encodings with $d$ tuned over the range 2-16. The average performance with DIGL positional encodings is 85.48, compared to 85.28 for hyperbolic encodings. Only in one case the DIGL encodings outperform the best hyperbolic encodings. Figure 4 (right) show the change in performance with the dimension $d ^ { \\prime }$ of the positional encoding using a fixed hyperparameter configuration. As expected, we observe monotonic increase in performance as function of $d ^ { \\prime }$ . Importantly most of the performance is captured by either 16 dimensional hyperbolic or positional encodings, with a small additional margin gained for going up to $d ^ { \\prime } = n$ , which is impractical for larger graphs. ",
1221
+ "bbox": [
1222
+ 174,
1223
+ 856,
1224
+ 823,
1225
+ 911
1226
+ ],
1227
+ "page_idx": 7
1228
+ },
1229
+ {
1230
+ "type": "image",
1231
+ "img_path": "images/cf2f3c6184ee4def22e15e97841984600ca4613036d2ee6fa001f1eb560430f0.jpg",
1232
+ "image_caption": [
1233
+ "Figure 4: Left: performance comparison between Euclidean and hyperbolic positional embeddings. Right: results of positional embeddings ablation. Hyperbolic or Euclidean embeddings with $d ^ { \\prime } = \\bar { 1 } 6$ allow to obtain performances comparable to euclidean embeddings with $d ^ { \\prime } = n > > 1 6$ . "
1234
+ ],
1235
+ "image_footnote": [],
1236
+ "bbox": [
1237
+ 179,
1238
+ 90,
1239
+ 789,
1240
+ 246
1241
+ ],
1242
+ "page_idx": 8
1243
+ },
1244
+ {
1245
+ "type": "text",
1246
+ "text": "",
1247
+ "bbox": [
1248
+ 174,
1249
+ 324,
1250
+ 825,
1251
+ 393
1252
+ ],
1253
+ "page_idx": 8
1254
+ },
1255
+ {
1256
+ "type": "text",
1257
+ "text": "5.3 Additional ablations ",
1258
+ "text_level": 1,
1259
+ "bbox": [
1260
+ 174,
1261
+ 414,
1262
+ 354,
1263
+ 429
1264
+ ],
1265
+ "page_idx": 8
1266
+ },
1267
+ {
1268
+ "type": "text",
1269
+ "text": "In addition to studying the affect of positional encodings , we performed ablation studies on the step size used in the integrator as well as different forms of attention. ",
1270
+ "bbox": [
1271
+ 174,
1272
+ 441,
1273
+ 452,
1274
+ 496
1275
+ ],
1276
+ "page_idx": 8
1277
+ },
1278
+ {
1279
+ "type": "text",
1280
+ "text": "In Figure 5 we studied the affect of changing the step size of the integrator using the explicit Euler method with a fixed terminal time set to be the optimal terminal time. The left hand side of the figure shows the performance using the adaptive stepsize Dopri5 for comparison. Dopri5 gives the most consistent performance and is best if three of the six datasets. Details of additional attention functions and their relative performance can be found in the Supplementary Material. ",
1281
+ "bbox": [
1282
+ 173,
1283
+ 503,
1284
+ 452,
1285
+ 642
1286
+ ],
1287
+ "page_idx": 8
1288
+ },
1289
+ {
1290
+ "type": "image",
1291
+ "img_path": "images/7e73b7317e03a716d1f4115508a1d2d2936e06fc5f2ae39ad94121efc2b6334d.jpg",
1292
+ "image_caption": [
1293
+ "Figure 5: step size against accurcy for explicit Euler compared to the adaptive dopri5. "
1294
+ ],
1295
+ "image_footnote": [],
1296
+ "bbox": [
1297
+ 480,
1298
+ 417,
1299
+ 789,
1300
+ 599
1301
+ ],
1302
+ "page_idx": 8
1303
+ },
1304
+ {
1305
+ "type": "text",
1306
+ "text": "",
1307
+ "bbox": [
1308
+ 173,
1309
+ 641,
1310
+ 549,
1311
+ 655
1312
+ ],
1313
+ "page_idx": 8
1314
+ },
1315
+ {
1316
+ "type": "text",
1317
+ "text": "6 Related work ",
1318
+ "text_level": 1,
1319
+ "bbox": [
1320
+ 174,
1321
+ 678,
1322
+ 316,
1323
+ 695
1324
+ ],
1325
+ "page_idx": 8
1326
+ },
1327
+ {
1328
+ "type": "text",
1329
+ "text": "Image processing, computer vision, and graphics. Following the Perona-Malik scheme [65], PDE-based approaches flourished in the 1990s with multiple versions of anisotropic [90] and nonEuclidean [82] diffusion used primarily for image denoising. The realisation of these ideas in the form of nonlinear filters [85, 12] was adopted in the industry. PDE-based methods were also used for low-level vision tasks including inpainting [6] and image segmentation [13, 18]. In computer graphics, fundamental solutions (‘heat kernels’) of diffusion equations were used as shape descriptors [83]. The closed-form expression of such solutions using the Laplace-Beltrami operator served as inspiration for some of the early approaches for GNNs [34, 23, 42, 45]. ",
1330
+ "bbox": [
1331
+ 174,
1332
+ 712,
1333
+ 825,
1334
+ 824
1335
+ ],
1336
+ "page_idx": 8
1337
+ },
1338
+ {
1339
+ "type": "text",
1340
+ "text": "Neural differential equations The interpretation of neural networks as discretised differential equations (‘neural ODEs’) [19] was an influential recent result with multiple follow-up works [25, 28, 53, 46]. In deep learning on graphs, this mindset has been applied to GNN architectures [2, 67] and continuous message passing [94]. Continuous GNNs were also explored in [31] who, similarly to [76], addressed the solutions of fixed point equations. Ordinary Differential Equations ",
1341
+ "bbox": [
1342
+ 174,
1343
+ 842,
1344
+ 825,
1345
+ 911
1346
+ ],
1347
+ "page_idx": 8
1348
+ },
1349
+ {
1350
+ "type": "text",
1351
+ "text": "on Graph Networks (GODE)[98] approach the problem using the technique of invertible ResNets. \nFinally, [75] used graph-based ODEs to generate physics simulations. ",
1352
+ "bbox": [
1353
+ 171,
1354
+ 92,
1355
+ 825,
1356
+ 119
1357
+ ],
1358
+ "page_idx": 9
1359
+ },
1360
+ {
1361
+ "type": "text",
1362
+ "text": "Physics-inspired learning Solving PDEs with deep learning has been explored by [72]. Neural networks appeared in [47] to accelerate PDE solvers with applications in the physical sciences. These have been applied to problems where the PDE can be described on a graph [48]. [4] consider the problem of predicting fluid flow and use a PDE inside a GNN. These approaches differ from ours in that they solve a given PDE, whereas we use the notion of discretising PDEs as a principle to understand and design GNNs. ",
1363
+ "bbox": [
1364
+ 174,
1365
+ 135,
1366
+ 825,
1367
+ 217
1368
+ ],
1369
+ "page_idx": 9
1370
+ },
1371
+ {
1372
+ "type": "text",
1373
+ "text": "Neural ODEs on graphs The most similar work to this is GRAND [16] of which BLEND can be considered a non-Euclidean extension. In addition there are several other works that apply the neural ODE framework to graphs. In GDE [67], GODE [98] and CGNN [94], the goal is to adapt neural ordinary differential equations to graphs. In contrast, we consider non-Euclidean partial differential equations and regard GNNs as particular instances of their discretisation (both spatial and temporal). We can naturally use spaces with any metric, in particular, extending recent works on hyperbolic graph embeddings. None of the previous techniques explore the link to differential geometry. More specifically, GDE, GODE, and CGNN consider Neural ODEs of the canonical form $\\begin{array} { r } { \\frac { \\partial x } { \\partial t } = f ( x , t , \\theta ) } \\end{array}$ where $f$ is a graph neural network (GODE), the message passing component of a GNN (CGNN), or restricting $f$ to be layers of bijective functions on graphs (GDE). Furthermore, in CGNN only ODEs with closed form solutions are considered. [75] on the other hand is quite distinct as they are not concerned with GNN design and instead use graph-based ODEs to generate physics simulations. ",
1374
+ "bbox": [
1375
+ 173,
1376
+ 232,
1377
+ 825,
1378
+ 400
1379
+ ],
1380
+ "page_idx": 9
1381
+ },
1382
+ {
1383
+ "type": "text",
1384
+ "text": "7 Conclusion ",
1385
+ "text_level": 1,
1386
+ "bbox": [
1387
+ 174,
1388
+ 419,
1389
+ 297,
1390
+ 435
1391
+ ],
1392
+ "page_idx": 9
1393
+ },
1394
+ {
1395
+ "type": "text",
1396
+ "text": "We developed a new class of graph neural networks based on the discretisation of a non-Euclidean diffusion PDE called Beltrami flow. We represent the graph structure and node features as a manifold in a joint continuous space, whose evolution by a parametric diffusion PDE (driven by the downstream learning task) gives rise to feature learning, positional encoding, and possibly also graph rewiring. Our experimental results show very good performance on popular benchmarks with a small fraction of parameters used by other GNN models. Perhaps most importantly, our framework establishes important links between GNNs, differential geometry, and PDEs – fields with a rich history and profound results that in our opinion are still insufficiently explored in the graph ML community. ",
1397
+ "bbox": [
1398
+ 174,
1399
+ 450,
1400
+ 825,
1401
+ 561
1402
+ ],
1403
+ "page_idx": 9
1404
+ },
1405
+ {
1406
+ "type": "text",
1407
+ "text": "Future directions While we show that our framework generalises many popular GNN architectures we seek to use the graph as a numerical representation of an underlying continuous space. This view of the graph as an approximation of a continuous latent structure is a common paradigm of manifold learning [84, 5, 20] and network geometry [8]. If adopted in graph ML, this mindset offers a rigorous mathematical framework for formalising and generalising some of the recent trends in the field, including the departure from the input graph as the basis for message passing [1], latent graph inference [41, 89, 29, 36] higher-order [7, 54] and directional [3] message passing (which can be expressed as anisotropic diffusion arising from additional structure of the underlying continuous space and different discretisation of the PDEs e.g. based on finite elements), and exploiting efficient numerical PDE solvers [19]. We leave these exciting directions for future research. ",
1408
+ "bbox": [
1409
+ 174,
1410
+ 575,
1411
+ 825,
1412
+ 715
1413
+ ],
1414
+ "page_idx": 9
1415
+ },
1416
+ {
1417
+ "type": "text",
1418
+ "text": "Societal impact GNNs have recently become increasingly utilized in industrial applications e.g. in recommender systems and social networks, and hence could potentially lead to a negative societal impact if used improperly. We would like to emphasize that our paper does not study such potential negative applications and the mathematical framework we develop could help to interpret and understand existing GNN models. We believe that better understanding of ML models is key to managing their potential societal implications and preventing their negative impact. ",
1419
+ "bbox": [
1420
+ 174,
1421
+ 729,
1422
+ 825,
1423
+ 813
1424
+ ],
1425
+ "page_idx": 9
1426
+ },
1427
+ {
1428
+ "type": "text",
1429
+ "text": "Limitations The assumption that the graph can be modelled as a discretisation of some continuous space makes our framework applicable only to cases where the edge and node features are continuous in nature. Applications e.g. to knowledge graphs with categorical attributes could only be addressed by first embedding such attributes in a continuous space. Finally, the structural result presented in Theorem 1 that link our model to the Polyakov action have not been implemented. This will be addressed in future works. ",
1430
+ "bbox": [
1431
+ 174,
1432
+ 828,
1433
+ 823,
1434
+ 911
1435
+ ],
1436
+ "page_idx": 9
1437
+ },
1438
+ {
1439
+ "type": "text",
1440
+ "text": "8 Acknowledgements ",
1441
+ "text_level": 1,
1442
+ "bbox": [
1443
+ 174,
1444
+ 89,
1445
+ 367,
1446
+ 107
1447
+ ],
1448
+ "page_idx": 10
1449
+ },
1450
+ {
1451
+ "type": "text",
1452
+ "text": "We thank Nils Hammerla and Gabriele Corso for feedback on early version of this manuscript. MB and JR are supported in part by ERC Consolidator grant no 724228 (LEMAN). ",
1453
+ "bbox": [
1454
+ 173,
1455
+ 121,
1456
+ 823,
1457
+ 150
1458
+ ],
1459
+ "page_idx": 10
1460
+ },
1461
+ {
1462
+ "type": "text",
1463
+ "text": "References ",
1464
+ "text_level": 1,
1465
+ "bbox": [
1466
+ 174,
1467
+ 169,
1468
+ 266,
1469
+ 185
1470
+ ],
1471
+ "page_idx": 10
1472
+ },
1473
+ {
1474
+ "type": "text",
1475
+ "text": "[1] Uri Alon and Eran Yahav. On the bottleneck of graph neural networks and its practical implications. In ICLR, 2021. \n[2] Pedro H. C. Avelar, Anderson R. Tavares, Marco Gori, and Luis C. Lamb. Discrete and continuous deep residual learning over graphs. arXiv:1911.09554, 2019. [3] Dominique Beaini, Saro Passaro, Vincent Létourneau, William L Hamilton, Gabriele Corso, and Pietro Liò. Directional graph networks. arXiv:2010.02863, 2020. [4] Filipe de Avila Belbute-Peres, Thomas Economon, and Zico Kolter. Combining differentiable pde solvers and graph neural networks for fluid flow prediction. In ICML, 2020. [5] Mikhail Belkin and Partha Niyogi. Laplacian eigenmaps for dimensionality reduction and data representation. Neural Computation, 15(6):1373–1396, 2003. \n[6] Marcelo Bertalmio, Guillermo Sapiro, Vincent Caselles, and Coloma Ballester. Image inpainting. In PACMCGIT, 2000. [7] Cristian Bodnar, Fabrizio Frasca, Yu Guang Wang, Nina Otter, Guido Montúfar, Pietro Liò, and Michael Bronstein. Weisfeiler and lehman go topological: Message passing simplicial networks. arXiv:2103.03212, 2021. \n[8] Marian Boguna, Ivan Bonamassa, Manlio De Domenico, Shlomo Havlin, Dmitri Krioukov, and M Ángeles Serrano. Network geometry. Nature Reviews Physics, pages 1–22, 2021. \n[9] Giorgos Bouritsas, Fabrizio Frasca, Stefanos Zafeiriou, and Michael M Bronstein. Improving graph neural network expressivity via subgraph isomorphism counting. arXiv:2006.09252, 2020. \n[10] Michael M Bronstein, Joan Bruna, Taco Cohen, and Petar Velickovi ˇ c. Geometric deep learning: ´ Grids, groups, graphs, geodesics, and gauges. arXiv:2104.13478, 2021. \n[11] Michael M Bronstein and Iasonas Kokkinos. Scale-invariant heat kernel signatures for non-rigid shape recognition. In CVPR, 2010. \n[12] Antoni Buades, Bartomeu Coll, and J-M Morel. A non-local algorithm for image denoising. In ICCV, 2005. \n[13] Vicent Caselles, Ron Kimmel, and Guillermo Sapiro. Geodesic active contours. IJCV, 22(1):61–79, 1997. \n[14] Soumen Chakrabarti. Dynamic personalized pagerank in entity-relation graphs. In WWW, 2007. \n[15] Benjamin P. Chamberlain, James Clough, and Marc Peter Deisenroth. Neural embeddings of graphs in hyperbolic space. arXiv:1705.10359, 2017. \n[16] Benjamin P. Chamberlain, James Rowbottom, Maria Gorinova, Michael M. Bronstein, Stefan Webb, and Emanuele Rossi. GRAND: graph neural diffusion. In ICML 2021, pages 1407–1418, 2021. \n[17] Ines Chami, Rex Ying, Christopher Ré, and Jure Leskovec. Hyperbolic graph convolutional neural networks. In NeurIPS, 2019. \n[18] Tony F Chan and Luminita A Vese. Active contours without edges. IEEE Trans. Image Processing, 10(2):266–277, 2001. \n[19] Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David Duvenaud. Neural ordinary differential equations. In NeurIPS, pages 6571–6583, 2018. \n[20] Ronald R Coifman and Stéphane Lafon. Diffusion maps. Applied and computational harmonic analysis, 21(1):5–30, 2006. \n[21] Miles Cranmer, Sam Greydanus, Stephan Hoyer, Peter Battaglia, David Spergel, and Shirley Ho. Lagrangian neural networks. 2020. \n[22] Miles Cranmer, Rui Xu, Peter Battaglia, and Shirley Ho. Learning symbolic physics with graph networks. arXiv:1909.05862, 2019. \n[23] Michaël Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. In NeurIPS, 2016. \n[24] John R Dormand and Peter J Prince. A family of embedded Runge-Kutta formulae. Journal of Computational and Applied Mathematics, 6(1):19–26, 1980. \n[25] Emilien Dupont, Arnaud Doucet, and Yee Whye Teh. Augmented neural ODEs. In NeurIPS, pages 3134–3144, 2019. \n[26] Frédo Durand and Julie Dorsey. Fast bilateral filtering for the display of high-dynamic-range images. In Proc Computer Graphics and Interactive Techniques, 2002. \n[27] Vijay Prakash Dwivedi and Xavier Bresson. A generalization of transformer networks to graphs. arXiv:2012.09699, 2020. \n[28] Chris Finlay, Jörn-Henrik Jacobsen, Levon Nurbekyan, and Adam M Oberman. How to train your neural ode: The world of Jacobian and kinetic regularization. In ICML, 2020. \n[29] Luca Franceschi, Mathias Niepert, Massimiliano Pontil, and Xiao He. Learning discrete structures for graph neural networks. In ICML, 2019. \n[30] Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E. Dahl. Neural message passing for quantum chemistry. In ICML, 2017. \n[31] Fangda Gu, Heng Chang, Wenwu Zhu, Somayeh Sojoudi, and Laurent El Ghaoui. Implicit graph neural networks. In NeurIPS, 2020. \n[32] William L Hamilton, Rex Ying, and Jure Leskovec. Inductive representation learning on large graphs. In NeurIPS, 2017. \n[33] DF Hawken, JJ Gottlieb, and JS Hansen. Review of some adaptive node-movement techniques in finite-element and finite-difference solutions of partial differential equations. Journal of Computational Physics, 95(2):254–302, 1991. \n[34] Mikael Henaff, Joan Bruna, and Yann LeCun. Deep convolutional networks on graph-structured data. arXiv:1506.05163, 2015. \n[35] Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and Jure Leskovec. Open graph benchmark: Datasets for machine learning on graphs. arXiv:2005.00687, 2020. \n[36] Anees Kazi, Luca Cosmo, Nassir Navab, and Michael Bronstein. Differentiable graph module (DGN) graph convolutional networks. arXiv:2002.04999, 2020. \n[37] Jacob Kelly, Jesse Bettencourt, Matthew James Johnson, and David Duvenaud. Learning differential equations that are easy to solve. In NeurIPS, 2020. \n[38] Ron Kimmel. Numerical geometry of images: Theory, algorithms, and applications. Springer, 2012. \n[39] Ron Kimmel, Ravi Malladi, and Nir Sochen. Images as embedded maps and minimal surfaces: movies, color, texture, and volumetric medical images. IJCV, 39(2):111–129, 2000. \n[40] Ron Kimmel, Nir Sochen, and Ravi Malladi. From high energy physics to low level vision. pages 236–247. Springer, 1997. \n[41] Thomas Kipf, Ethan Fetaya, Kuan-Chieh Wang, Max Welling, and Richard Zemel. Neural relational inference for interacting systems. In ICML, 2018. \n[42] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In ICLR, 2017. \n[43] Johannes Klicpera, Stefan Weißenberger, and Stephan Günnemann. Diffusion improves graph learning. In NeurIPS, volume 32, 2019. \n[44] Wilhelm Kutta. Beitrag zur naherungsweisen integration totaler differentialgleichungen. Z. Math. Phys., 46:435–453, 1901. \n[45] Ron Levie, Federico Monti, Xavier Bresson, and Michael M Bronstein. Cayleynets: Graph convolutional neural networks with complex rational spectral filters. IEEE Trans. Signal Processing, 2017. \n[46] Xuechen Li, Ricky Tian Qi Chen, Ting-Kam Leonard Wong, and David Duvenaud. Scalable gradients for stochastic differential equations. In Artificial Intelligence and Statistics, 2020. \n[47] Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. 2020. \n[48] Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Multipole graph neural operator for parametric partial differential equations. In NeurIPS, 2020. \n[49] Renjie Liao, Zhizhen Zhao, Raquel Urtasun, and Richard S Zemel. Lanczosnet: Multi-scale deep graph convo-lutional networks. In 7th International Conference on Learning Representations, ICLR 2019, 2019. \n[50] Richard Liaw, Eric Liang, Robert Nishihara, Philipp Moritz, Joseph E. Gonzalez, and Ion Stoica. Tune: A research platform for distributed model selection and training. In arXiv: 1807.05118., 2018. \n[51] Roee Litman and Alexander M Bronstein. Learning spectral descriptors for deformable shape correspondence. PAMI, 36(1):171–180, 2013. \n[52] Qi Liu, Maximilian Nickel, and Douwe Kiela. Hyperbolic graph neural networks. arXiv preprint arXiv:1910.12892, 2019. \n[53] Xuanqing Liu, Tesi Xiao, Si Si, Qin Cao, Sanjiv Kumar, and Cho-Jui Hsieh. Neural SDE: Stabilizing neural ODE networks with stochastic noise. (2), 2019. \n[54] Haggai Maron, Heli Ben-Hamu, Hadar Serviansky, and Yaron Lipman. Provably powerful graph networks. In NeurIPS, pages 2153–2164, 2019. \n[55] Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton Van Den Hengel. Image-based recommendations on styles and substitutes. In Proceedings Information Retrieval, 2015. \n[56] Andrew Kachites McCallum, Kamal Nigam, Jason Rennie, and Kristie Seymore. Automating the construction of internet portals with machine learning. Information Retrieval, 3(2):127–163, 2000. \n[57] Federico Monti, Davide Boscaini, Jonathan Masci, Emanuele Rodolà, Jan Svoboda, and Michael M. Bronstein. Geometric deep learning on graphs and manifolds using mixture model CNNs. In CVPR, 2017. \n[58] Galileo Namata, Ben London, Lise Getoor, Bert Huang, and UMD EDU. Query-driven active surveying for collective classification. In Proceedings Mining and Learning with Graphs, 2012. \n[59] Maximillian Nickel and Douwe Kiela. Learning continuous hierarchies in the lorentz model of hyperbolic geometry. In International Conference on Machine Learning, pages 3779–3788. PMLR, 2018. \n[60] Hoang NT and Takanori Maehara. Revisiting graph neural networks: All we have is low-pass filters. 2019. \n[61] Lars Onsager. Reciprocal relations in irreversible processes. i. Physical Review, 37(4):405, 1931. \n[62] Kenta Oono and Taiji Suzuki. Graph neural networks exponentially lose expressive power for node classification. In ICLR, 2020. \n[63] Lawrence Page, Sergey Brin, Rajeev Motwani, and Terry Winograd. The pagerank citation ranking: Bringing order to the web. Technical report, 1999. \n[64] Giuseppe Patané. Laplacian spectral kernels and distances for geometry processing and shape analysis. Computer Graphics Forum, 35(2):599–624, 2016. \n[65] Pietro Perona and Jitendra Malik. Scale-space and edge detection using anisotropic diffusion. PAMI, 12(7):629–639, 1990. \n[66] Bryan Perozzi, Rami Al-Rfou, and Steven Skiena. Deepwalk: Online learning of social representations. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 701–710, 2014. \n[67] Michael Poli, Stefano Massaroli, Junyoung Park, Atsushi Yamashita, Hajime Asama, and Jinkyoo Park. Graph neural ordinary differential equations. pages 6571–6583, 2019. \n[68] Alexander M Polyakov. Quantum geometry of bosonic strings. Physics Letters B, 103(3):207–210, 1981. \n[69] Lev Semenovich Pontryagin. Mathematical theory of optimal processes. Routledge, 2018. \n[70] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. \n[71] Alejandro F. Queiruga, N. Benjamin Erichson, Dane Taylor, and Michael W. Mahoney. Continuous-in-depth neural networks. CoRR, 2020. \n[72] Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Physics informed deep learning (Part II): Data-driven discovery of nonlinear partial differential equations. CoRR, (Part II), 2017. \n[73] Emanuele Rossi, Fabrizio Frasca, Ben Chamberlain, Davide Eynard, Michael Bronstein, and Federico Monti. SIGN: Scalable inception graph neural networks. arXiv:2004.11198, 2020. \n[74] Carl Runge. Über die numerische auflösung von differentialgleichungen. Mathematische Annalen, 46(2):167–178, 1895. \n[75] Alvaro Sanchez-Gonzalez, Victor Bapst, Kyle Cranmer, and Peter Battaglia. Hamiltonian graph networks with ODE integrators. 2019. \n[76] Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE Trans. Neural Networks, 27(8):61–80, 2009. \n[77] Prithviraj Sen, Galileo Namata, Mustafa Bilgic, Lise Getoor, Brian Galligher, and Tina EliassiRad. Collective classification in network data. AI Magazine, 29(3):93–93, 2008. \n[78] Lawrence F Shampine. Some practical Runge-Kutta formulas. Mathematics of Computation, 46(173):135–150, 1986. \n[79] Oleksandr Shchur, Maximilian Mumme, Aleksandar Bojchevski, and Stephan Günnemann. Pitfalls of graph neural network evaluation. arXiv:1811.05868, 2018. \n[80] Nino Shervashidze, Pascal Schweitzer, Erik Jan Van Leeuwen, Kurt Mehlhorn, and Karsten M Borgwardt. Weisfeiler-lehman graph kernels. JMLR, 12:2539–2561, 2011. \n[81] Jonathan Shlomi, Peter Battaglia, and Jean-Roch Vlimant. Graph neural networks in particle physics. Machine Learning: Science and Technology, 2(2):021001, 2020. \n[82] Nir Sochen, Ron Kimmel, and Ravi Malladi. A general framework for low level vision. IEEE Trans. Image Processing, 7(3):310–318, 1998. \n[83] Jian Sun, Maks Ovsjanikov, and Leonidas Guibas. A concise and provably informative multiscale signature based on heat diffusion. Computer Graphics Forum, 28(5):1383–1392, 2009. \n[84] Joshua B Tenenbaum, Vin De Silva, and John C Langford. A global geometric framework for nonlinear dimensionality reduction. Dcience, 290(5500):2319–2323, 2000. \n[85] Carlo Tomasi and Roberto Manduchi. Bilateral filtering for gray and color images. In ICCV, 1998. \n[86] Ashish Vaswani, Noam Shazeer, Niki Parmar, Akob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998–6008, 2017. \n[87] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. 2017. \n[88] Petar Velickovi ˇ c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua ´ Bengio. Graph attention networks. In ICLR, 2018. \n[89] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph CNN for learning on point clouds. ACM Trans. Graphics, 38(5):1–12, 2019. \n[90] Joachim Weickert. Anisotropic diffusion in image processing. Teubner Stuttgart, 1998. \n[91] Joachim Weickert, Bart M. Ter Haar Romeny, and Max A. Viergever. Efficient and reliable schemes for nonlinear diffusion filtering. IEEE Trans. Image Processing, 7(3):398–410, 1998. \n[92] Felix Wu, Amauri Souza, Tianyi Zhang, Christopher Fifty, Tao Yu, and Kilian Weinberger. Simplifying graph convolutional networks. In ICML. PMLR, 2019. \n[93] Tailin Wu, Hongyu Ren, Pan Li, and Jure Leskovec. Graph information bottleneck. NeurIPS, 2020. \n[94] Louis-pascal A C Xhonneux, Meng Qu, and Jian Tang. Continuous graph neural networks. In ICML, 2020. \n[95] Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? In ICLR, 2019. \n[96] Zhilin Yang, William Cohen, and Ruslan Salakhudinov. Revisiting semi-supervised learning with graph embeddings. In International conference on machine learning, pages 40–48. PMLR, 2016. \n[97] Manzil Zaheer, Satwik Kottur, Siamak Ravanbakhsh, Barnabas Poczos, Russ R Salakhutdinov, and Alexander J Smola. Deep sets. In NeurIPS, 2017. \n[98] Juntang Zhuang, Nicha Dvornek, Xiaoxiao Li, and James S. Duncan. Ordinary differential equations on graph networks. Technical Report 1, 2020. ",
1476
+ "bbox": [
1477
+ 174,
1478
+ 184,
1479
+ 828,
1480
+ 920
1481
+ ],
1482
+ "page_idx": 10
1483
+ },
1484
+ {
1485
+ "type": "text",
1486
+ "text": "",
1487
+ "bbox": [
1488
+ 171,
1489
+ 53,
1490
+ 828,
1491
+ 920
1492
+ ],
1493
+ "page_idx": 11
1494
+ },
1495
+ {
1496
+ "type": "text",
1497
+ "text": "",
1498
+ "bbox": [
1499
+ 171,
1500
+ 61,
1501
+ 828,
1502
+ 919
1503
+ ],
1504
+ "page_idx": 12
1505
+ },
1506
+ {
1507
+ "type": "text",
1508
+ "text": "",
1509
+ "bbox": [
1510
+ 171,
1511
+ 83,
1512
+ 828,
1513
+ 915
1514
+ ],
1515
+ "page_idx": 13
1516
+ },
1517
+ {
1518
+ "type": "text",
1519
+ "text": "",
1520
+ "bbox": [
1521
+ 171,
1522
+ 90,
1523
+ 828,
1524
+ 794
1525
+ ],
1526
+ "page_idx": 14
1527
+ },
1528
+ {
1529
+ "type": "text",
1530
+ "text": "Checklist ",
1531
+ "text_level": 1,
1532
+ "bbox": [
1533
+ 174,
1534
+ 89,
1535
+ 254,
1536
+ 106
1537
+ ],
1538
+ "page_idx": 15
1539
+ },
1540
+ {
1541
+ "type": "text",
1542
+ "text": "1. For all authors... ",
1543
+ "bbox": [
1544
+ 214,
1545
+ 116,
1546
+ 339,
1547
+ 130
1548
+ ],
1549
+ "page_idx": 15
1550
+ },
1551
+ {
1552
+ "type": "text",
1553
+ "text": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes] \n(b) Did you describe the limitations of your work? [Yes] In limitations (S6) and an extensions (S3.4) \n(c) Did you discuss any potential negative societal impacts of your work? [Yes] Section 6 \n(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes] ",
1554
+ "bbox": [
1555
+ 238,
1556
+ 135,
1557
+ 825,
1558
+ 239
1559
+ ],
1560
+ "page_idx": 15
1561
+ },
1562
+ {
1563
+ "type": "text",
1564
+ "text": "2. If you are including theoretical results... ",
1565
+ "bbox": [
1566
+ 214,
1567
+ 243,
1568
+ 493,
1569
+ 258
1570
+ ],
1571
+ "page_idx": 15
1572
+ },
1573
+ {
1574
+ "type": "text",
1575
+ "text": "(a) Did you state the full set of assumptions of all theoretical results? Theorem 1 is dealt with in the supplementary material [Yes] Supplementary Material \n(b) Did you include complete proofs of all theoretical results? [Yes] Supplementary Material ",
1576
+ "bbox": [
1577
+ 238,
1578
+ 263,
1579
+ 825,
1580
+ 320
1581
+ ],
1582
+ "page_idx": 15
1583
+ },
1584
+ {
1585
+ "type": "text",
1586
+ "text": "3. If you ran experiments... ",
1587
+ "bbox": [
1588
+ 212,
1589
+ 325,
1590
+ 393,
1591
+ 340
1592
+ ],
1593
+ "page_idx": 15
1594
+ },
1595
+ {
1596
+ "type": "text",
1597
+ "text": "(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] In supplementary material \n(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] hyperparameters are included in Supplementary Material and splits are addressed in section 5 \n(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [Yes] Results in Section 5 include error bars \n(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] AWS, details at the end of 5.1 and supplementary material ",
1598
+ "bbox": [
1599
+ 238,
1600
+ 344,
1601
+ 825,
1602
+ 503
1603
+ ],
1604
+ "page_idx": 15
1605
+ },
1606
+ {
1607
+ "type": "text",
1608
+ "text": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets... ",
1609
+ "bbox": [
1610
+ 210,
1611
+ 508,
1612
+ 823,
1613
+ 523
1614
+ ],
1615
+ "page_idx": 15
1616
+ },
1617
+ {
1618
+ "type": "text",
1619
+ "text": "(a) If your work uses existing assets, did you cite the creators? [Yes] papers have been cited for the most relevant OS libraries \n(b) Did you mention the license of the assets? [No] \n(c) Did you include any new assets either in the supplemental material or as a URL? [N/A] \n(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [N/A] \n(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [N/A] ",
1620
+ "bbox": [
1621
+ 235,
1622
+ 526,
1623
+ 826,
1624
+ 662
1625
+ ],
1626
+ "page_idx": 15
1627
+ },
1628
+ {
1629
+ "type": "text",
1630
+ "text": "5. If you used crowdsourcing or conducted research with human subjects... ",
1631
+ "bbox": [
1632
+ 215,
1633
+ 666,
1634
+ 705,
1635
+ 681
1636
+ ],
1637
+ "page_idx": 15
1638
+ },
1639
+ {
1640
+ "type": "text",
1641
+ "text": "(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [N/A] \n(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A] \n(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [N/A] ",
1642
+ "bbox": [
1643
+ 238,
1644
+ 685,
1645
+ 825,
1646
+ 773
1647
+ ],
1648
+ "page_idx": 15
1649
+ }
1650
+ ]
parse/train/4YlE2huxEsl/4YlE2huxEsl_middle.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/4YlE2huxEsl/4YlE2huxEsl_model.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/8qDwejCuCN/8qDwejCuCN.md ADDED
@@ -0,0 +1,314 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # UNSUPERVISED REPRESENTATION LEARNING FOR TIME SERIES WITH TEMPORAL NEIGHBORHOOD CODING
2
+
3
+ Sana Tonekaboni∗ University of Toronto & Vector Institute The Hospital for Sick Children stonekaboni@cs.toronto.edu
4
+
5
+ Danny Eytan The Hospital for Sick Children biliary.colic@gmail.com
6
+
7
+ Anna Goldengerg University of Toronto & Vector Institute The Hospital for Sick Children anna.goldenberg@utoronto.ca
8
+
9
+ # ABSTRACT
10
+
11
+ Time series are often complex and rich in information but sparsely labeled and therefore challenging to model. In this paper, we propose a self-supervised framework for learning generalizable representations for non-stationary time series. Our approach, called Temporal Neighborhood Coding (TNC), takes advantage of the local smoothness of a signal’s generative process to define neighborhoods in time with stationary properties. Using a debiased contrastive objective, our framework learns time series representations by ensuring that in the encoding space, the distribution of signals from within a neighborhood is distinguishable from the distribution of non-neighboring signals. Our motivation stems from the medical field, where the ability to model the dynamic nature of time series data is especially valuable for identifying, tracking, and predicting the underlying patients’ latent states in settings where labeling data is practically impossible. We compare our method to recently developed unsupervised representation learning approaches and demonstrate superior performance on clustering and classification tasks for multiple datasets.
12
+
13
+ # 1 INTRODUCTION
14
+
15
+ Real-world time-series data is high dimensional, complex, and has unique properties that bring about many challenges for data modeling (Yang & Wu, 2006). In addition, these signals are often sparsely labeled, making it even more challenging for supervised learning tasks. Unsupervised representation learning can extract informative low-dimensional representations from raw time series by leveraging the data’s inherent structure, without the need for explicit supervision. These representations are more generalizable and robust, as they are less specialized for solving a single supervised task. Unsupervised representation learning is well studied in domains such as vision (Donahue & Simonyan, 2019; Denton et al., 2017; Radford et al., 2015) and natural language processing (Radford et al., 2017; Young et al., 2018; Mikolov et al., 2013), but has been underexplored in the literature for time series settings. Frameworks designed for time series need to be efficient and scalable because signals encountered in practice can be long, high dimensional, and high frequency. Moreover, it should account for and be able to model dynamic changes that occur within samples, i.e., non-stationarity of signals.
16
+
17
+ The ability to model the dynamic nature of time series data is especially valuable in medicine. Health care data is often organized as a time series, with multiple data types, collected from various sources at different sampling frequencies, and riddled with artifacts and missing values. Throughout their stay at the hospital or within the disease progression period, patients transition gradually between distinct clinical states, with periods of relative stability, improvement, or unexpected deterioration, requiring escalation of care that alters the patient’s trajectory. A particular challenge in medical time-series data is the lack of well-defined or available labels that are needed for identifying the underlying clinical state of an individual or for training models aimed at extracting low-dimensional representations of these states. For instance, in the context of critical-care, a patient’s stay in the critical care unit (CCU) is captured continuously via streaming physiological signals by the bedside monitor. Obtaining labels for the patient’s state for extended periods of these signals is practically impossible as the underlying physiological state can be unknown even to the clinicians. This further motivates the use of unsupervised representation learning in these contexts. Learning rich representations can be crucial in facilitating the tracking of disease progression, predicting the future trajectories of the patients, and tailoring treatments to these underlying states.
18
+
19
+ In this paper, we propose a self-supervised framework for learning representations for complex multivariate non-stationary time series. This approach, called Temporal Neighborhood Coding (TNC), is designed for temporal settings where the latent distribution of the signals changes over time, and it aims to capture the progression of the underlying temporal dynamics. TNC is efficient, easily scalable to high dimensions, and can be used in different time series settings. We assess the quality of the learned representations on multiple datasets and show that the representations are general and transferable to many downstream tasks such as classification and clustering. We further demonstrate that our method outperforms existing approaches for unsupervised representation learning, and it even performs closely to supervised techniques in classification tasks. The contributions of this work are three-fold:
20
+
21
+ 1. We present a novel neighborhood-based unsupervised learning framework for non-stationary multivariate time series data.
22
+ 2. We introduce the concept of a temporal neighborhood with stationary properties as the distribution of similar windows in time. The neighborhood boundaries are determined automatically using the properties of the signal and statistical testing.
23
+ 3. We incorporate concepts from Positive Unlabeled Learning, specifically, sample weight adjustment, to account for potential bias introduced in sampling negative examples for the contrastive loss.
24
+
25
+ # 2 METHOD
26
+
27
+ We introduce a framework for learning representations that encode the underlying state of a multivariate, non-stationary time series. Our self-supervised approach, TNC, takes advantage of the local smoothness of the generative process of signals to learn generalizable representations for windows of time series. This is done by ensuring that in the representation space, the distribution of signals proximal in time is distinguishable from the distribution of signals far away, i.e., proximity in time is identifiable in the encoding space. We represent our multivariate time series signals as $\dot { X } \in { \cal R } ^ { D \times T }$ , where D is the number of features and T is the number of measurements over time. X[t− δ ,t+ δ ] represents a window of time series of length $\delta$ , centered around time $t$ , that includes measurements of all features taken in the interval $\begin{array} { r } { [ t - \frac { \delta } { 2 } , t + \frac { \delta } { 2 } ] } \end{array}$ . Throughout the paper, we refer to this window as $W _ { t }$ for notational simplicity. Our goal is to learn the underlying representation of $W _ { t }$ , and by sliding this window over time, we can obtain the trajectory of the underlying states of the signal.
28
+
29
+ We define the temporal neighborhood $( N _ { t } )$ of a window $W _ { t }$ as the set of all windows with centroids $t ^ { * }$ , sampled from a normal distribution $t ^ { * } \sim \mathcal { N } ( t , \eta \cdot \delta )$ . Where $\mathcal { N }$ is a Gaussian centered at $t , \delta$ is the size of the window, and $\eta$ is the parameter that defines the range of the neighborhood. Relying on the local smoothness of a signal’s generative process, the neighborhood distribution is characterized as a Gaussian to model the gradual transition in temporal data, and intuitively, it approximates the distribution of samples that are similar to $W _ { t }$ . The $\eta$ parameter determines the neighborhood range and depends on the signal characteristics and how gradual the time series’s statistical properties change over time. This can be set by domain experts based on prior knowledge of the signal behavior, or for more robust estimation, it can be determined by analyzing the stationarity properties of the signal for every $W _ { t }$ . Since the neighborhood represents similar samples, the range should identify the approximate time span within which the signal remains stationary, and the generative process does not change. For this purpose, we use the Augmented Dickey-Fuller (ADF) statistical test to determine this region for every window. Proper estimation of the neighborhood range is an integral part of the TNC framework. If $\eta$ is too small, many samples from within a neighborhood will overlap, and therefore the encoder would only learn to encode the overlapping information. On the other hand, if $\eta$ is too big, the neighborhood would span over multiple underlying states, and therefore the encoder would fail to distinguish the variation among these states. Using the ADF test, we can automatically adjust the neighborhood for every window based on the signal behavior. More details on this test and how it is used to estimate $\eta$ is described in section 2.
30
+
31
+ Now, assuming windows within a neighborhood possess similar properties, signals outside of this neighborhood, denoted as $\bar { N } _ { t }$ , are considered non-neighboring windows. Samples from $\bar { N } _ { t }$ are likely to be different from $W _ { t }$ , and can be considered as negative samples in a context of a contrastive learning framework. However, this assumption can suffer from the problem of sampling bias, common in most contrastive learning approaches (Chuang et al., 2020; Saunshi et al., 2019). This bias occurs because randomly drawing negative examples from the data distribution may result in negative samples that are actually similar to the reference. This can significantly impact the learning framework’s performance, but little work has been done on addressing this issue (Chuang et al., 2020). In our context, this can happen when there are windows from $\bar { N _ { t } }$ that are far away from $W _ { t }$ , but have the same underlying state. To alleviate this bias in the TNC framework, we consider samples from $\bar { N } _ { t }$ as unlabeled samples, as opposed to negative, and use ideas from Positive-Unlabeled (PU) learning to accurately measure the loss function. In reality, even though samples within a neighborhood are all similar, we cannot make the assumption that samples outside this region are necessarily different. For instance, in the presence of long term seasonalities, signals can exhibit similar properties at distant times. In a healthcare context, this can be like a stable patient that undergoes a critical condition, but returns back to a stable state afterwards.
32
+
33
+ In PU learning, a classifier is learned using labeled data drawn from the positive class $( P )$ and unlabeled data $( U )$ that is a mixture of positive and negative samples with a positive class prior $\pi$ (Du Plessis et al., 2014; Kiryo et al., 2017; Du Plessis & Sugiyama, 2014). Existing PU learning methods fall under two categories based on how they handle the unlabeled data: 1) methods that identify negative samples from the unlabeled cohort (Li & Liu, 2003); 2) methods that treat the unlabeled data as negative samples with smaller weights (Lee & Liu, 2003; Elkan & Noto, 2008). In the second category, unlabeled samples should be properly weighted in the loss term in order to train an unbiased classifier. Elkan & Noto (2008) introduces a simple and efficient way of approximating the expectation of a loss function by assigning individual weights $w$ to training examples from the unlabeled cohort. This means each sample from the neighborhood is treated as a positive example with unit weight, while each sample from $\bar { N }$ is treated as a combination of a positive example with weight $w$ and a negative example with complementary weight ${ 1 - w }$ . In the original paper (Elkan & Noto, 2008), the weight is defined as the probability for a sample from the unlabeled set to be a positive sample, i.e. $w = p ( y = 1 | x )$ for $x \in U$ . In the TNC framework, this weight represents the probability of having samples similar to $W _ { t }$ in $\bar { N }$ . By incorporating weight adjustment into the TNC loss (Equation 1), we account for possible positive samples that occur in the non-neighboring distribution. $w$ can be approximated using prior knowledge of the underlying state distribution or tuned as a hyperparameter. Appendix A.6 explains how the weight parameter is selected for our different experiment setups and also demonstrates the impact of weight adjustment on performance for downstream tasks.
34
+
35
+ After defining the neighborhood distribution, we train an objective function that encourages a distinction between the representation of samples of the same neighborhood from the outside samples. An ideal encoder preserves the neighborhood properties in the encoding space. Therefore representations $Z _ { l } = E n c ( \bar { W _ { l } } )$ of samples from a neighborhood $W _ { l } \in N _ { t }$ , can be distinguished from representation $Z _ { k } = E n c ( W _ { k } )$ of samples from outside the neighborhood $W _ { k } \in \bar { N } _ { t }$ . TNC is composed of two main components:
36
+
37
+ 1. An Encoder $E n c ( W _ { t } )$ that maps $W _ { t } \in \mathbb { R } ^ { D \times \delta }$ to a representation $Z _ { t } \in \mathbb { R } ^ { M }$ , in a lower dimensional space $( M \ll D \times \delta )$ , where $D \times \delta$ is the total number of measurements in $W _ { t }$
38
+
39
+ 2. A Discriminator $\mathcal { D } ( Z _ { t } , Z )$ that approximates the probability of $Z$ being the representation of a window in $N _ { t }$ . More specifically, it receives two samples from the encoding space and predicts the probability of those samples belonging to the same temporal neighborhood.
40
+
41
+ ![](images/398f866909f9701d5c0a6c7583ec244c49b4701a0c9562b08987de6f3e9de2a0.jpg)
42
+ Figure 1: Overview of the TNC framework components. For each sample window $W _ { t }$ (indicated with the dashed black box), we first define the neighborhood distribution. The encoder learns the distribution of windows sampled from $N _ { t }$ and $\bar { N } _ { t }$ , in the representation space. Then samples from Weightingthese distributions are fed into the discriminator alongside $Z _ { t }$ , to predict the probability of the Representationwindows being in the same neighborhood.
43
+
44
+ TNC is a general framework; therefore, it is agnostic to the nature of the time series and the architecture of the encoder. The encoder can be any parametric model that is well-suited to the signal properties (Oord et al., 2016; Bai et al., 2018; Fawaz et al., 2019). For the Discriminator $\mathcal { D } ( Z _ { t } , Z )$ we use a simple multi-headed binary classifier that outputs 1 if $Z$ and $Z _ { t }$ are representations of neighbors in time, and 0 otherwise. In the experiment section, we describe the architectural details of the models used for our experiments in more depth.
45
+
46
+ Figure 1 provides a summary overview of the TNC framework. We formalize the objective function of our unsupervised learning framework in Equation 1. In essence, we would like the probability likelihood estimation of the Discriminator to be accurate, i.e., close to 1 for the representation of neighboring samples and close to 0 for windows far apart. Samples from the non-neighboring region $( \bar { N } )$ are weight-adjusted using the $w$ parameters to account for positive samples in this distribution.
47
+
48
+ $$
49
+ \begin{array} { r } { \boldsymbol { \Sigma } = - \mathbb { E } _ { W _ { t } \sim X } [ \mathbb { E } _ { W _ { t } \sim N _ { t } } [ \log \underbrace { \mathcal { D } ( Z _ { t } , Z _ { l } ) } _ { \mathrm { ~ } \forall \mathrm { ~ } W _ { t } \sim \bar { N } _ { t } } ] + \mathbb { E } _ { W _ { t } \sim \bar { N } _ { t } } [ ( 1 - w _ { t } ) \times \log \underbrace { ( 1 - \mathcal { D } ( Z _ { t } , Z _ { k } ) ) } _ { \mathcal { D } ( E n c ( W _ { t } ) , E n c ( W _ { k } ) ) } + w _ { t } \times \log \mathcal { D } ( Z _ { t } , Z _ { k } ) ] } \\ { \mathcal { D } ( E n c ( W _ { t } ) , E n c ( W _ { l } ) ) \quad \quad \quad \quad { \mathcal { D } } ( E n c ( W _ { t } ) , E n c ( W _ { k } ) ) } \end{array}
50
+ $$
51
+
52
+ We train the encoder and the discriminator hand in hand by optimizing for this objective. Note that the Discriminator is only part of training and will not be used during inference. Similar to the encoder, it can be approximated using any parametric model. However, the more complex the Discriminator, the harder it becomes to interpret the latent space’s decision boundaries since it allows similarities to be mapped on complex nonlinear relationships.
53
+
54
+ Defining the neighborhood parameter using the ADF test: As mentioned earlier, the neighborhood range can be specified using the characteristics of the data. In non-stationary time series, the generative process of the signals changes over time. We define the temporal neighborhood around every window as the region where the signal is relatively stationary. Since a signal may remain in an underlying state for an unknown amount of time, each window’s neighborhood range may vary in size and must be adjusted to signal behavior. To that end, we use the Augmented Dickey-Fuller (ADF) statistical test to derive the neighborhood range $\eta$ . The ADF test belongs to a category of tests called "Unit Root Test", and is a method for testing the stationarity of a time series. For every $W _ { t }$ , we want to find the neighborhood range around that window that indicates a stationary region. To determine this, we start from $\eta = 1$ and gradually increase the neighborhood size $\eta$ , measuring the $p$ -value from the test at every step. Once $p$ -value is above a threshold (in our setting 0.01), it means that it fails to reject the null hypothesis and suggests that within this neighborhood region, the signal is no longer stationary. This way, we find the widest neighborhood within which the signal remains relatively stationary. Note that the window size $\delta$ is constant throughout the experiment, and during ADF testing, we only adjust the neighborhood’s width.
55
+
56
+ # 3 EXPERIMENTS
57
+
58
+ We evaluate our framework’s usability on multiple time series datasets with dynamic latent states that change over time. We compare classification performance and clusterability against two state-ofthe-art approaches for unsupervised representation learning for time series: 1. Contrastive Predictive Coding (CPC) (Oord et al., 2018) that uses predictive coding principles to train the encoder on a probabilistic contrastive loss. 2. Triplet-Loss (T-Loss), introduced in (Franceschi et al., 2019), which employs time-based negative sampling and a triplet loss to learn representations for time series windows. The triplet loss objective ensures similar time series have similar representations by minimizing the pairwise distance between positive samples (subseries) while maximizing it for negative ones. (See Appendix A.2 for more details on each baseline.)
59
+
60
+ For a fair comparison and to ensure that the difference in performance is not due to the differences in the models’ architecture, the same encoder network is used across all compared baselines. Our objective is to compare the performance of the learning frameworks, agnostic to the encoder’s choice. Therefore, we selected simple architectures to evaluate how each framework can use a simple encoder’s limited capacity to learn meaningful representations. We assess the generalizability of the representations by 1) evaluating clusterability in the encoding space and 2) using the representations for a downstream classification task. In addition to the baselines mentioned above, we also compare clusterability performance with unsupervised K-means and classification with a K-Nearest Neighbor classifier, using Dynamic Time Warping (DTW) to measure time series distance. All models are implemented using Pytorch 1.3.1 and trained on a machine with Quadro 400 GPU 1. Below we describe the datasets for our experiments in more detail.
61
+
62
+ # 3.1 SIMULATED DATA
63
+
64
+ The simulated dataset is designed to replicate very long, non-stationary, and high-frequency time series for which the underlying dynamics change over time. Our generated time series consists of 2000 measurements for 3 features, generated from 4 different underlying states. We use a Hidden Markov Model (HMM) to generate the random latent states over time, and in each state, the time series is generated from a different generative process, including Gaussian Processes (GPs) with different kernel functions and Nonlinear Auto-regressive Moving Average models with different sets of parameters $\overset { \cdot } { \alpha }$ and $\beta$ ). Besides, for it to further resemble realistic (e.g., clinical) time series, two features are always correlated. More details about this dataset are provided in the Appendix A.1. For this experimental setup, we use a two-directional, single-layer recurrent neural network encoder. We have selected this simple architecture because it handles time series with variable lengths, and it easily extends to higher-dimensional inputs. The encoder model encodes multi-dimensional signal windows of $\delta = 5 0$ into 10 dimensional representation vectors. The window size is selected such that it is long enough to contain information about the underlying state but not too long to span over multiple underlying states. A more detailed discussion on the window size choice is presented in Appendix A.4.
65
+
66
+ # 3.2 CLINICAL WAVEFORM DATA
67
+
68
+ For a real-world clinical experiment, we use the MIT-BIH Atrial Fibrillation dataset (Moody, 1983). This dataset includes 25 long-term Electrocardiogram (ECG) recordings (10 hours in duration) of human subjects with atrial fibrillation. It consists of two ECG signals; each sampled at $2 5 0 \mathrm { H z }$ . The signals are annotated over time for the following different rhythm types: 1) Atrial fibrillation, 2) Atrial flutter, 3) AV junctional rhythm, and 4) all other rhythms. Our goal in this experiment is to identify the underlying type of arrhythmia for each sample without any information about the labels. This dataset is particularly interesting and makes this experiment challenging due to the following special properties:
69
+
70
+ • The underlying heart rhythm changes over time in each sample. This is an opportunity to evaluate how different representation learning frameworks can handle alternating classes in non-stationary settings;
71
+
72
+ • The dataset is highly imbalanced, with atrial flutter and AV junctional rhythm being present in fewer than $0 . 1 \%$ of the measurements. Data imbalance poses many challenges for downstream classification, further motivating the use of unsupervised representation learning; The dataset has samples from a small number of individuals, but over an extended period (around 5 million data points). This realistic scenario, common in healthcare data, shows that our framework is still powerful in settings with a limited number of samples.
73
+
74
+ The simple RNN encoder architecture used for other experiment setups cannot model the highfrequency ECG measurements. Therefore, inspired by state-of-the-art architectures for ECG classification problems, the encoder Enc used in this experiment is a 2-channel, 1-dimensional strided convolutional neural network that runs directly on the ECG waveforms. We use six convolutional layers with a total down-sampling factor of 16. The window size is 2500 samples, meaning that each convolutional filter covers at least half a second of ECG recording, and the representations are summarized in a 64-dimensional vector.
75
+
76
+ # 3.3 HUMAN ACTIVITY RECOGNITION (HAR) DATA
77
+
78
+ Human Activity Recognition (HAR) is the problem of predicting the type of activity using temporal data from accelerometer and gyroscope measurements. We use the HAR dataset from the UCI Machine Learning Repository 2 that includes data collected from 30 individuals using a smartwatch. Each person performs six activities: 1) walking, 2) walking upstairs, 3) walking downstairs, 4) sitting, 5) standing, and 6) laying. The time-series measurements are pre-processed to extract 561 features. For our purpose, we concatenate the activity samples from every individual over time using the subject identifier to build the full-time series for each subject, which includes continuous activity change. Similar to the simulated data setting, we use a single-layer RNN encoder. The selected window size is 4, representing about 15 seconds of recording, and the representations are encoded in a 10-dimensional vector space.
79
+
80
+ # 4 RESULTS
81
+
82
+ In this section we present the results for clusterability of the latent representations and downstream classification performance for all datasets and across all baselines. Clusterability indicates how well each method recovers appropriate states, and classification assesses how informative our representations are for downstream tasks.
83
+
84
+ # 4.1 EVALUATION: CLUSTERABILITY
85
+
86
+ Many real-world time series data have underlying multi-category structure, naturally leading to representations with clustering properties. Encoding such general priors is a property of a good representation (Bengio et al., 2013). In this section, we assess the distribution of the representations in the encoding space. If information of the latent state is properly learned and encoded by the framework, the representations of signals from the same underlying state should cluster together. Figures 2a, 2b, and 2c show an example of this distribution for simulated data across compared approaches. Each plot is a 2-dimensional t-SNE visualization of the representations where each data point in the scatter plot is an encoding $Z \in R ^ { 1 0 }$ that represents a window of size $\delta = 5 0$ of a simulated time series. We can see that without any information about the hidden states, representations learned using TNC cluster windows from the same hidden state better than the alternative approaches. The results show that CPC and Triplet Loss have difficulty separating time series that are generated from non-linear auto-regressive moving average (NARMA) models with variable regression parameters.
87
+
88
+ To compare the representation clusters’ consistency for each baseline, we use two very common cluster validity indices, namely, the Silhouette score and the Davies-Bouldin index. We use K-means clustering in the representation space to measure these clusterability scores. The Silhouette score measures the similarity of each sample to its own cluster, compared to other clusters. The values can range from $- 1$ to $+ 1$ , and a greater score implies a better cohesion. The Davies-Bouldin Index measures intra-cluster similarity and inter-cluster differences. This is a positive index score, where smaller values indicate low within-cluster scatter and large separation between clusters. Therefore, a lower score represents better clusterability (more details on the cluster validity scores and how they are calculated can be found in Appendix A.5).
89
+
90
+ ![](images/1e54dcf85574cebb36161a9083e97999516e415efc834be6219305e5d274ca9d.jpg)
91
+ Figure 2: T-SNE visualization of signal representations for the simulated dataset across all baselines. Each data point in the plot presents a 10-dimensional representation of a window of time series of size $\delta = 5 0$ , and the color indicates the latent state of the signal window. See Appendix A.7 for similar plots from different datasets.
92
+
93
+ <table><tr><td></td><td colspan="2">Simulation</td><td colspan="2">ECG Waveform</td><td colspan="2">HAR</td></tr><tr><td>Method</td><td>Silhouette 个</td><td>DBI↓</td><td>Silhouette 个</td><td>DBI↓</td><td>Silhouette↑</td><td>DBI↓</td></tr><tr><td>TNC</td><td>0.71±0.01</td><td>0.36±0.01</td><td>0.44±0.02</td><td>0.74±0.04</td><td>0.61±0.02</td><td>0.52±0.04</td></tr><tr><td>CPC</td><td>0.51±0.03</td><td>0.84±0.06</td><td>0.26±0.02</td><td>1.44±0.04</td><td>0.58±0.02</td><td>0.57±0.05</td></tr><tr><td>T-Loss</td><td>0.61±0.08</td><td>0.64±0.12</td><td>0.25±0.01</td><td>1.30±0.03</td><td>0.17±0.01</td><td>1.76±0.20</td></tr><tr><td>K-means</td><td>0.01±0.019</td><td>7.23±0.14</td><td>0.19±0.11</td><td>3.65±0.48</td><td>0.12±0.40</td><td>2.66±0.05</td></tr></table>
94
+
95
+ Table 1: Clustering quality of representations in the encoding space for multiple datasets.
96
+
97
+ Table 1 summarizes the scores for all baselines and across all datasets, demonstrating that TNC is superior in learning representations that can distinguish the latent dynamics of time series. CPC performs closely to Triplet loss on waveform data but performs poorly on the simulated dataset, where signals are highly non-stationary, and transitions are less predictable. However, for the HAR dataset, CPC clusters the states very well because most activities are recorded in a specific order, empowering predictive coding. Triplet loss performs reasonably well in the simulated setting; however, it fails to distinguish states 0 and 2, where signals come from autoregressive models with different parameters and have a relatively similar generative process. Performing K-means on the original time series generally does not generate coherent clusters, as demonstrated by the scores. However, the performance is slightly better in time series like the ECG waveforms, where the signals are formed by consistent shapelets, and therefore the DTW measures similarity more accurately.
98
+
99
+ # 4.2 EVALUATION: CLASSIFICATION
100
+
101
+ We further evaluate the quality of the encodings using a classification task. We train a linear classifier to evaluate how well the representations can be used to classify hidden states. The performance of all baselines is compared to a supervised classifier, composed of an encoder and a classifier with identical architectures to that of the unsupervised models, and a K-nearest neighbor classifier that uses DTW metric. The performance is reported as the prediction accuracy and the area under the precision-recall curve (AUPRC) score since AUPRC is a more accurate reflection of model performance for imbalance classification settings like the waveform dataset.
102
+
103
+ Table 2 demonstrates the classification performance for all datasets. The performance of the classifiers that use TNC representations are closer to the end-to-end supervised model in comparison to CPC and Triplet Loss. This provides further evidence that our encodings capture informative parts of the time series and are generalizable to be used for downstream tasks. In datasets like the HAR, where an inherent ordering usually exists in the time series, CPC performs reasonably. However, in datasets with increased non-stationarity, the performance drops. Triplet Loss is also a powerful framework, but since it samples positive examples from overlapping windows of time series, it is vulnerable to map the overlaps into the encoding and, therefore, fail to learn more general representations. TNC, on the other hand, samples similar windows from a wider distribution, defined by the temporal neighborhood, where many of the neighboring signals do not necessarily overlap. The lower performance of the CPC and Triplet Loss methods can also be partly because none of these methods explicitly account for the potential sampling bias that happens when randomly selected negative examples are similar to the reference $W _ { t }$ .
104
+
105
+ <table><tr><td colspan="3">Simulation</td><td colspan="2">ECG Waveform</td><td colspan="2">HAR</td></tr><tr><td>Method</td><td>AUPRC</td><td>Accuracy</td><td>AUPRC</td><td>Accuracy</td><td>AUPRC</td><td>Accuracy</td></tr><tr><td>TNC</td><td>0.99±0.00</td><td>97.52±0.13</td><td>0.55±0.01</td><td>77.79±0.84</td><td>0.94±0.007</td><td>88.32±0.12</td></tr><tr><td>CPC</td><td>0.69±0.06</td><td>70.26±6.48</td><td>0.42±0.01</td><td>68.64±0.49</td><td>0.93±0.006</td><td>86.43±1.41</td></tr><tr><td>T-Loss</td><td>0.78±0.01</td><td>76.66±1.40</td><td>0.47±0.00</td><td>75.51±1.26</td><td>0.71±0.007</td><td>63.60±3.37</td></tr><tr><td>KNN</td><td>0.42±0.00</td><td>55.53±0.65</td><td>0.38±0.06</td><td>54.76±5.46</td><td>0.75±0.01</td><td>84.85±0.84</td></tr><tr><td>Supervised</td><td>0.99±0.00</td><td>98.56±0.13</td><td>0.67±0.01</td><td>94.81±0.28</td><td>0.98±0.00</td><td>92.03±2.48</td></tr></table>
106
+
107
+ Table 2: Performance of all baselines in classifying the underlying hidden states of the time series, measured as the accuracy and AUPRC score.
108
+
109
+ # 4.3 EVALUATION: TRAJECTORY
110
+
111
+ ![](images/57ec11acc26cc0da64f569141bc5f07e4d7032734b4804873f3d1735a8eacdcc.jpg)
112
+ Figure 3: Trajectory of a signal encoding from the simulated dataset. The top plot shows the original time series with shaded regions indicating the underlying state. The bottom plot shows the 10 dimensional encoding of the sliding windows $W _ { t }$ where $\delta = 5 0$ .
113
+
114
+ This section investigates the trajectories of our learned encodings over time to understand how the state transitions are captured and modeled in the representation space. This is an important property for non-stationary time series where underlying states change over time, and capturing those changes is critical in many application domains such as healthcare. Figure 3 shows a sample from the simulated dataset. The top panel shows the signal measurements over time, and the shaded regions indicate the underlying latent states. The bottom panel illustrates the 10-dimensional representation of a sliding window $W _ { t }$ estimated over time. From the bottom panel of Figure 3, we can see that the encoding pattern changes at state transitions and settle into a different pattern, corresponding to the new state. This change happens at every transition, and we can see the distinct patterns for all 4 underlying states in the representations. This analysis of the trajectory of change could be very informative for the users’ post-analysis; for instance, in clinical applications, it could help clinicians visualize the evolution of the patient state over time and plan treatment based on the state progression
115
+
116
+ # 5 RELATED WORK
117
+
118
+ While integral for many applications, unsupervised representation learning has been far less studied for time series (Längkvist et al., 2014), compared to other domains such as vision or natural language processing (Denton et al., 2017; Radford et al., 2015; Gutmann & Hyvärinen, 2012; Wang & Gupta, 2015). One of the earliest approaches to unsupervised end-to-end representation learning in time series is the use of auto-encoders (Choi et al., 2016a; Amiriparian et al., 2017; Malhotra et al., 2017) and seq-to-seq models (Lyu et al., 2018), with the objective to train an encoder jointly with a decoder that reconstructs the input signal from its learned representation. Using fully generative models like variational auto-encoders is also useful for imposing properties like disentanglement, which help with the interpretability of the representations (Dezfouli et al., 2019). However, in many cases, like for high-frequency physiological signals, the reconstruction of complex time series can be challenging; therefore, more novel approaches are designed to avoid this step. Contrastive Predictive Coding (Oord et al., 2018; Löwe et al., 2019) learns representations by predicting the future in latent space, eliminating the need to reconstruct the full input. The representations are such that the mutual information between the original signal and the concept vector is maximally preserve using a lower bound approximation and a contrastive loss. Very similarly, in Time Contrastive Learning (Hyvarinen & Morioka, 2016), a contrastive loss is used to predict the segment-ID of multivariate time-series as a way to extract representation. Franceschi et al. (2019) employs time-based negative sampling and a triplet loss to learn scalable representations for multivariate time series. Some other approaches use inherent similarities in temporal data to learn representations without supervision. For instance, in similarity-preserving representation learning (Lei et al., 2019), learned encodings are constrained to preserve the pairwise similarities that exist in the time domain, measured by DTW distance. Another group of approaches combines reconstruction loss with clustering objectives to cluster similar temporal patterns in the encoding space (Ma et al., 2019).
119
+
120
+ In healthcare, learning representation of rich temporal medical data is extremely important for understanding patients’ underlying health conditions. However, most of the existing approaches for learning representations are designed for specific downstream tasks and require labeling by experts (Choi et al., 2016b;c; Tonekaboni et al., 2020). Examples of similar works to representation learning in the field of clinical ML include computational phenotyping for discovering subgroups of patients with similar underlying disease mechanisms from temporal clinical data (Lasko et al., 2013; Suresh et al., 2018; Schulam et al., 2015), and disease progression modeling, for learning the hidden vector of comorbidities representing a disease over time (Wang et al., 2014; Alaa & van der Schaar, 2019).
121
+
122
+ # 6 CONCLUSION
123
+
124
+ This paper presents a novel unsupervised representation learning framework for complex multivariate time series, called Temporal Neighborhood Coding (TNC). This framework is designed to learn the underlying dynamics of non-stationary signals and to model the progression over time by defining a temporal neighborhood. The problem is motivated by the medical field, where patients transition between distinct clinical states over time, and obtaining labels to define these underlying states is challenging. We evaluate the performance of TNC on multiple datasets and show that our representations are generalizable and can easily be used for diverse tasks such as classification and clustering. We finally note that TNC is flexible to be used with arbitrary encoder architectures; therefore, the framework is applicable to many time series data domains. Moreover, in addition to tasks presented in this paper, general representations can be used for several other downstream tasks, such as anomaly detection, which is challenging in supervised learning settings for time series data in sparsely labeled contexts.
125
+
126
+ # ACKNOWLEDGMENTS
127
+
128
+ Resources used in preparing this research were provided, in part, by the Government of Canada through CIFAR, and companies sponsoring the Vector Institute. This research was undertaken, in part, thanks to funding from the Canadian Institute of Health Research (CIHR) and the Natural Sciences and Engineering Research Council of Canada (NSERC).
129
+
130
+ # REFERENCES
131
+
132
+ Ahmed Alaa and Mihaela van der Schaar. Attentive state-space modeling of disease progression. 2019.
133
+
134
+ Shahin Amiriparian, Michael Freitag, Nicholas Cummins, and Björn Schuller. Sequence to sequence autoencoders for unsupervised representation learning from audio. In Proc. of the DCASE 2017 Workshop, 2017.
135
+
136
+ Shaojie Bai, J Zico Kolter, and Vladlen Koltun. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv preprint arXiv:1803.01271, 2018.
137
+
138
+ Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence, 35(8):1798–1828, 2013.
139
+
140
+ Edward Choi, Mohammad Taha Bahadori, Elizabeth Searles, Catherine Coffey, Michael Thompson, James Bost, Javier Tejedor-Sojo, and Jimeng Sun. Multi-layer representation learning for medical concepts. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 1495–1504, 2016a.
141
+
142
+ Edward Choi, Andy Schuetz, Walter F Stewart, and Jimeng Sun. Medical concept representation learning from electronic health records and its application on heart failure prediction. arXiv preprint arXiv:1602.03686, 2016b.
143
+
144
+ Youngduck Choi, Chill Yi-I Chiu, and David Sontag. Learning low-dimensional representations of medical concepts. AMIA Summits on Translational Science Proceedings, 2016:41, 2016c.
145
+
146
+ Ching-Yao Chuang, Joshua Robinson, Lin Yen-Chen, Antonio Torralba, and Stefanie Jegelka. Debiased contrastive learning. arXiv preprint arXiv:2007.00224, 2020.
147
+
148
+ Emily L Denton et al. Unsupervised learning of disentangled representations from video. In Advances in Neural Information Processing Systems, pp. 4414–4423, 2017.
149
+
150
+ Amir Dezfouli, Hassan Ashtiani, Omar Ghattas, Richard Nock, Peter Dayan, and Cheng Soon Ong. Disentangled behavioural representations. In Advances in Neural Information Processing Systems, pp. 2251–2260, 2019.
151
+
152
+ Jeff Donahue and Karen Simonyan. Large scale adversarial representation learning. In Advances in Neural Information Processing Systems, pp. 10541–10551, 2019.
153
+
154
+ Marthinus C Du Plessis, Gang Niu, and Masashi Sugiyama. Analysis of learning from positive and unlabeled data. In Advances in Neural Information Processing Systems, pp. 703–711, 2014.
155
+
156
+ Marthinus Christoffel Du Plessis and Masashi Sugiyama. Class prior estimation from positive and unlabeled data. IEICE TRANSACTIONS on Information and Systems, 97(5):1358–1362, 2014.
157
+
158
+ Charles Elkan and Keith Noto. Learning classifiers from only positive and unlabeled data. In Proceedings of the 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 213–220, 2008.
159
+
160
+ Hassan Ismail Fawaz, Germain Forestier, Jonathan Weber, Lhassane Idoumghar, and Pierre-Alain Muller. Deep learning for time series classification: a review. Data Mining and Knowledge Discovery, 33(4):917–963, 2019.
161
+
162
+ Jean-Yves Franceschi, Aymeric Dieuleveut, and Martin Jaggi. Unsupervised scalable representation learning for multivariate time series. In Advances in Neural Information Processing Systems, pp. 4652–4663, 2019.
163
+
164
+ Michael U Gutmann and Aapo Hyvärinen. Noise-contrastive estimation of unnormalized statistical models, with applications to natural image statistics. Journal of Machine Learning Research, 13 (Feb):307–361, 2012.
165
+
166
+ Aapo Hyvarinen and Hiroshi Morioka. Unsupervised feature extraction by time-contrastive learning and nonlinear ICA. In Advances in Neural Information Processing Systems, pp. 3765–3773, 2016.
167
+
168
+ Ryuichi Kiryo, Gang Niu, Marthinus C Du Plessis, and Masashi Sugiyama. Positive-unlabeled learning with non-negative risk estimator. In Advances in Neural Information Processing Systems, pp. 1675–1685, 2017.
169
+
170
+ Martin Längkvist, Lars Karlsson, and Amy Loutfi. A review of unsupervised feature learning and deep learning for time-series modeling. Pattern Recognition Letters, 42:11–24, 2014.
171
+
172
+ Thomas A Lasko, Joshua C Denny, and Mia A Levy. Computational phenotype discovery using unsupervised feature learning over noisy, sparse, and irregular clinical data. PloS one, 8(6), 2013.
173
+
174
+ Wee Sun Lee and Bing Liu. Learning with positive and unlabeled examples using weighted logistic regression. In International Conference on Machine Learning, volume 3, pp. 448–455, 2003.
175
+
176
+ Qi Lei, Jinfeng Yi, Roman Vaculin, Lingfei Wu, and Inderjit S Dhillon. Similarity preserving representation learning for time series clustering. In IJCAI, volume 19, pp. 2845–2851, 2019.
177
+
178
+ Xiaoli Li and Bing Liu. Learning to classify texts using positive and unlabeled data. In IJCAI, volume 3, pp. 587–592, 2003.
179
+
180
+ Sindy Löwe, Peter O’Connor, and Bastiaan Veeling. Putting an end to end-to-end: Gradient-isolated learning of representations. In Advances in Neural Information Processing Systems, pp. 3039–3051, 2019.
181
+
182
+ Xinrui Lyu, Matthias Hueser, Stephanie L Hyland, George Zerveas, and Gunnar Rätsch. Improving clinical predictions through unsupervised time series representation learning. arXiv preprint arXiv:1812.00490, 2018.
183
+
184
+ Qianli Ma, Jiawei Zheng, Sen Li, and Gary W Cottrell. Learning representations for time series clustering. In Advances in Neural Information Processing Systems, pp. 3776–3786, 2019.
185
+
186
+ Pankaj Malhotra, Vishnu TV, Lovekesh Vig, Puneet Agarwal, and Gautam Shroff. Timenet: Pre-trained deep recurrent neural network for time series classification. arXiv preprint arXiv:1706.08838, 2017.
187
+
188
+ Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781, 2013.
189
+
190
+ George Moody. A new method for detecting atrial fibrillation using RR intervals. Computers in Cardiology, pp. 227–230, 1983.
191
+
192
+ Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw audio. arXiv preprint arXiv:1609.03499, 2016.
193
+
194
+ Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018.
195
+
196
+ Slobodan Petrovic. A comparison between the silhouette index and the davies-bouldin index in labelling ids clusters. In Proceedings of the 11th Nordic Workshop of Secure IT Systems, pp. 53–64, 2006.
197
+
198
+ Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015.
199
+
200
+ Alec Radford, Rafal Jozefowicz, and Ilya Sutskever. Learning to generate reviews and discovering sentiment. arXiv preprint arXiv:1704.01444, 2017.
201
+
202
+ Nikunj Saunshi, Orestis Plevrakis, Sanjeev Arora, Mikhail Khodak, and Hrishikesh Khandeparkar. A theoretical analysis of contrastive unsupervised representation learning. In International Conference on Machine Learning, pp. 5628–5637. PMLR, 2019.
203
+
204
+ Peter Schulam, Fredrick Wigley, and Suchi Saria. Clustering longitudinal clinical marker trajectories from electronic health data: Applications to phenotyping and endotype discovery. In Twenty-Ninth AAAI Conference on Artificial Intelligence, 2015.
205
+
206
+ Harini Suresh, Jen J Gong, and John V Guttag. Learning tasks for multitask learning: Heterogenous patient populations in the icu. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pp. 802–810, 2018.
207
+
208
+ Sana Tonekaboni, Shalmali Joshi, Kieran Campbell, David K Duvenaud, and Anna Goldenberg. What went wrong and when? instance-wise feature importance for time-series black-box models. Advances in Neural Information Processing Systems, 33, 2020.
209
+
210
+ Xiang Wang, David Sontag, and Fei Wang. Unsupervised learning of disease progression models. In Proceedings of the 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 85–94, 2014.
211
+
212
+ Xiaolong Wang and Abhinav Gupta. Unsupervised learning of visual representations using videos. In Proceedings of the IEEE International Conference on Computer Vision, pp. 2794–2802, 2015.
213
+
214
+ Qiang Yang and Xindong Wu. 10 challenging problems in data mining research. International Journal of Information Technology & Decision Making, 5(04):597–604, 2006.
215
+
216
+ Tom Young, Devamanyu Hazarika, Soujanya Poria, and Erik Cambria. Recent trends in deep learning based natural language processing. IEEE Computational Intelligence Magazine, 13(3):55–75, 2018.
217
+
218
+ A APPENDIX
219
+
220
+ # A.1 SIMULATED DATASET
221
+
222
+ ![](images/50d00ce6f7b9850d9be2d0710b8c08d5aea991a9f2270a96461961fad39a337e.jpg)
223
+ Figure A.1: A normalized time series sample from the simulated dataset. Each row represents a single feature, and the shaded regions indicate one of the 4 underllying simulated states.
224
+
225
+ The simulated time series consists of 3 features generated from different underlying hidden states. Figure A.1 shows a sample from this dataset. Each panel in the figure shows one of the features, and the shaded regions indicate the underlying state of the signal in that period. We use a Hidden Markov Model (HMM) to generate these random latent states over time. The transition probability is set equal to $\% 5$ for switching to an alternating state, and $\% 8 5$ for not changing state. In each state, the time series is generated from a different signal distribution. Table 3 describes the generative process of each signal feature in each state. Note that feature 1 and 2 are always correlated, mainly to mimic realistic clinical time series. As an example, physiological measurements like pulse rate and heart rate are always correlated.
226
+
227
+ Table 3: Signal distributions for each time series feature of the simulated dataset
228
+
229
+ <table><tr><td></td><td>State 1</td><td>State 2</td><td>State 3</td><td>State 4</td></tr><tr><td>Feature 1</td><td>GP (periodic)</td><td>NARMAα</td><td>GP (Squared Exp.)</td><td>NARMAβ</td></tr><tr><td>Feature 2</td><td>GP (periodic)</td><td>NARMAα</td><td>GP (Squared Exp.)</td><td>NARMAβ</td></tr><tr><td>Feature 3</td><td>GP (Squared Exp.)</td><td>NARMAβ</td><td>GP (periodic)</td><td>NARMAα</td></tr></table>
230
+
231
+ In the state 1, the correlated features are generated by a Gaussian Process (GP) with a periodic kernel. Feature 3, which is uncorrelated with the other two features, comes from another GP with a squared exponential kernel. In addition to GPs, we also have multiple Non-Linear Auto-Regressive Moving Average (NARMA) time series models. The linear function of $\mathrm { N A R M A } _ { \alpha }$ and $\mathrm { N A R M A } _ { \beta }$ are shown in Equation 2 and 3.
232
+
233
+ $$
234
+ \mathrm { N A R M A } _ { \alpha } : y ( k + 1 ) = 0 . 3 y ( k ) + 0 . 0 5 y ( k ) \sum _ { i = 0 } ^ { n - 1 } y ( k - i ) + 1 . 5 u ( k - ( n - 1 ) ) u ( k ) + 0 . 1 1
235
+ $$
236
+
237
+ $$
238
+ \mathbf { N A R M A } _ { \beta } : y ( k + 1 ) = 0 . 1 y ( k ) + 0 . 2 5 y ( k ) \sum _ { i = 0 } ^ { n - 1 } y ( k - i ) + 2 . 5 u ( k - ( n - 1 ) ) u ( k ) - 0 . 0 0 5
239
+ $$
240
+
241
+ A white Gaussian noise with $\sigma = 0 . 3$ is added to all signals, and overall, the dataset consists of 500 instances of $T = 2 0 0 0$ measurements.
242
+
243
+ # A.2 BASELINE IMPLEMENTATION DETAILS
244
+
245
+ Implementation of all baselines are included in the code base for reproducibility purposes, and hyper-parameters for all baselines are tuned using cross-validation.
246
+
247
+ Contrastive Predictive Coding (CPC): The CPC baseline first processes the sequential signal windows using an encoder $Z _ { t } = E n c ( X _ { t } )$ , with a similar architecture to the encoders of other baselines. Next, an autoregressive model $g _ { a r }$ aggregates all the information in $Z _ { \le t }$ and summarizes it into a context latent representation $c _ { t } = g _ { a r } ( Z _ { \leq t } )$ . In our implementation, we have used a single layer, a one-directional recurrent neural network with GRU cell and hidden size equal to the encoding size as the auto-regressor. Like the original paper, the density ratio is estimated using a linear transformation, and the model is trained for 1 step ahead prediction.
248
+
249
+ Triplet-Loss (T-Loss): The triplet loss baseline is implemented using the original code made available by the authors on Github3.
250
+
251
+ KNN and K-means: These two baselines for classification and clustering are implemented using the tslearn library 4, that integrates distance metrics such as DTW. Note that evaluating DTW is computationally expensive, and the tslearn implementation is not optimized. Therefore, for the waveform data with windows of size 2500, we had to down-sampled the signal frequency by a factor of two.
252
+
253
+ # A.3 TNC IMPLEMENTATION EXTRA DETAILS
254
+
255
+ To define the neighborhood range in the TNC framework, as mentioned earlier, we use the AugmentedDickey Fuller (ADF) statistical test to determine this range $( \eta )$ as the region for which the signals remain stationary. More precisely, we gradually increase the range, from a single window size up to 3 times the window size (the upper limit we set), and repeatedly perform the ADF test. We use the $p$ -value from this statistical test to determine whether the Null hypothesis can be rejected, meaning that the signal is stationary. At the point where the $p$ -value is above our defined threshold (0.01), we can no longer assume that the signal is stationary, and this is where we set the $\eta$ parameter. Now, once the neighborhood is defined, we make sure the non-neighboring samples are taken from the distribution of windows with at least $4 \times \eta$ away from $W _ { t }$ , ensuring a low likelihood of belonging to the neighborhood. Note that for implementation of ADF, we use the stats model library 5. Unfortunately, this implementation is not optimized and does not support GPU computation, therefore evaluating the neighborhood range using ADF slows down TNC framework training. As a future direction, we are working on an optimized implementation of the ADF score for our framework.
256
+
257
+ # A.4 SELECTING THE WINDOW SIZE
258
+
259
+ The window size $\delta$ is an important factor in the performance of a representation learning framework, not only for TNC but also for similar baselines such as CPC and triplet loss. Overall, the window size should be selected such that it is long enough to contain information about the underlying state and not too long to span over multiple underlying states. In our settings, we have selected the window sizes based on our prior knowledge of the signals. For instance, in the case of an ECG signal, the selected window size is equivalent to 7 seconds of recording, which is small enough such that the ECG remains in a stable state and yet has enough information to determine that underlying state. Our understanding of the time series data can help us select an appropriate window size, but we can also experiment with different $\delta$ to learn this parameter. Table 4 shows classification performance results for the simulation setups, under different window sizes. We can clearly see the drop in performance for all baseline methods when the window size is too small or too large.
260
+
261
+ <table><tr><td></td><td colspan="2">δ=10</td><td colspan="2">8=50</td><td colspan="2">δ=100</td></tr><tr><td></td><td>AUPRC</td><td> Accuracy</td><td>AUPRC</td><td>Accuracy</td><td>AUPRC</td><td>Accuracy</td></tr><tr><td>TNC</td><td>0.74 ± 0.01</td><td>71.60 ± 0.59</td><td>0.99 ± 0.00</td><td>97.52 ± 0.13</td><td>0.84 ± 0.11</td><td>84.25 ± 9.08</td></tr><tr><td>CPC</td><td>0.49 ± 0.02</td><td>51.85 ± 1.81</td><td>0.69 ± 0.06</td><td>70.26 ± 6.48</td><td>0.49 ± 0.05</td><td>56.65 ± 0.81</td></tr><tr><td>T-Loss</td><td>0.48 ± 0.06</td><td>56.70 ± 1.07</td><td>0.78 ± 0.01</td><td>76.66 ± 1.14</td><td>0.73 ± 0.008</td><td>73.29 ± 1.58</td></tr></table>
262
+
263
+ Table 4: Downstream classification performance for different window size $\delta$ on the simulated dataset
264
+
265
+ # A.5 CLUSTERING METRICS
266
+
267
+ Most cluster validity measures assess certain structural properties of a clustering result. In our evaluation, we have used two measures, namely the Silhouette score and Davies-Bouldin index, to evaluate the representations’ clustering quality. Davies-Bouldin measures intra-cluster similarity (coherence) and inter-cluster differences (separation). Let $\mathcal { C } = \{ \mathcal { C } _ { 1 } , . . . , \mathcal { C } _ { k } \}$ be a clustering of a set $D$ of objects. The Davies-Bouldin score is evaluated as follows:
268
+
269
+ $$
270
+ D B = \frac { 1 } { k } \sum _ { i = 1 } ^ { k } m a x _ { j } \frac { s ( \mathcal { C } ) + s ( \mathcal { C } ) } { \delta \mathcal { C } _ { i } \mathcal { C } _ { j } }
271
+ $$
272
+
273
+ Where $s ( \mathcal { C } )$ measures the scatter within a cluster, and $\delta$ is a cluster to cluster distance measure. On the other hand, the silhouette score measures how similar an object is to its cluster compared to other clusters. Both measures are commonly used for the evaluation of clustering algorithms. A comparison of 2 metrics has shown that the Silhouette index produces slightly more accurate results in some cases. However, the Davies-Bouldin index is generally much less complex to compute Petrovic (2006).
274
+
275
+ # A.6 SETTING THE WEIGHTS FOR PU LEARNING
276
+
277
+ As mentioned in the Experiment section, the weight parameter in the loss is the probability of sampling a positive window from the non-neighboring region. One way to set this parameter is using prior knowledge of the number and the distribution of underlying states. Another way is to learn it as a hyperparameter. Table 5 shows the TNC loss for different weight parameters. The loss column reports the value measured in Equation 1, and the accuracy shows how well the discriminator identifies the neighboring samples from non-neighboring ones for settings with different weight parameters. To also assess the impact of re-weighting the loss on downstream classification performance, we compared these performance measures for weighted and non-weighted settings. Table 6 demonstrates these results and confirms that weight adjusting the loss for non-neighboring samples improves the quality of learned representations.
278
+
279
+ Table 5: Training the TNC framework using different weight parameters. The loss is the measured value determined in Equation 1, and the Accuracy is the accuracy of the discriminator.
280
+
281
+ <table><tr><td></td><td colspan="2">Simulation</td><td colspan="2">ECG Waveform</td><td colspan="2">HAR</td></tr><tr><td>Weight</td><td>Loss</td><td>Accuracy</td><td>Loss</td><td>Accuracy</td><td>Loss</td><td>Accuracy</td></tr><tr><td>0.2</td><td>0.582±0.002</td><td>74.29±0.61</td><td>0.631±0.011</td><td>60.44±2.56</td><td>0.475±0.004</td><td>85.75±0.5</td></tr><tr><td>0.1</td><td>0.571±0.011</td><td>75.41±0.37</td><td>0.637±0.011</td><td>63.67±1.29</td><td>0.413±0.003</td><td>88.21±1.29</td></tr><tr><td>0.05</td><td>0.576±0.002</td><td>75.73±0.24</td><td>0.622±0.023</td><td>66.04±3.46</td><td>0.383±0.001</td><td>87.33±0.17</td></tr></table>
282
+
283
+ <table><tr><td>Weighting?</td><td>Simulation</td><td>ECG Waveform</td><td>HAR</td></tr><tr><td>True</td><td>97.52±0.13</td><td>77.79±0.84</td><td>88.32±0.12</td></tr><tr><td>False</td><td>97.17±0.44</td><td>75.26±1.48</td><td>75.25±13.6</td></tr></table>
284
+
285
+ Table 6: Downstream classification accuracy on simulated data with the TNC frameworks, using 2 different weighting strategies: 1)Trained with weight adjustment, 2)Trained with $w = 0$ .
286
+
287
+ # A.7 SUPPLEMENTARY FIGURE
288
+
289
+ # A.7.1 CLINICAL WAVEFORM DATA
290
+
291
+ In order to understand what TNC framework encodes from the high dimensional ECG signals, we visualize the trajectory of the representations of an individual sample over time. Figure A.3 demonstrates this example, where the top 2 rows are ECG signals from two recording leads and the bottom row demonstrates the representation vectors. We see that around second 40 the pattern in the representations change as a result of an artifact happening in one of the signals. With the help from our clinical expert, we also tried to interpret different patterns in the encoding space. For instance, between time 80 and 130, where features 0-10 become more activated, the heart rate (HR) has increased. Increase in HR can be seen as increased frequency in the ECG signals and is one of the indicators of arrhythmia that we believe TNC has captured. Figure A.2 shows the distribution of the latent encoding of ECG signals for different baselines, with colors indicating the arrhythmia class.
292
+
293
+ ![](images/f8c41d9d7b7f3ca2dee2d86c0465aaca53aa42304010fcb04480e641114baedc.jpg)
294
+ Figure A.2: T-SNE visualization of waveform signal representations for unsupervised representation learning baselines. Each point in the plot is a 64 dimensional representation of a window of time series, with the color indicating the latent state.
295
+
296
+ ![](images/cdf9f7bf417aea0ea9e860d91a4ecbbb954f0389143f6bc77bc251a49ec2386f.jpg)
297
+ Figure A.3: Trajectory of a waveform signal encoding. The top two plots show the ECG recordings from 2 ECG leads. The bottom plot shows the 64 dimensional encoding of the sliding windows $W _ { t }$ where $\delta = 2 5 0 0$ .
298
+
299
+ # A.7.2 HAR DATA
300
+
301
+ Figure A.4 and A.5 are similar plots to the ones demonstrated in the previous section, but for the HAR dataset. As shown in Figure A.5, the underlying states of the signal are clearly captured by the TNC framework as different patterns in the latent representations.
302
+
303
+ ![](images/38c699134734a4f23855bab05972dcbbbc7b6990cd92481f77433a1d84d0a101.jpg)
304
+ Figure A.4: T-SNE visualization of HAR signal representations for all baselines. Each point in the plot is a 10 dimensional representation of a window of $\delta = 4$ , with colors indicating latent states.
305
+
306
+ ![](images/216f68c8384e1cbf4eb2ddfdaf789ae535631b1609fb8be97fdb673ecdfa0ab9.jpg)
307
+ Figure A.5: Trajectory of a HAR signal encoding. The top plot shows the original time series with shaded regions indicating the underlying state. The bottom plot shows the 10 dimensional encoding of the sliding windows $W _ { t }$ where $\delta = 4$ .
308
+
309
+ # A.7.3 SIMULATION DATA
310
+
311
+ In addition to the initial experiment, we also show the trajectory of the encoding for a smaller encoding size (3). In this setting, we have 4 underlying states in the signal, and only 3 dimensions for the encoding.
312
+
313
+ ![](images/074b1e338add5ee800eec2f4ff2c88c3d91ab6b2e3e944728fe7d5a7e281b535.jpg)
314
+ Figure A.6: Trajectory of a simulation signal encoding. The top plots shows the signals and the bottom plot shows the 3 dimensional encoding of the sliding windows $W _ { t }$ where $\delta = 5 0$ .
parse/train/8qDwejCuCN/8qDwejCuCN_content_list.json ADDED
@@ -0,0 +1,1735 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "type": "text",
4
+ "text": "UNSUPERVISED REPRESENTATION LEARNING FOR TIME SERIES WITH TEMPORAL NEIGHBORHOOD CODING ",
5
+ "text_level": 1,
6
+ "bbox": [
7
+ 174,
8
+ 99,
9
+ 722,
10
+ 171
11
+ ],
12
+ "page_idx": 0
13
+ },
14
+ {
15
+ "type": "text",
16
+ "text": "Sana Tonekaboni∗ University of Toronto & Vector Institute The Hospital for Sick Children stonekaboni@cs.toronto.edu ",
17
+ "bbox": [
18
+ 184,
19
+ 195,
20
+ 449,
21
+ 251
22
+ ],
23
+ "page_idx": 0
24
+ },
25
+ {
26
+ "type": "text",
27
+ "text": "Danny Eytan The Hospital for Sick Children biliary.colic@gmail.com ",
28
+ "bbox": [
29
+ 586,
30
+ 195,
31
+ 813,
32
+ 237
33
+ ],
34
+ "page_idx": 0
35
+ },
36
+ {
37
+ "type": "text",
38
+ "text": "Anna Goldengerg University of Toronto & Vector Institute The Hospital for Sick Children anna.goldenberg@utoronto.ca ",
39
+ "bbox": [
40
+ 183,
41
+ 272,
42
+ 449,
43
+ 328
44
+ ],
45
+ "page_idx": 0
46
+ },
47
+ {
48
+ "type": "text",
49
+ "text": "ABSTRACT ",
50
+ "text_level": 1,
51
+ "bbox": [
52
+ 454,
53
+ 364,
54
+ 544,
55
+ 380
56
+ ],
57
+ "page_idx": 0
58
+ },
59
+ {
60
+ "type": "text",
61
+ "text": "Time series are often complex and rich in information but sparsely labeled and therefore challenging to model. In this paper, we propose a self-supervised framework for learning generalizable representations for non-stationary time series. Our approach, called Temporal Neighborhood Coding (TNC), takes advantage of the local smoothness of a signal’s generative process to define neighborhoods in time with stationary properties. Using a debiased contrastive objective, our framework learns time series representations by ensuring that in the encoding space, the distribution of signals from within a neighborhood is distinguishable from the distribution of non-neighboring signals. Our motivation stems from the medical field, where the ability to model the dynamic nature of time series data is especially valuable for identifying, tracking, and predicting the underlying patients’ latent states in settings where labeling data is practically impossible. We compare our method to recently developed unsupervised representation learning approaches and demonstrate superior performance on clustering and classification tasks for multiple datasets. ",
62
+ "bbox": [
63
+ 233,
64
+ 395,
65
+ 766,
66
+ 602
67
+ ],
68
+ "page_idx": 0
69
+ },
70
+ {
71
+ "type": "text",
72
+ "text": "1 INTRODUCTION ",
73
+ "text_level": 1,
74
+ "bbox": [
75
+ 176,
76
+ 627,
77
+ 336,
78
+ 643
79
+ ],
80
+ "page_idx": 0
81
+ },
82
+ {
83
+ "type": "text",
84
+ "text": "Real-world time-series data is high dimensional, complex, and has unique properties that bring about many challenges for data modeling (Yang & Wu, 2006). In addition, these signals are often sparsely labeled, making it even more challenging for supervised learning tasks. Unsupervised representation learning can extract informative low-dimensional representations from raw time series by leveraging the data’s inherent structure, without the need for explicit supervision. These representations are more generalizable and robust, as they are less specialized for solving a single supervised task. Unsupervised representation learning is well studied in domains such as vision (Donahue & Simonyan, 2019; Denton et al., 2017; Radford et al., 2015) and natural language processing (Radford et al., 2017; Young et al., 2018; Mikolov et al., 2013), but has been underexplored in the literature for time series settings. Frameworks designed for time series need to be efficient and scalable because signals encountered in practice can be long, high dimensional, and high frequency. Moreover, it should account for and be able to model dynamic changes that occur within samples, i.e., non-stationarity of signals. ",
85
+ "bbox": [
86
+ 174,
87
+ 659,
88
+ 825,
89
+ 839
90
+ ],
91
+ "page_idx": 0
92
+ },
93
+ {
94
+ "type": "text",
95
+ "text": "The ability to model the dynamic nature of time series data is especially valuable in medicine. Health care data is often organized as a time series, with multiple data types, collected from various sources at different sampling frequencies, and riddled with artifacts and missing values. Throughout their stay at the hospital or within the disease progression period, patients transition gradually between distinct clinical states, with periods of relative stability, improvement, or unexpected deterioration, requiring escalation of care that alters the patient’s trajectory. A particular challenge in medical time-series data is the lack of well-defined or available labels that are needed for identifying the underlying clinical state of an individual or for training models aimed at extracting low-dimensional representations of these states. For instance, in the context of critical-care, a patient’s stay in the critical care unit (CCU) is captured continuously via streaming physiological signals by the bedside monitor. Obtaining labels for the patient’s state for extended periods of these signals is practically impossible as the underlying physiological state can be unknown even to the clinicians. This further motivates the use of unsupervised representation learning in these contexts. Learning rich representations can be crucial in facilitating the tracking of disease progression, predicting the future trajectories of the patients, and tailoring treatments to these underlying states. ",
96
+ "bbox": [
97
+ 174,
98
+ 847,
99
+ 825,
100
+ 904
101
+ ],
102
+ "page_idx": 0
103
+ },
104
+ {
105
+ "type": "text",
106
+ "text": "",
107
+ "bbox": [
108
+ 174,
109
+ 103,
110
+ 825,
111
+ 256
112
+ ],
113
+ "page_idx": 1
114
+ },
115
+ {
116
+ "type": "text",
117
+ "text": "In this paper, we propose a self-supervised framework for learning representations for complex multivariate non-stationary time series. This approach, called Temporal Neighborhood Coding (TNC), is designed for temporal settings where the latent distribution of the signals changes over time, and it aims to capture the progression of the underlying temporal dynamics. TNC is efficient, easily scalable to high dimensions, and can be used in different time series settings. We assess the quality of the learned representations on multiple datasets and show that the representations are general and transferable to many downstream tasks such as classification and clustering. We further demonstrate that our method outperforms existing approaches for unsupervised representation learning, and it even performs closely to supervised techniques in classification tasks. The contributions of this work are three-fold: ",
118
+ "bbox": [
119
+ 174,
120
+ 263,
121
+ 825,
122
+ 402
123
+ ],
124
+ "page_idx": 1
125
+ },
126
+ {
127
+ "type": "text",
128
+ "text": "1. We present a novel neighborhood-based unsupervised learning framework for non-stationary multivariate time series data. \n2. We introduce the concept of a temporal neighborhood with stationary properties as the distribution of similar windows in time. The neighborhood boundaries are determined automatically using the properties of the signal and statistical testing. \n3. We incorporate concepts from Positive Unlabeled Learning, specifically, sample weight adjustment, to account for potential bias introduced in sampling negative examples for the contrastive loss. ",
129
+ "bbox": [
130
+ 210,
131
+ 415,
132
+ 825,
133
+ 540
134
+ ],
135
+ "page_idx": 1
136
+ },
137
+ {
138
+ "type": "text",
139
+ "text": "2 METHOD ",
140
+ "text_level": 1,
141
+ "bbox": [
142
+ 174,
143
+ 561,
144
+ 282,
145
+ 578
146
+ ],
147
+ "page_idx": 1
148
+ },
149
+ {
150
+ "type": "text",
151
+ "text": "We introduce a framework for learning representations that encode the underlying state of a multivariate, non-stationary time series. Our self-supervised approach, TNC, takes advantage of the local smoothness of the generative process of signals to learn generalizable representations for windows of time series. This is done by ensuring that in the representation space, the distribution of signals proximal in time is distinguishable from the distribution of signals far away, i.e., proximity in time is identifiable in the encoding space. We represent our multivariate time series signals as $\\dot { X } \\in { \\cal R } ^ { D \\times T }$ , where D is the number of features and T is the number of measurements over time. X[t− δ ,t+ δ ] represents a window of time series of length $\\delta$ , centered around time $t$ , that includes measurements of all features taken in the interval $\\begin{array} { r } { [ t - \\frac { \\delta } { 2 } , t + \\frac { \\delta } { 2 } ] } \\end{array}$ . Throughout the paper, we refer to this window as $W _ { t }$ for notational simplicity. Our goal is to learn the underlying representation of $W _ { t }$ , and by sliding this window over time, we can obtain the trajectory of the underlying states of the signal. ",
152
+ "bbox": [
153
+ 174,
154
+ 594,
155
+ 825,
156
+ 750
157
+ ],
158
+ "page_idx": 1
159
+ },
160
+ {
161
+ "type": "text",
162
+ "text": "We define the temporal neighborhood $( N _ { t } )$ of a window $W _ { t }$ as the set of all windows with centroids $t ^ { * }$ , sampled from a normal distribution $t ^ { * } \\sim \\mathcal { N } ( t , \\eta \\cdot \\delta )$ . Where $\\mathcal { N }$ is a Gaussian centered at $t , \\delta$ is the size of the window, and $\\eta$ is the parameter that defines the range of the neighborhood. Relying on the local smoothness of a signal’s generative process, the neighborhood distribution is characterized as a Gaussian to model the gradual transition in temporal data, and intuitively, it approximates the distribution of samples that are similar to $W _ { t }$ . The $\\eta$ parameter determines the neighborhood range and depends on the signal characteristics and how gradual the time series’s statistical properties change over time. This can be set by domain experts based on prior knowledge of the signal behavior, or for more robust estimation, it can be determined by analyzing the stationarity properties of the signal for every $W _ { t }$ . Since the neighborhood represents similar samples, the range should identify the approximate time span within which the signal remains stationary, and the generative process does not change. For this purpose, we use the Augmented Dickey-Fuller (ADF) statistical test to determine this region for every window. Proper estimation of the neighborhood range is an integral part of the TNC framework. If $\\eta$ is too small, many samples from within a neighborhood will overlap, and therefore the encoder would only learn to encode the overlapping information. On the other hand, if $\\eta$ is too big, the neighborhood would span over multiple underlying states, and therefore the encoder would fail to distinguish the variation among these states. Using the ADF test, we can automatically adjust the neighborhood for every window based on the signal behavior. More details on this test and how it is used to estimate $\\eta$ is described in section 2. ",
163
+ "bbox": [
164
+ 174,
165
+ 757,
166
+ 825,
167
+ 924
168
+ ],
169
+ "page_idx": 1
170
+ },
171
+ {
172
+ "type": "text",
173
+ "text": "",
174
+ "bbox": [
175
+ 174,
176
+ 103,
177
+ 825,
178
+ 200
179
+ ],
180
+ "page_idx": 2
181
+ },
182
+ {
183
+ "type": "text",
184
+ "text": "Now, assuming windows within a neighborhood possess similar properties, signals outside of this neighborhood, denoted as $\\bar { N } _ { t }$ , are considered non-neighboring windows. Samples from $\\bar { N } _ { t }$ are likely to be different from $W _ { t }$ , and can be considered as negative samples in a context of a contrastive learning framework. However, this assumption can suffer from the problem of sampling bias, common in most contrastive learning approaches (Chuang et al., 2020; Saunshi et al., 2019). This bias occurs because randomly drawing negative examples from the data distribution may result in negative samples that are actually similar to the reference. This can significantly impact the learning framework’s performance, but little work has been done on addressing this issue (Chuang et al., 2020). In our context, this can happen when there are windows from $\\bar { N _ { t } }$ that are far away from $W _ { t }$ , but have the same underlying state. To alleviate this bias in the TNC framework, we consider samples from $\\bar { N } _ { t }$ as unlabeled samples, as opposed to negative, and use ideas from Positive-Unlabeled (PU) learning to accurately measure the loss function. In reality, even though samples within a neighborhood are all similar, we cannot make the assumption that samples outside this region are necessarily different. For instance, in the presence of long term seasonalities, signals can exhibit similar properties at distant times. In a healthcare context, this can be like a stable patient that undergoes a critical condition, but returns back to a stable state afterwards. ",
185
+ "bbox": [
186
+ 174,
187
+ 208,
188
+ 825,
189
+ 429
190
+ ],
191
+ "page_idx": 2
192
+ },
193
+ {
194
+ "type": "text",
195
+ "text": "In PU learning, a classifier is learned using labeled data drawn from the positive class $( P )$ and unlabeled data $( U )$ that is a mixture of positive and negative samples with a positive class prior $\\pi$ (Du Plessis et al., 2014; Kiryo et al., 2017; Du Plessis & Sugiyama, 2014). Existing PU learning methods fall under two categories based on how they handle the unlabeled data: 1) methods that identify negative samples from the unlabeled cohort (Li & Liu, 2003); 2) methods that treat the unlabeled data as negative samples with smaller weights (Lee & Liu, 2003; Elkan & Noto, 2008). In the second category, unlabeled samples should be properly weighted in the loss term in order to train an unbiased classifier. Elkan & Noto (2008) introduces a simple and efficient way of approximating the expectation of a loss function by assigning individual weights $w$ to training examples from the unlabeled cohort. This means each sample from the neighborhood is treated as a positive example with unit weight, while each sample from $\\bar { N }$ is treated as a combination of a positive example with weight $w$ and a negative example with complementary weight ${ 1 - w }$ . In the original paper (Elkan & Noto, 2008), the weight is defined as the probability for a sample from the unlabeled set to be a positive sample, i.e. $w = p ( y = 1 | x )$ for $x \\in U$ . In the TNC framework, this weight represents the probability of having samples similar to $W _ { t }$ in $\\bar { N }$ . By incorporating weight adjustment into the TNC loss (Equation 1), we account for possible positive samples that occur in the non-neighboring distribution. $w$ can be approximated using prior knowledge of the underlying state distribution or tuned as a hyperparameter. Appendix A.6 explains how the weight parameter is selected for our different experiment setups and also demonstrates the impact of weight adjustment on performance for downstream tasks. ",
196
+ "bbox": [
197
+ 173,
198
+ 435,
199
+ 825,
200
+ 714
201
+ ],
202
+ "page_idx": 2
203
+ },
204
+ {
205
+ "type": "text",
206
+ "text": "After defining the neighborhood distribution, we train an objective function that encourages a distinction between the representation of samples of the same neighborhood from the outside samples. An ideal encoder preserves the neighborhood properties in the encoding space. Therefore representations $Z _ { l } = E n c ( \\bar { W _ { l } } )$ of samples from a neighborhood $W _ { l } \\in N _ { t }$ , can be distinguished from representation $Z _ { k } = E n c ( W _ { k } )$ of samples from outside the neighborhood $W _ { k } \\in \\bar { N } _ { t }$ . TNC is composed of two main components: ",
207
+ "bbox": [
208
+ 174,
209
+ 722,
210
+ 825,
211
+ 805
212
+ ],
213
+ "page_idx": 2
214
+ },
215
+ {
216
+ "type": "text",
217
+ "text": "1. An Encoder $E n c ( W _ { t } )$ that maps $W _ { t } \\in \\mathbb { R } ^ { D \\times \\delta }$ to a representation $Z _ { t } \\in \\mathbb { R } ^ { M }$ , in a lower dimensional space $( M \\ll D \\times \\delta )$ , where $D \\times \\delta$ is the total number of measurements in $W _ { t }$ ",
218
+ "bbox": [
219
+ 207,
220
+ 830,
221
+ 825,
222
+ 859
223
+ ],
224
+ "page_idx": 2
225
+ },
226
+ {
227
+ "type": "text",
228
+ "text": "2. A Discriminator $\\mathcal { D } ( Z _ { t } , Z )$ that approximates the probability of $Z$ being the representation of a window in $N _ { t }$ . More specifically, it receives two samples from the encoding space and predicts the probability of those samples belonging to the same temporal neighborhood. ",
229
+ "bbox": [
230
+ 214,
231
+ 881,
232
+ 825,
233
+ 924
234
+ ],
235
+ "page_idx": 2
236
+ },
237
+ {
238
+ "type": "image",
239
+ "img_path": "images/398f866909f9701d5c0a6c7583ec244c49b4701a0c9562b08987de6f3e9de2a0.jpg",
240
+ "image_caption": [
241
+ "Figure 1: Overview of the TNC framework components. For each sample window $W _ { t }$ (indicated with the dashed black box), we first define the neighborhood distribution. The encoder learns the distribution of windows sampled from $N _ { t }$ and $\\bar { N } _ { t }$ , in the representation space. Then samples from Weightingthese distributions are fed into the discriminator alongside $Z _ { t }$ , to predict the probability of the Representationwindows being in the same neighborhood. "
242
+ ],
243
+ "image_footnote": [],
244
+ "bbox": [
245
+ 207,
246
+ 99,
247
+ 823,
248
+ 263
249
+ ],
250
+ "page_idx": 3
251
+ },
252
+ {
253
+ "type": "text",
254
+ "text": "TNC is a general framework; therefore, it is agnostic to the nature of the time series and the architecture of the encoder. The encoder can be any parametric model that is well-suited to the signal properties (Oord et al., 2016; Bai et al., 2018; Fawaz et al., 2019). For the Discriminator $\\mathcal { D } ( Z _ { t } , Z )$ we use a simple multi-headed binary classifier that outputs 1 if $Z$ and $Z _ { t }$ are representations of neighbors in time, and 0 otherwise. In the experiment section, we describe the architectural details of the models used for our experiments in more depth. ",
255
+ "bbox": [
256
+ 173,
257
+ 372,
258
+ 825,
259
+ 457
260
+ ],
261
+ "page_idx": 3
262
+ },
263
+ {
264
+ "type": "text",
265
+ "text": "Figure 1 provides a summary overview of the TNC framework. We formalize the objective function of our unsupervised learning framework in Equation 1. In essence, we would like the probability likelihood estimation of the Discriminator to be accurate, i.e., close to 1 for the representation of neighboring samples and close to 0 for windows far apart. Samples from the non-neighboring region $( \\bar { N } )$ are weight-adjusted using the $w$ parameters to account for positive samples in this distribution. ",
266
+ "bbox": [
267
+ 173,
268
+ 462,
269
+ 825,
270
+ 534
271
+ ],
272
+ "page_idx": 3
273
+ },
274
+ {
275
+ "type": "equation",
276
+ "img_path": "images/623cc7dc20a38141085e11093e43884cda000dc98a519ddbb8a58ccb501cd1e8.jpg",
277
+ "text": "$$\n\\begin{array} { r } { \\boldsymbol { \\Sigma } = - \\mathbb { E } _ { W _ { t } \\sim X } [ \\mathbb { E } _ { W _ { t } \\sim N _ { t } } [ \\log \\underbrace { \\mathcal { D } ( Z _ { t } , Z _ { l } ) } _ { \\mathrm { ~ } \\forall \\mathrm { ~ } W _ { t } \\sim \\bar { N } _ { t } } ] + \\mathbb { E } _ { W _ { t } \\sim \\bar { N } _ { t } } [ ( 1 - w _ { t } ) \\times \\log \\underbrace { ( 1 - \\mathcal { D } ( Z _ { t } , Z _ { k } ) ) } _ { \\mathcal { D } ( E n c ( W _ { t } ) , E n c ( W _ { k } ) ) } + w _ { t } \\times \\log \\mathcal { D } ( Z _ { t } , Z _ { k } ) ] } \\\\ { \\mathcal { D } ( E n c ( W _ { t } ) , E n c ( W _ { l } ) ) \\quad \\quad \\quad \\quad { \\mathcal { D } } ( E n c ( W _ { t } ) , E n c ( W _ { k } ) ) } \\end{array}\n$$",
278
+ "text_format": "latex",
279
+ "bbox": [
280
+ 181,
281
+ 559,
282
+ 839,
283
+ 609
284
+ ],
285
+ "page_idx": 3
286
+ },
287
+ {
288
+ "type": "text",
289
+ "text": "We train the encoder and the discriminator hand in hand by optimizing for this objective. Note that the Discriminator is only part of training and will not be used during inference. Similar to the encoder, it can be approximated using any parametric model. However, the more complex the Discriminator, the harder it becomes to interpret the latent space’s decision boundaries since it allows similarities to be mapped on complex nonlinear relationships. ",
290
+ "bbox": [
291
+ 173,
292
+ 630,
293
+ 826,
294
+ 700
295
+ ],
296
+ "page_idx": 3
297
+ },
298
+ {
299
+ "type": "text",
300
+ "text": "Defining the neighborhood parameter using the ADF test: As mentioned earlier, the neighborhood range can be specified using the characteristics of the data. In non-stationary time series, the generative process of the signals changes over time. We define the temporal neighborhood around every window as the region where the signal is relatively stationary. Since a signal may remain in an underlying state for an unknown amount of time, each window’s neighborhood range may vary in size and must be adjusted to signal behavior. To that end, we use the Augmented Dickey-Fuller (ADF) statistical test to derive the neighborhood range $\\eta$ . The ADF test belongs to a category of tests called \"Unit Root Test\", and is a method for testing the stationarity of a time series. For every $W _ { t }$ , we want to find the neighborhood range around that window that indicates a stationary region. To determine this, we start from $\\eta = 1$ and gradually increase the neighborhood size $\\eta$ , measuring the $p$ -value from the test at every step. Once $p$ -value is above a threshold (in our setting 0.01), it means that it fails to reject the null hypothesis and suggests that within this neighborhood region, the signal is no longer stationary. This way, we find the widest neighborhood within which the signal remains relatively stationary. Note that the window size $\\delta$ is constant throughout the experiment, and during ADF testing, we only adjust the neighborhood’s width. ",
301
+ "bbox": [
302
+ 173,
303
+ 715,
304
+ 826,
305
+ 924
306
+ ],
307
+ "page_idx": 3
308
+ },
309
+ {
310
+ "type": "text",
311
+ "text": "3 EXPERIMENTS ",
312
+ "text_level": 1,
313
+ "bbox": [
314
+ 176,
315
+ 102,
316
+ 326,
317
+ 117
318
+ ],
319
+ "page_idx": 4
320
+ },
321
+ {
322
+ "type": "text",
323
+ "text": "We evaluate our framework’s usability on multiple time series datasets with dynamic latent states that change over time. We compare classification performance and clusterability against two state-ofthe-art approaches for unsupervised representation learning for time series: 1. Contrastive Predictive Coding (CPC) (Oord et al., 2018) that uses predictive coding principles to train the encoder on a probabilistic contrastive loss. 2. Triplet-Loss (T-Loss), introduced in (Franceschi et al., 2019), which employs time-based negative sampling and a triplet loss to learn representations for time series windows. The triplet loss objective ensures similar time series have similar representations by minimizing the pairwise distance between positive samples (subseries) while maximizing it for negative ones. (See Appendix A.2 for more details on each baseline.) ",
324
+ "bbox": [
325
+ 174,
326
+ 135,
327
+ 825,
328
+ 258
329
+ ],
330
+ "page_idx": 4
331
+ },
332
+ {
333
+ "type": "text",
334
+ "text": "For a fair comparison and to ensure that the difference in performance is not due to the differences in the models’ architecture, the same encoder network is used across all compared baselines. Our objective is to compare the performance of the learning frameworks, agnostic to the encoder’s choice. Therefore, we selected simple architectures to evaluate how each framework can use a simple encoder’s limited capacity to learn meaningful representations. We assess the generalizability of the representations by 1) evaluating clusterability in the encoding space and 2) using the representations for a downstream classification task. In addition to the baselines mentioned above, we also compare clusterability performance with unsupervised K-means and classification with a K-Nearest Neighbor classifier, using Dynamic Time Warping (DTW) to measure time series distance. All models are implemented using Pytorch 1.3.1 and trained on a machine with Quadro 400 GPU 1. Below we describe the datasets for our experiments in more detail. ",
335
+ "bbox": [
336
+ 174,
337
+ 266,
338
+ 825,
339
+ 419
340
+ ],
341
+ "page_idx": 4
342
+ },
343
+ {
344
+ "type": "text",
345
+ "text": "3.1 SIMULATED DATA ",
346
+ "text_level": 1,
347
+ "bbox": [
348
+ 176,
349
+ 438,
350
+ 338,
351
+ 452
352
+ ],
353
+ "page_idx": 4
354
+ },
355
+ {
356
+ "type": "text",
357
+ "text": "The simulated dataset is designed to replicate very long, non-stationary, and high-frequency time series for which the underlying dynamics change over time. Our generated time series consists of 2000 measurements for 3 features, generated from 4 different underlying states. We use a Hidden Markov Model (HMM) to generate the random latent states over time, and in each state, the time series is generated from a different generative process, including Gaussian Processes (GPs) with different kernel functions and Nonlinear Auto-regressive Moving Average models with different sets of parameters $\\overset { \\cdot } { \\alpha }$ and $\\beta$ ). Besides, for it to further resemble realistic (e.g., clinical) time series, two features are always correlated. More details about this dataset are provided in the Appendix A.1. For this experimental setup, we use a two-directional, single-layer recurrent neural network encoder. We have selected this simple architecture because it handles time series with variable lengths, and it easily extends to higher-dimensional inputs. The encoder model encodes multi-dimensional signal windows of $\\delta = 5 0$ into 10 dimensional representation vectors. The window size is selected such that it is long enough to contain information about the underlying state but not too long to span over multiple underlying states. A more detailed discussion on the window size choice is presented in Appendix A.4. ",
358
+ "bbox": [
359
+ 174,
360
+ 464,
361
+ 825,
362
+ 672
363
+ ],
364
+ "page_idx": 4
365
+ },
366
+ {
367
+ "type": "text",
368
+ "text": "3.2 CLINICAL WAVEFORM DATA ",
369
+ "text_level": 1,
370
+ "bbox": [
371
+ 176,
372
+ 691,
373
+ 406,
374
+ 704
375
+ ],
376
+ "page_idx": 4
377
+ },
378
+ {
379
+ "type": "text",
380
+ "text": "For a real-world clinical experiment, we use the MIT-BIH Atrial Fibrillation dataset (Moody, 1983). This dataset includes 25 long-term Electrocardiogram (ECG) recordings (10 hours in duration) of human subjects with atrial fibrillation. It consists of two ECG signals; each sampled at $2 5 0 \\mathrm { H z }$ . The signals are annotated over time for the following different rhythm types: 1) Atrial fibrillation, 2) Atrial flutter, 3) AV junctional rhythm, and 4) all other rhythms. Our goal in this experiment is to identify the underlying type of arrhythmia for each sample without any information about the labels. This dataset is particularly interesting and makes this experiment challenging due to the following special properties: ",
381
+ "bbox": [
382
+ 173,
383
+ 717,
384
+ 825,
385
+ 829
386
+ ],
387
+ "page_idx": 4
388
+ },
389
+ {
390
+ "type": "text",
391
+ "text": "• The underlying heart rhythm changes over time in each sample. This is an opportunity to evaluate how different representation learning frameworks can handle alternating classes in non-stationary settings; ",
392
+ "bbox": [
393
+ 212,
394
+ 843,
395
+ 823,
396
+ 886
397
+ ],
398
+ "page_idx": 4
399
+ },
400
+ {
401
+ "type": "text",
402
+ "text": "• The dataset is highly imbalanced, with atrial flutter and AV junctional rhythm being present in fewer than $0 . 1 \\%$ of the measurements. Data imbalance poses many challenges for downstream classification, further motivating the use of unsupervised representation learning; The dataset has samples from a small number of individuals, but over an extended period (around 5 million data points). This realistic scenario, common in healthcare data, shows that our framework is still powerful in settings with a limited number of samples. ",
403
+ "bbox": [
404
+ 215,
405
+ 103,
406
+ 826,
407
+ 193
408
+ ],
409
+ "page_idx": 5
410
+ },
411
+ {
412
+ "type": "text",
413
+ "text": "The simple RNN encoder architecture used for other experiment setups cannot model the highfrequency ECG measurements. Therefore, inspired by state-of-the-art architectures for ECG classification problems, the encoder Enc used in this experiment is a 2-channel, 1-dimensional strided convolutional neural network that runs directly on the ECG waveforms. We use six convolutional layers with a total down-sampling factor of 16. The window size is 2500 samples, meaning that each convolutional filter covers at least half a second of ECG recording, and the representations are summarized in a 64-dimensional vector. ",
414
+ "bbox": [
415
+ 174,
416
+ 207,
417
+ 825,
418
+ 304
419
+ ],
420
+ "page_idx": 5
421
+ },
422
+ {
423
+ "type": "text",
424
+ "text": "3.3 HUMAN ACTIVITY RECOGNITION (HAR) DATA ",
425
+ "text_level": 1,
426
+ "bbox": [
427
+ 174,
428
+ 321,
429
+ 542,
430
+ 337
431
+ ],
432
+ "page_idx": 5
433
+ },
434
+ {
435
+ "type": "text",
436
+ "text": "Human Activity Recognition (HAR) is the problem of predicting the type of activity using temporal data from accelerometer and gyroscope measurements. We use the HAR dataset from the UCI Machine Learning Repository 2 that includes data collected from 30 individuals using a smartwatch. Each person performs six activities: 1) walking, 2) walking upstairs, 3) walking downstairs, 4) sitting, 5) standing, and 6) laying. The time-series measurements are pre-processed to extract 561 features. For our purpose, we concatenate the activity samples from every individual over time using the subject identifier to build the full-time series for each subject, which includes continuous activity change. Similar to the simulated data setting, we use a single-layer RNN encoder. The selected window size is 4, representing about 15 seconds of recording, and the representations are encoded in a 10-dimensional vector space. ",
437
+ "bbox": [
438
+ 174,
439
+ 348,
440
+ 825,
441
+ 488
442
+ ],
443
+ "page_idx": 5
444
+ },
445
+ {
446
+ "type": "text",
447
+ "text": "4 RESULTS ",
448
+ "text_level": 1,
449
+ "bbox": [
450
+ 176,
451
+ 508,
452
+ 281,
453
+ 525
454
+ ],
455
+ "page_idx": 5
456
+ },
457
+ {
458
+ "type": "text",
459
+ "text": "In this section we present the results for clusterability of the latent representations and downstream classification performance for all datasets and across all baselines. Clusterability indicates how well each method recovers appropriate states, and classification assesses how informative our representations are for downstream tasks. ",
460
+ "bbox": [
461
+ 174,
462
+ 541,
463
+ 825,
464
+ 597
465
+ ],
466
+ "page_idx": 5
467
+ },
468
+ {
469
+ "type": "text",
470
+ "text": "4.1 EVALUATION: CLUSTERABILITY ",
471
+ "text_level": 1,
472
+ "bbox": [
473
+ 176,
474
+ 616,
475
+ 437,
476
+ 630
477
+ ],
478
+ "page_idx": 5
479
+ },
480
+ {
481
+ "type": "text",
482
+ "text": "Many real-world time series data have underlying multi-category structure, naturally leading to representations with clustering properties. Encoding such general priors is a property of a good representation (Bengio et al., 2013). In this section, we assess the distribution of the representations in the encoding space. If information of the latent state is properly learned and encoded by the framework, the representations of signals from the same underlying state should cluster together. Figures 2a, 2b, and 2c show an example of this distribution for simulated data across compared approaches. Each plot is a 2-dimensional t-SNE visualization of the representations where each data point in the scatter plot is an encoding $Z \\in R ^ { 1 0 }$ that represents a window of size $\\delta = 5 0$ of a simulated time series. We can see that without any information about the hidden states, representations learned using TNC cluster windows from the same hidden state better than the alternative approaches. The results show that CPC and Triplet Loss have difficulty separating time series that are generated from non-linear auto-regressive moving average (NARMA) models with variable regression parameters. ",
483
+ "bbox": [
484
+ 174,
485
+ 642,
486
+ 825,
487
+ 809
488
+ ],
489
+ "page_idx": 5
490
+ },
491
+ {
492
+ "type": "text",
493
+ "text": "To compare the representation clusters’ consistency for each baseline, we use two very common cluster validity indices, namely, the Silhouette score and the Davies-Bouldin index. We use K-means clustering in the representation space to measure these clusterability scores. The Silhouette score measures the similarity of each sample to its own cluster, compared to other clusters. The values can range from $- 1$ to $+ 1$ , and a greater score implies a better cohesion. The Davies-Bouldin Index measures intra-cluster similarity and inter-cluster differences. This is a positive index score, where smaller values indicate low within-cluster scatter and large separation between clusters. Therefore, a lower score represents better clusterability (more details on the cluster validity scores and how they are calculated can be found in Appendix A.5). ",
494
+ "bbox": [
495
+ 174,
496
+ 815,
497
+ 823,
498
+ 886
499
+ ],
500
+ "page_idx": 5
501
+ },
502
+ {
503
+ "type": "image",
504
+ "img_path": "images/1e54dcf85574cebb36161a9083e97999516e415efc834be6219305e5d274ca9d.jpg",
505
+ "image_caption": [
506
+ "Figure 2: T-SNE visualization of signal representations for the simulated dataset across all baselines. Each data point in the plot presents a 10-dimensional representation of a window of time series of size $\\delta = 5 0$ , and the color indicates the latent state of the signal window. See Appendix A.7 for similar plots from different datasets. "
507
+ ],
508
+ "image_footnote": [],
509
+ "bbox": [
510
+ 184,
511
+ 106,
512
+ 823,
513
+ 262
514
+ ],
515
+ "page_idx": 6
516
+ },
517
+ {
518
+ "type": "text",
519
+ "text": "",
520
+ "bbox": [
521
+ 174,
522
+ 354,
523
+ 825,
524
+ 410
525
+ ],
526
+ "page_idx": 6
527
+ },
528
+ {
529
+ "type": "table",
530
+ "img_path": "images/3d25f460d8db1110a5d2945507df7eed301335ca67a1e769cba171e705ca2fb4.jpg",
531
+ "table_caption": [],
532
+ "table_footnote": [],
533
+ "table_body": "<table><tr><td></td><td colspan=\"2\">Simulation</td><td colspan=\"2\">ECG Waveform</td><td colspan=\"2\">HAR</td></tr><tr><td>Method</td><td>Silhouette 个</td><td>DBI↓</td><td>Silhouette 个</td><td>DBI↓</td><td>Silhouette↑</td><td>DBI↓</td></tr><tr><td>TNC</td><td>0.71±0.01</td><td>0.36±0.01</td><td>0.44±0.02</td><td>0.74±0.04</td><td>0.61±0.02</td><td>0.52±0.04</td></tr><tr><td>CPC</td><td>0.51±0.03</td><td>0.84±0.06</td><td>0.26±0.02</td><td>1.44±0.04</td><td>0.58±0.02</td><td>0.57±0.05</td></tr><tr><td>T-Loss</td><td>0.61±0.08</td><td>0.64±0.12</td><td>0.25±0.01</td><td>1.30±0.03</td><td>0.17±0.01</td><td>1.76±0.20</td></tr><tr><td>K-means</td><td>0.01±0.019</td><td>7.23±0.14</td><td>0.19±0.11</td><td>3.65±0.48</td><td>0.12±0.40</td><td>2.66±0.05</td></tr></table>",
534
+ "bbox": [
535
+ 176,
536
+ 421,
537
+ 823,
538
+ 529
539
+ ],
540
+ "page_idx": 6
541
+ },
542
+ {
543
+ "type": "text",
544
+ "text": "Table 1: Clustering quality of representations in the encoding space for multiple datasets. ",
545
+ "bbox": [
546
+ 200,
547
+ 539,
548
+ 790,
549
+ 553
550
+ ],
551
+ "page_idx": 6
552
+ },
553
+ {
554
+ "type": "text",
555
+ "text": "Table 1 summarizes the scores for all baselines and across all datasets, demonstrating that TNC is superior in learning representations that can distinguish the latent dynamics of time series. CPC performs closely to Triplet loss on waveform data but performs poorly on the simulated dataset, where signals are highly non-stationary, and transitions are less predictable. However, for the HAR dataset, CPC clusters the states very well because most activities are recorded in a specific order, empowering predictive coding. Triplet loss performs reasonably well in the simulated setting; however, it fails to distinguish states 0 and 2, where signals come from autoregressive models with different parameters and have a relatively similar generative process. Performing K-means on the original time series generally does not generate coherent clusters, as demonstrated by the scores. However, the performance is slightly better in time series like the ECG waveforms, where the signals are formed by consistent shapelets, and therefore the DTW measures similarity more accurately. ",
556
+ "bbox": [
557
+ 173,
558
+ 568,
559
+ 826,
560
+ 720
561
+ ],
562
+ "page_idx": 6
563
+ },
564
+ {
565
+ "type": "text",
566
+ "text": "4.2 EVALUATION: CLASSIFICATION ",
567
+ "text_level": 1,
568
+ "bbox": [
569
+ 176,
570
+ 737,
571
+ 431,
572
+ 751
573
+ ],
574
+ "page_idx": 6
575
+ },
576
+ {
577
+ "type": "text",
578
+ "text": "We further evaluate the quality of the encodings using a classification task. We train a linear classifier to evaluate how well the representations can be used to classify hidden states. The performance of all baselines is compared to a supervised classifier, composed of an encoder and a classifier with identical architectures to that of the unsupervised models, and a K-nearest neighbor classifier that uses DTW metric. The performance is reported as the prediction accuracy and the area under the precision-recall curve (AUPRC) score since AUPRC is a more accurate reflection of model performance for imbalance classification settings like the waveform dataset. ",
579
+ "bbox": [
580
+ 174,
581
+ 763,
582
+ 825,
583
+ 861
584
+ ],
585
+ "page_idx": 6
586
+ },
587
+ {
588
+ "type": "text",
589
+ "text": "Table 2 demonstrates the classification performance for all datasets. The performance of the classifiers that use TNC representations are closer to the end-to-end supervised model in comparison to CPC and Triplet Loss. This provides further evidence that our encodings capture informative parts of the time series and are generalizable to be used for downstream tasks. In datasets like the HAR, where an inherent ordering usually exists in the time series, CPC performs reasonably. However, in datasets with increased non-stationarity, the performance drops. Triplet Loss is also a powerful framework, but since it samples positive examples from overlapping windows of time series, it is vulnerable to map the overlaps into the encoding and, therefore, fail to learn more general representations. TNC, on the other hand, samples similar windows from a wider distribution, defined by the temporal neighborhood, where many of the neighboring signals do not necessarily overlap. The lower performance of the CPC and Triplet Loss methods can also be partly because none of these methods explicitly account for the potential sampling bias that happens when randomly selected negative examples are similar to the reference $W _ { t }$ . ",
590
+ "bbox": [
591
+ 174,
592
+ 867,
593
+ 823,
594
+ 924
595
+ ],
596
+ "page_idx": 6
597
+ },
598
+ {
599
+ "type": "table",
600
+ "img_path": "images/39f9c9c8e816efdeff9b00fea5b1ac99567a485179e4d9967b11f1d540db7df0.jpg",
601
+ "table_caption": [],
602
+ "table_footnote": [
603
+ "Table 2: Performance of all baselines in classifying the underlying hidden states of the time series, measured as the accuracy and AUPRC score. "
604
+ ],
605
+ "table_body": "<table><tr><td colspan=\"3\">Simulation</td><td colspan=\"2\">ECG Waveform</td><td colspan=\"2\">HAR</td></tr><tr><td>Method</td><td>AUPRC</td><td>Accuracy</td><td>AUPRC</td><td>Accuracy</td><td>AUPRC</td><td>Accuracy</td></tr><tr><td>TNC</td><td>0.99±0.00</td><td>97.52±0.13</td><td>0.55±0.01</td><td>77.79±0.84</td><td>0.94±0.007</td><td>88.32±0.12</td></tr><tr><td>CPC</td><td>0.69±0.06</td><td>70.26±6.48</td><td>0.42±0.01</td><td>68.64±0.49</td><td>0.93±0.006</td><td>86.43±1.41</td></tr><tr><td>T-Loss</td><td>0.78±0.01</td><td>76.66±1.40</td><td>0.47±0.00</td><td>75.51±1.26</td><td>0.71±0.007</td><td>63.60±3.37</td></tr><tr><td>KNN</td><td>0.42±0.00</td><td>55.53±0.65</td><td>0.38±0.06</td><td>54.76±5.46</td><td>0.75±0.01</td><td>84.85±0.84</td></tr><tr><td>Supervised</td><td>0.99±0.00</td><td>98.56±0.13</td><td>0.67±0.01</td><td>94.81±0.28</td><td>0.98±0.00</td><td>92.03±2.48</td></tr></table>",
606
+ "bbox": [
607
+ 171,
608
+ 99,
609
+ 838,
610
+ 227
611
+ ],
612
+ "page_idx": 7
613
+ },
614
+ {
615
+ "type": "text",
616
+ "text": "",
617
+ "bbox": [
618
+ 173,
619
+ 300,
620
+ 825,
621
+ 426
622
+ ],
623
+ "page_idx": 7
624
+ },
625
+ {
626
+ "type": "text",
627
+ "text": "4.3 EVALUATION: TRAJECTORY ",
628
+ "text_level": 1,
629
+ "bbox": [
630
+ 176,
631
+ 452,
632
+ 406,
633
+ 467
634
+ ],
635
+ "page_idx": 7
636
+ },
637
+ {
638
+ "type": "image",
639
+ "img_path": "images/57ec11acc26cc0da64f569141bc5f07e4d7032734b4804873f3d1735a8eacdcc.jpg",
640
+ "image_caption": [
641
+ "Figure 3: Trajectory of a signal encoding from the simulated dataset. The top plot shows the original time series with shaded regions indicating the underlying state. The bottom plot shows the 10 dimensional encoding of the sliding windows $W _ { t }$ where $\\delta = 5 0$ . "
642
+ ],
643
+ "image_footnote": [],
644
+ "bbox": [
645
+ 184,
646
+ 493,
647
+ 815,
648
+ 675
649
+ ],
650
+ "page_idx": 7
651
+ },
652
+ {
653
+ "type": "text",
654
+ "text": "This section investigates the trajectories of our learned encodings over time to understand how the state transitions are captured and modeled in the representation space. This is an important property for non-stationary time series where underlying states change over time, and capturing those changes is critical in many application domains such as healthcare. Figure 3 shows a sample from the simulated dataset. The top panel shows the signal measurements over time, and the shaded regions indicate the underlying latent states. The bottom panel illustrates the 10-dimensional representation of a sliding window $W _ { t }$ estimated over time. From the bottom panel of Figure 3, we can see that the encoding pattern changes at state transitions and settle into a different pattern, corresponding to the new state. This change happens at every transition, and we can see the distinct patterns for all 4 underlying states in the representations. This analysis of the trajectory of change could be very informative for the users’ post-analysis; for instance, in clinical applications, it could help clinicians visualize the evolution of the patient state over time and plan treatment based on the state progression ",
655
+ "bbox": [
656
+ 173,
657
+ 757,
658
+ 825,
659
+ 924
660
+ ],
661
+ "page_idx": 7
662
+ },
663
+ {
664
+ "type": "text",
665
+ "text": "5 RELATED WORK ",
666
+ "text_level": 1,
667
+ "bbox": [
668
+ 176,
669
+ 102,
670
+ 339,
671
+ 117
672
+ ],
673
+ "page_idx": 8
674
+ },
675
+ {
676
+ "type": "text",
677
+ "text": "While integral for many applications, unsupervised representation learning has been far less studied for time series (Längkvist et al., 2014), compared to other domains such as vision or natural language processing (Denton et al., 2017; Radford et al., 2015; Gutmann & Hyvärinen, 2012; Wang & Gupta, 2015). One of the earliest approaches to unsupervised end-to-end representation learning in time series is the use of auto-encoders (Choi et al., 2016a; Amiriparian et al., 2017; Malhotra et al., 2017) and seq-to-seq models (Lyu et al., 2018), with the objective to train an encoder jointly with a decoder that reconstructs the input signal from its learned representation. Using fully generative models like variational auto-encoders is also useful for imposing properties like disentanglement, which help with the interpretability of the representations (Dezfouli et al., 2019). However, in many cases, like for high-frequency physiological signals, the reconstruction of complex time series can be challenging; therefore, more novel approaches are designed to avoid this step. Contrastive Predictive Coding (Oord et al., 2018; Löwe et al., 2019) learns representations by predicting the future in latent space, eliminating the need to reconstruct the full input. The representations are such that the mutual information between the original signal and the concept vector is maximally preserve using a lower bound approximation and a contrastive loss. Very similarly, in Time Contrastive Learning (Hyvarinen & Morioka, 2016), a contrastive loss is used to predict the segment-ID of multivariate time-series as a way to extract representation. Franceschi et al. (2019) employs time-based negative sampling and a triplet loss to learn scalable representations for multivariate time series. Some other approaches use inherent similarities in temporal data to learn representations without supervision. For instance, in similarity-preserving representation learning (Lei et al., 2019), learned encodings are constrained to preserve the pairwise similarities that exist in the time domain, measured by DTW distance. Another group of approaches combines reconstruction loss with clustering objectives to cluster similar temporal patterns in the encoding space (Ma et al., 2019). ",
678
+ "bbox": [
679
+ 173,
680
+ 140,
681
+ 825,
682
+ 459
683
+ ],
684
+ "page_idx": 8
685
+ },
686
+ {
687
+ "type": "text",
688
+ "text": "In healthcare, learning representation of rich temporal medical data is extremely important for understanding patients’ underlying health conditions. However, most of the existing approaches for learning representations are designed for specific downstream tasks and require labeling by experts (Choi et al., 2016b;c; Tonekaboni et al., 2020). Examples of similar works to representation learning in the field of clinical ML include computational phenotyping for discovering subgroups of patients with similar underlying disease mechanisms from temporal clinical data (Lasko et al., 2013; Suresh et al., 2018; Schulam et al., 2015), and disease progression modeling, for learning the hidden vector of comorbidities representing a disease over time (Wang et al., 2014; Alaa & van der Schaar, 2019). ",
689
+ "bbox": [
690
+ 174,
691
+ 465,
692
+ 825,
693
+ 577
694
+ ],
695
+ "page_idx": 8
696
+ },
697
+ {
698
+ "type": "text",
699
+ "text": "6 CONCLUSION ",
700
+ "text_level": 1,
701
+ "bbox": [
702
+ 176,
703
+ 608,
704
+ 318,
705
+ 625
706
+ ],
707
+ "page_idx": 8
708
+ },
709
+ {
710
+ "type": "text",
711
+ "text": "This paper presents a novel unsupervised representation learning framework for complex multivariate time series, called Temporal Neighborhood Coding (TNC). This framework is designed to learn the underlying dynamics of non-stationary signals and to model the progression over time by defining a temporal neighborhood. The problem is motivated by the medical field, where patients transition between distinct clinical states over time, and obtaining labels to define these underlying states is challenging. We evaluate the performance of TNC on multiple datasets and show that our representations are generalizable and can easily be used for diverse tasks such as classification and clustering. We finally note that TNC is flexible to be used with arbitrary encoder architectures; therefore, the framework is applicable to many time series data domains. Moreover, in addition to tasks presented in this paper, general representations can be used for several other downstream tasks, such as anomaly detection, which is challenging in supervised learning settings for time series data in sparsely labeled contexts. ",
712
+ "bbox": [
713
+ 174,
714
+ 646,
715
+ 825,
716
+ 813
717
+ ],
718
+ "page_idx": 8
719
+ },
720
+ {
721
+ "type": "text",
722
+ "text": "ACKNOWLEDGMENTS ",
723
+ "text_level": 1,
724
+ "bbox": [
725
+ 176,
726
+ 840,
727
+ 326,
728
+ 853
729
+ ],
730
+ "page_idx": 8
731
+ },
732
+ {
733
+ "type": "text",
734
+ "text": "Resources used in preparing this research were provided, in part, by the Government of Canada through CIFAR, and companies sponsoring the Vector Institute. This research was undertaken, in part, thanks to funding from the Canadian Institute of Health Research (CIHR) and the Natural Sciences and Engineering Research Council of Canada (NSERC). ",
735
+ "bbox": [
736
+ 176,
737
+ 867,
738
+ 825,
739
+ 922
740
+ ],
741
+ "page_idx": 8
742
+ },
743
+ {
744
+ "type": "text",
745
+ "text": "REFERENCES ",
746
+ "text_level": 1,
747
+ "bbox": [
748
+ 176,
749
+ 102,
750
+ 287,
751
+ 117
752
+ ],
753
+ "page_idx": 9
754
+ },
755
+ {
756
+ "type": "text",
757
+ "text": "Ahmed Alaa and Mihaela van der Schaar. Attentive state-space modeling of disease progression. 2019. ",
758
+ "bbox": [
759
+ 173,
760
+ 126,
761
+ 823,
762
+ 154
763
+ ],
764
+ "page_idx": 9
765
+ },
766
+ {
767
+ "type": "text",
768
+ "text": "Shahin Amiriparian, Michael Freitag, Nicholas Cummins, and Björn Schuller. Sequence to sequence autoencoders for unsupervised representation learning from audio. In Proc. of the DCASE 2017 Workshop, 2017. ",
769
+ "bbox": [
770
+ 174,
771
+ 162,
772
+ 826,
773
+ 205
774
+ ],
775
+ "page_idx": 9
776
+ },
777
+ {
778
+ "type": "text",
779
+ "text": "Shaojie Bai, J Zico Kolter, and Vladlen Koltun. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv preprint arXiv:1803.01271, 2018. ",
780
+ "bbox": [
781
+ 173,
782
+ 213,
783
+ 825,
784
+ 242
785
+ ],
786
+ "page_idx": 9
787
+ },
788
+ {
789
+ "type": "text",
790
+ "text": "Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence, 35(8):1798–1828, 2013. ",
791
+ "bbox": [
792
+ 174,
793
+ 251,
794
+ 823,
795
+ 292
796
+ ],
797
+ "page_idx": 9
798
+ },
799
+ {
800
+ "type": "text",
801
+ "text": "Edward Choi, Mohammad Taha Bahadori, Elizabeth Searles, Catherine Coffey, Michael Thompson, James Bost, Javier Tejedor-Sojo, and Jimeng Sun. Multi-layer representation learning for medical concepts. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 1495–1504, 2016a. ",
802
+ "bbox": [
803
+ 173,
804
+ 301,
805
+ 826,
806
+ 359
807
+ ],
808
+ "page_idx": 9
809
+ },
810
+ {
811
+ "type": "text",
812
+ "text": "Edward Choi, Andy Schuetz, Walter F Stewart, and Jimeng Sun. Medical concept representation learning from electronic health records and its application on heart failure prediction. arXiv preprint arXiv:1602.03686, 2016b. ",
813
+ "bbox": [
814
+ 174,
815
+ 366,
816
+ 825,
817
+ 409
818
+ ],
819
+ "page_idx": 9
820
+ },
821
+ {
822
+ "type": "text",
823
+ "text": "Youngduck Choi, Chill Yi-I Chiu, and David Sontag. Learning low-dimensional representations of medical concepts. AMIA Summits on Translational Science Proceedings, 2016:41, 2016c. ",
824
+ "bbox": [
825
+ 171,
826
+ 416,
827
+ 825,
828
+ 446
829
+ ],
830
+ "page_idx": 9
831
+ },
832
+ {
833
+ "type": "text",
834
+ "text": "Ching-Yao Chuang, Joshua Robinson, Lin Yen-Chen, Antonio Torralba, and Stefanie Jegelka. Debiased contrastive learning. arXiv preprint arXiv:2007.00224, 2020. ",
835
+ "bbox": [
836
+ 171,
837
+ 454,
838
+ 825,
839
+ 484
840
+ ],
841
+ "page_idx": 9
842
+ },
843
+ {
844
+ "type": "text",
845
+ "text": "Emily L Denton et al. Unsupervised learning of disentangled representations from video. In Advances in Neural Information Processing Systems, pp. 4414–4423, 2017. ",
846
+ "bbox": [
847
+ 171,
848
+ 491,
849
+ 825,
850
+ 521
851
+ ],
852
+ "page_idx": 9
853
+ },
854
+ {
855
+ "type": "text",
856
+ "text": "Amir Dezfouli, Hassan Ashtiani, Omar Ghattas, Richard Nock, Peter Dayan, and Cheng Soon Ong. Disentangled behavioural representations. In Advances in Neural Information Processing Systems, pp. 2251–2260, 2019. ",
857
+ "bbox": [
858
+ 174,
859
+ 529,
860
+ 826,
861
+ 571
862
+ ],
863
+ "page_idx": 9
864
+ },
865
+ {
866
+ "type": "text",
867
+ "text": "Jeff Donahue and Karen Simonyan. Large scale adversarial representation learning. In Advances in Neural Information Processing Systems, pp. 10541–10551, 2019. ",
868
+ "bbox": [
869
+ 171,
870
+ 579,
871
+ 823,
872
+ 609
873
+ ],
874
+ "page_idx": 9
875
+ },
876
+ {
877
+ "type": "text",
878
+ "text": "Marthinus C Du Plessis, Gang Niu, and Masashi Sugiyama. Analysis of learning from positive and unlabeled data. In Advances in Neural Information Processing Systems, pp. 703–711, 2014. ",
879
+ "bbox": [
880
+ 173,
881
+ 616,
882
+ 825,
883
+ 646
884
+ ],
885
+ "page_idx": 9
886
+ },
887
+ {
888
+ "type": "text",
889
+ "text": "Marthinus Christoffel Du Plessis and Masashi Sugiyama. Class prior estimation from positive and unlabeled data. IEICE TRANSACTIONS on Information and Systems, 97(5):1358–1362, 2014. ",
890
+ "bbox": [
891
+ 171,
892
+ 654,
893
+ 823,
894
+ 684
895
+ ],
896
+ "page_idx": 9
897
+ },
898
+ {
899
+ "type": "text",
900
+ "text": "Charles Elkan and Keith Noto. Learning classifiers from only positive and unlabeled data. In Proceedings of the 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 213–220, 2008. ",
901
+ "bbox": [
902
+ 176,
903
+ 690,
904
+ 825,
905
+ 734
906
+ ],
907
+ "page_idx": 9
908
+ },
909
+ {
910
+ "type": "text",
911
+ "text": "Hassan Ismail Fawaz, Germain Forestier, Jonathan Weber, Lhassane Idoumghar, and Pierre-Alain Muller. Deep learning for time series classification: a review. Data Mining and Knowledge Discovery, 33(4):917–963, 2019. ",
912
+ "bbox": [
913
+ 174,
914
+ 742,
915
+ 825,
916
+ 785
917
+ ],
918
+ "page_idx": 9
919
+ },
920
+ {
921
+ "type": "text",
922
+ "text": "Jean-Yves Franceschi, Aymeric Dieuleveut, and Martin Jaggi. Unsupervised scalable representation learning for multivariate time series. In Advances in Neural Information Processing Systems, pp. 4652–4663, 2019. ",
923
+ "bbox": [
924
+ 171,
925
+ 792,
926
+ 825,
927
+ 835
928
+ ],
929
+ "page_idx": 9
930
+ },
931
+ {
932
+ "type": "text",
933
+ "text": "Michael U Gutmann and Aapo Hyvärinen. Noise-contrastive estimation of unnormalized statistical models, with applications to natural image statistics. Journal of Machine Learning Research, 13 (Feb):307–361, 2012. ",
934
+ "bbox": [
935
+ 173,
936
+ 844,
937
+ 825,
938
+ 887
939
+ ],
940
+ "page_idx": 9
941
+ },
942
+ {
943
+ "type": "text",
944
+ "text": "Aapo Hyvarinen and Hiroshi Morioka. Unsupervised feature extraction by time-contrastive learning and nonlinear ICA. In Advances in Neural Information Processing Systems, pp. 3765–3773, 2016. ",
945
+ "bbox": [
946
+ 174,
947
+ 895,
948
+ 823,
949
+ 924
950
+ ],
951
+ "page_idx": 9
952
+ },
953
+ {
954
+ "type": "text",
955
+ "text": "Ryuichi Kiryo, Gang Niu, Marthinus C Du Plessis, and Masashi Sugiyama. Positive-unlabeled learning with non-negative risk estimator. In Advances in Neural Information Processing Systems, pp. 1675–1685, 2017. ",
956
+ "bbox": [
957
+ 173,
958
+ 103,
959
+ 823,
960
+ 146
961
+ ],
962
+ "page_idx": 10
963
+ },
964
+ {
965
+ "type": "text",
966
+ "text": "Martin Längkvist, Lars Karlsson, and Amy Loutfi. A review of unsupervised feature learning and deep learning for time-series modeling. Pattern Recognition Letters, 42:11–24, 2014. ",
967
+ "bbox": [
968
+ 173,
969
+ 155,
970
+ 823,
971
+ 184
972
+ ],
973
+ "page_idx": 10
974
+ },
975
+ {
976
+ "type": "text",
977
+ "text": "Thomas A Lasko, Joshua C Denny, and Mia A Levy. Computational phenotype discovery using unsupervised feature learning over noisy, sparse, and irregular clinical data. PloS one, 8(6), 2013. ",
978
+ "bbox": [
979
+ 173,
980
+ 193,
981
+ 825,
982
+ 222
983
+ ],
984
+ "page_idx": 10
985
+ },
986
+ {
987
+ "type": "text",
988
+ "text": "Wee Sun Lee and Bing Liu. Learning with positive and unlabeled examples using weighted logistic regression. In International Conference on Machine Learning, volume 3, pp. 448–455, 2003. ",
989
+ "bbox": [
990
+ 173,
991
+ 231,
992
+ 825,
993
+ 260
994
+ ],
995
+ "page_idx": 10
996
+ },
997
+ {
998
+ "type": "text",
999
+ "text": "Qi Lei, Jinfeng Yi, Roman Vaculin, Lingfei Wu, and Inderjit S Dhillon. Similarity preserving representation learning for time series clustering. In IJCAI, volume 19, pp. 2845–2851, 2019. ",
1000
+ "bbox": [
1001
+ 176,
1002
+ 267,
1003
+ 821,
1004
+ 297
1005
+ ],
1006
+ "page_idx": 10
1007
+ },
1008
+ {
1009
+ "type": "text",
1010
+ "text": "Xiaoli Li and Bing Liu. Learning to classify texts using positive and unlabeled data. In IJCAI, volume 3, pp. 587–592, 2003. ",
1011
+ "bbox": [
1012
+ 173,
1013
+ 306,
1014
+ 825,
1015
+ 335
1016
+ ],
1017
+ "page_idx": 10
1018
+ },
1019
+ {
1020
+ "type": "text",
1021
+ "text": "Sindy Löwe, Peter O’Connor, and Bastiaan Veeling. Putting an end to end-to-end: Gradient-isolated learning of representations. In Advances in Neural Information Processing Systems, pp. 3039–3051, 2019. ",
1022
+ "bbox": [
1023
+ 176,
1024
+ 343,
1025
+ 823,
1026
+ 386
1027
+ ],
1028
+ "page_idx": 10
1029
+ },
1030
+ {
1031
+ "type": "text",
1032
+ "text": "Xinrui Lyu, Matthias Hueser, Stephanie L Hyland, George Zerveas, and Gunnar Rätsch. Improving clinical predictions through unsupervised time series representation learning. arXiv preprint arXiv:1812.00490, 2018. ",
1033
+ "bbox": [
1034
+ 173,
1035
+ 395,
1036
+ 826,
1037
+ 439
1038
+ ],
1039
+ "page_idx": 10
1040
+ },
1041
+ {
1042
+ "type": "text",
1043
+ "text": "Qianli Ma, Jiawei Zheng, Sen Li, and Gary W Cottrell. Learning representations for time series clustering. In Advances in Neural Information Processing Systems, pp. 3776–3786, 2019. ",
1044
+ "bbox": [
1045
+ 173,
1046
+ 446,
1047
+ 823,
1048
+ 477
1049
+ ],
1050
+ "page_idx": 10
1051
+ },
1052
+ {
1053
+ "type": "text",
1054
+ "text": "Pankaj Malhotra, Vishnu TV, Lovekesh Vig, Puneet Agarwal, and Gautam Shroff. Timenet: Pre-trained deep recurrent neural network for time series classification. arXiv preprint arXiv:1706.08838, 2017. ",
1055
+ "bbox": [
1056
+ 174,
1057
+ 484,
1058
+ 823,
1059
+ 527
1060
+ ],
1061
+ "page_idx": 10
1062
+ },
1063
+ {
1064
+ "type": "text",
1065
+ "text": "Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781, 2013. ",
1066
+ "bbox": [
1067
+ 173,
1068
+ 536,
1069
+ 825,
1070
+ 566
1071
+ ],
1072
+ "page_idx": 10
1073
+ },
1074
+ {
1075
+ "type": "text",
1076
+ "text": "George Moody. A new method for detecting atrial fibrillation using RR intervals. Computers in Cardiology, pp. 227–230, 1983. ",
1077
+ "bbox": [
1078
+ 173,
1079
+ 574,
1080
+ 823,
1081
+ 604
1082
+ ],
1083
+ "page_idx": 10
1084
+ },
1085
+ {
1086
+ "type": "text",
1087
+ "text": "Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw audio. arXiv preprint arXiv:1609.03499, 2016. ",
1088
+ "bbox": [
1089
+ 173,
1090
+ 612,
1091
+ 825,
1092
+ 655
1093
+ ],
1094
+ "page_idx": 10
1095
+ },
1096
+ {
1097
+ "type": "text",
1098
+ "text": "Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. ",
1099
+ "bbox": [
1100
+ 169,
1101
+ 664,
1102
+ 825,
1103
+ 694
1104
+ ],
1105
+ "page_idx": 10
1106
+ },
1107
+ {
1108
+ "type": "text",
1109
+ "text": "Slobodan Petrovic. A comparison between the silhouette index and the davies-bouldin index in labelling ids clusters. In Proceedings of the 11th Nordic Workshop of Secure IT Systems, pp. 53–64, 2006. ",
1110
+ "bbox": [
1111
+ 176,
1112
+ 702,
1113
+ 823,
1114
+ 744
1115
+ ],
1116
+ "page_idx": 10
1117
+ },
1118
+ {
1119
+ "type": "text",
1120
+ "text": "Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. ",
1121
+ "bbox": [
1122
+ 171,
1123
+ 753,
1124
+ 823,
1125
+ 784
1126
+ ],
1127
+ "page_idx": 10
1128
+ },
1129
+ {
1130
+ "type": "text",
1131
+ "text": "Alec Radford, Rafal Jozefowicz, and Ilya Sutskever. Learning to generate reviews and discovering sentiment. arXiv preprint arXiv:1704.01444, 2017. ",
1132
+ "bbox": [
1133
+ 171,
1134
+ 791,
1135
+ 823,
1136
+ 820
1137
+ ],
1138
+ "page_idx": 10
1139
+ },
1140
+ {
1141
+ "type": "text",
1142
+ "text": "Nikunj Saunshi, Orestis Plevrakis, Sanjeev Arora, Mikhail Khodak, and Hrishikesh Khandeparkar. A theoretical analysis of contrastive unsupervised representation learning. In International Conference on Machine Learning, pp. 5628–5637. PMLR, 2019. ",
1143
+ "bbox": [
1144
+ 176,
1145
+ 829,
1146
+ 823,
1147
+ 872
1148
+ ],
1149
+ "page_idx": 10
1150
+ },
1151
+ {
1152
+ "type": "text",
1153
+ "text": "Peter Schulam, Fredrick Wigley, and Suchi Saria. Clustering longitudinal clinical marker trajectories from electronic health data: Applications to phenotyping and endotype discovery. In Twenty-Ninth AAAI Conference on Artificial Intelligence, 2015. ",
1154
+ "bbox": [
1155
+ 174,
1156
+ 881,
1157
+ 825,
1158
+ 924
1159
+ ],
1160
+ "page_idx": 10
1161
+ },
1162
+ {
1163
+ "type": "text",
1164
+ "text": "Harini Suresh, Jen J Gong, and John V Guttag. Learning tasks for multitask learning: Heterogenous patient populations in the icu. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pp. 802–810, 2018. ",
1165
+ "bbox": [
1166
+ 176,
1167
+ 103,
1168
+ 823,
1169
+ 146
1170
+ ],
1171
+ "page_idx": 11
1172
+ },
1173
+ {
1174
+ "type": "text",
1175
+ "text": "Sana Tonekaboni, Shalmali Joshi, Kieran Campbell, David K Duvenaud, and Anna Goldenberg. What went wrong and when? instance-wise feature importance for time-series black-box models. Advances in Neural Information Processing Systems, 33, 2020. ",
1176
+ "bbox": [
1177
+ 178,
1178
+ 155,
1179
+ 823,
1180
+ 198
1181
+ ],
1182
+ "page_idx": 11
1183
+ },
1184
+ {
1185
+ "type": "text",
1186
+ "text": "Xiang Wang, David Sontag, and Fei Wang. Unsupervised learning of disease progression models. In Proceedings of the 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 85–94, 2014. ",
1187
+ "bbox": [
1188
+ 178,
1189
+ 205,
1190
+ 823,
1191
+ 250
1192
+ ],
1193
+ "page_idx": 11
1194
+ },
1195
+ {
1196
+ "type": "text",
1197
+ "text": "Xiaolong Wang and Abhinav Gupta. Unsupervised learning of visual representations using videos. In Proceedings of the IEEE International Conference on Computer Vision, pp. 2794–2802, 2015. ",
1198
+ "bbox": [
1199
+ 174,
1200
+ 257,
1201
+ 823,
1202
+ 287
1203
+ ],
1204
+ "page_idx": 11
1205
+ },
1206
+ {
1207
+ "type": "text",
1208
+ "text": "Qiang Yang and Xindong Wu. 10 challenging problems in data mining research. International Journal of Information Technology & Decision Making, 5(04):597–604, 2006. ",
1209
+ "bbox": [
1210
+ 173,
1211
+ 295,
1212
+ 821,
1213
+ 325
1214
+ ],
1215
+ "page_idx": 11
1216
+ },
1217
+ {
1218
+ "type": "text",
1219
+ "text": "Tom Young, Devamanyu Hazarika, Soujanya Poria, and Erik Cambria. Recent trends in deep learning based natural language processing. IEEE Computational Intelligence Magazine, 13(3):55–75, 2018. ",
1220
+ "bbox": [
1221
+ 176,
1222
+ 333,
1223
+ 825,
1224
+ 376
1225
+ ],
1226
+ "page_idx": 11
1227
+ },
1228
+ {
1229
+ "type": "text",
1230
+ "text": "A APPENDIX ",
1231
+ "bbox": [
1232
+ 176,
1233
+ 102,
1234
+ 299,
1235
+ 117
1236
+ ],
1237
+ "page_idx": 12
1238
+ },
1239
+ {
1240
+ "type": "text",
1241
+ "text": "A.1 SIMULATED DATASET ",
1242
+ "text_level": 1,
1243
+ "bbox": [
1244
+ 176,
1245
+ 132,
1246
+ 370,
1247
+ 147
1248
+ ],
1249
+ "page_idx": 12
1250
+ },
1251
+ {
1252
+ "type": "image",
1253
+ "img_path": "images/50d00ce6f7b9850d9be2d0710b8c08d5aea991a9f2270a96461961fad39a337e.jpg",
1254
+ "image_caption": [
1255
+ "Figure A.1: A normalized time series sample from the simulated dataset. Each row represents a single feature, and the shaded regions indicate one of the 4 underllying simulated states. "
1256
+ ],
1257
+ "image_footnote": [],
1258
+ "bbox": [
1259
+ 243,
1260
+ 185,
1261
+ 738,
1262
+ 386
1263
+ ],
1264
+ "page_idx": 12
1265
+ },
1266
+ {
1267
+ "type": "text",
1268
+ "text": "The simulated time series consists of 3 features generated from different underlying hidden states. Figure A.1 shows a sample from this dataset. Each panel in the figure shows one of the features, and the shaded regions indicate the underlying state of the signal in that period. We use a Hidden Markov Model (HMM) to generate these random latent states over time. The transition probability is set equal to $\\% 5$ for switching to an alternating state, and $\\% 8 5$ for not changing state. In each state, the time series is generated from a different signal distribution. Table 3 describes the generative process of each signal feature in each state. Note that feature 1 and 2 are always correlated, mainly to mimic realistic clinical time series. As an example, physiological measurements like pulse rate and heart rate are always correlated. ",
1269
+ "bbox": [
1270
+ 173,
1271
+ 446,
1272
+ 826,
1273
+ 573
1274
+ ],
1275
+ "page_idx": 12
1276
+ },
1277
+ {
1278
+ "type": "table",
1279
+ "img_path": "images/9474fffad00eb9f67f2f66fb95c77435872a402bb29d1975d4b8bcc6df0d3954.jpg",
1280
+ "table_caption": [
1281
+ "Table 3: Signal distributions for each time series feature of the simulated dataset "
1282
+ ],
1283
+ "table_footnote": [],
1284
+ "table_body": "<table><tr><td></td><td>State 1</td><td>State 2</td><td>State 3</td><td>State 4</td></tr><tr><td>Feature 1</td><td>GP (periodic)</td><td>NARMAα</td><td>GP (Squared Exp.)</td><td>NARMAβ</td></tr><tr><td>Feature 2</td><td>GP (periodic)</td><td>NARMAα</td><td>GP (Squared Exp.)</td><td>NARMAβ</td></tr><tr><td>Feature 3</td><td>GP (Squared Exp.)</td><td>NARMAβ</td><td>GP (periodic)</td><td>NARMAα</td></tr></table>",
1285
+ "bbox": [
1286
+ 225,
1287
+ 583,
1288
+ 774,
1289
+ 656
1290
+ ],
1291
+ "page_idx": 12
1292
+ },
1293
+ {
1294
+ "type": "text",
1295
+ "text": "In the state 1, the correlated features are generated by a Gaussian Process (GP) with a periodic kernel. Feature 3, which is uncorrelated with the other two features, comes from another GP with a squared exponential kernel. In addition to GPs, we also have multiple Non-Linear Auto-Regressive Moving Average (NARMA) time series models. The linear function of $\\mathrm { N A R M A } _ { \\alpha }$ and $\\mathrm { N A R M A } _ { \\beta }$ are shown in Equation 2 and 3. ",
1296
+ "bbox": [
1297
+ 173,
1298
+ 694,
1299
+ 826,
1300
+ 763
1301
+ ],
1302
+ "page_idx": 12
1303
+ },
1304
+ {
1305
+ "type": "equation",
1306
+ "img_path": "images/8d7907530af614dfe10e56303d96bdf4ef0097c0d5dded0a8853746446ead612.jpg",
1307
+ "text": "$$\n\\mathrm { N A R M A } _ { \\alpha } : y ( k + 1 ) = 0 . 3 y ( k ) + 0 . 0 5 y ( k ) \\sum _ { i = 0 } ^ { n - 1 } y ( k - i ) + 1 . 5 u ( k - ( n - 1 ) ) u ( k ) + 0 . 1 1\n$$",
1308
+ "text_format": "latex",
1309
+ "bbox": [
1310
+ 187,
1311
+ 784,
1312
+ 787,
1313
+ 829
1314
+ ],
1315
+ "page_idx": 12
1316
+ },
1317
+ {
1318
+ "type": "equation",
1319
+ "img_path": "images/d3e74b4e971e93caeec5a4b057dbdc38375279daa56e09ca2ad8e8ed3725dd57.jpg",
1320
+ "text": "$$\n\\mathbf { N A R M A } _ { \\beta } : y ( k + 1 ) = 0 . 1 y ( k ) + 0 . 2 5 y ( k ) \\sum _ { i = 0 } ^ { n - 1 } y ( k - i ) + 2 . 5 u ( k - ( n - 1 ) ) u ( k ) - 0 . 0 0 5\n$$",
1321
+ "text_format": "latex",
1322
+ "bbox": [
1323
+ 181,
1324
+ 843,
1325
+ 797,
1326
+ 886
1327
+ ],
1328
+ "page_idx": 12
1329
+ },
1330
+ {
1331
+ "type": "text",
1332
+ "text": "A white Gaussian noise with $\\sigma = 0 . 3$ is added to all signals, and overall, the dataset consists of 500 instances of $T = 2 0 0 0$ measurements. ",
1333
+ "bbox": [
1334
+ 173,
1335
+ 895,
1336
+ 823,
1337
+ 924
1338
+ ],
1339
+ "page_idx": 12
1340
+ },
1341
+ {
1342
+ "type": "text",
1343
+ "text": "A.2 BASELINE IMPLEMENTATION DETAILS ",
1344
+ "text_level": 1,
1345
+ "bbox": [
1346
+ 176,
1347
+ 104,
1348
+ 480,
1349
+ 117
1350
+ ],
1351
+ "page_idx": 13
1352
+ },
1353
+ {
1354
+ "type": "text",
1355
+ "text": "Implementation of all baselines are included in the code base for reproducibility purposes, and hyper-parameters for all baselines are tuned using cross-validation. ",
1356
+ "bbox": [
1357
+ 174,
1358
+ 132,
1359
+ 823,
1360
+ 160
1361
+ ],
1362
+ "page_idx": 13
1363
+ },
1364
+ {
1365
+ "type": "text",
1366
+ "text": "Contrastive Predictive Coding (CPC): The CPC baseline first processes the sequential signal windows using an encoder $Z _ { t } = E n c ( X _ { t } )$ , with a similar architecture to the encoders of other baselines. Next, an autoregressive model $g _ { a r }$ aggregates all the information in $Z _ { \\le t }$ and summarizes it into a context latent representation $c _ { t } = g _ { a r } ( Z _ { \\leq t } )$ . In our implementation, we have used a single layer, a one-directional recurrent neural network with GRU cell and hidden size equal to the encoding size as the auto-regressor. Like the original paper, the density ratio is estimated using a linear transformation, and the model is trained for 1 step ahead prediction. ",
1367
+ "bbox": [
1368
+ 174,
1369
+ 183,
1370
+ 825,
1371
+ 280
1372
+ ],
1373
+ "page_idx": 13
1374
+ },
1375
+ {
1376
+ "type": "text",
1377
+ "text": "Triplet-Loss (T-Loss): The triplet loss baseline is implemented using the original code made available by the authors on Github3. ",
1378
+ "bbox": [
1379
+ 176,
1380
+ 295,
1381
+ 820,
1382
+ 323
1383
+ ],
1384
+ "page_idx": 13
1385
+ },
1386
+ {
1387
+ "type": "text",
1388
+ "text": "KNN and K-means: These two baselines for classification and clustering are implemented using the tslearn library 4, that integrates distance metrics such as DTW. Note that evaluating DTW is computationally expensive, and the tslearn implementation is not optimized. Therefore, for the waveform data with windows of size 2500, we had to down-sampled the signal frequency by a factor of two. ",
1389
+ "bbox": [
1390
+ 174,
1391
+ 339,
1392
+ 825,
1393
+ 409
1394
+ ],
1395
+ "page_idx": 13
1396
+ },
1397
+ {
1398
+ "type": "text",
1399
+ "text": "A.3 TNC IMPLEMENTATION EXTRA DETAILS ",
1400
+ "text_level": 1,
1401
+ "bbox": [
1402
+ 174,
1403
+ 433,
1404
+ 496,
1405
+ 446
1406
+ ],
1407
+ "page_idx": 13
1408
+ },
1409
+ {
1410
+ "type": "text",
1411
+ "text": "To define the neighborhood range in the TNC framework, as mentioned earlier, we use the AugmentedDickey Fuller (ADF) statistical test to determine this range $( \\eta )$ as the region for which the signals remain stationary. More precisely, we gradually increase the range, from a single window size up to 3 times the window size (the upper limit we set), and repeatedly perform the ADF test. We use the $p$ -value from this statistical test to determine whether the Null hypothesis can be rejected, meaning that the signal is stationary. At the point where the $p$ -value is above our defined threshold (0.01), we can no longer assume that the signal is stationary, and this is where we set the $\\eta$ parameter. Now, once the neighborhood is defined, we make sure the non-neighboring samples are taken from the distribution of windows with at least $4 \\times \\eta$ away from $W _ { t }$ , ensuring a low likelihood of belonging to the neighborhood. Note that for implementation of ADF, we use the stats model library 5. Unfortunately, this implementation is not optimized and does not support GPU computation, therefore evaluating the neighborhood range using ADF slows down TNC framework training. As a future direction, we are working on an optimized implementation of the ADF score for our framework. ",
1412
+ "bbox": [
1413
+ 173,
1414
+ 462,
1415
+ 825,
1416
+ 642
1417
+ ],
1418
+ "page_idx": 13
1419
+ },
1420
+ {
1421
+ "type": "text",
1422
+ "text": "A.4 SELECTING THE WINDOW SIZE ",
1423
+ "text_level": 1,
1424
+ "bbox": [
1425
+ 178,
1426
+ 666,
1427
+ 429,
1428
+ 679
1429
+ ],
1430
+ "page_idx": 13
1431
+ },
1432
+ {
1433
+ "type": "text",
1434
+ "text": "The window size $\\delta$ is an important factor in the performance of a representation learning framework, not only for TNC but also for similar baselines such as CPC and triplet loss. Overall, the window size should be selected such that it is long enough to contain information about the underlying state and not too long to span over multiple underlying states. In our settings, we have selected the window sizes based on our prior knowledge of the signals. For instance, in the case of an ECG signal, the selected window size is equivalent to 7 seconds of recording, which is small enough such that the ECG remains in a stable state and yet has enough information to determine that underlying state. Our understanding of the time series data can help us select an appropriate window size, but we can also experiment with different $\\delta$ to learn this parameter. Table 4 shows classification performance results for the simulation setups, under different window sizes. We can clearly see the drop in performance for all baseline methods when the window size is too small or too large. ",
1435
+ "bbox": [
1436
+ 174,
1437
+ 694,
1438
+ 825,
1439
+ 847
1440
+ ],
1441
+ "page_idx": 13
1442
+ },
1443
+ {
1444
+ "type": "table",
1445
+ "img_path": "images/5271d9548957259ca08afa67e8a61c837aec90150140c91122ef9b4c928d7e6f.jpg",
1446
+ "table_caption": [],
1447
+ "table_footnote": [],
1448
+ "table_body": "<table><tr><td></td><td colspan=\"2\">δ=10</td><td colspan=\"2\">8=50</td><td colspan=\"2\">δ=100</td></tr><tr><td></td><td>AUPRC</td><td> Accuracy</td><td>AUPRC</td><td>Accuracy</td><td>AUPRC</td><td>Accuracy</td></tr><tr><td>TNC</td><td>0.74 ± 0.01</td><td>71.60 ± 0.59</td><td>0.99 ± 0.00</td><td>97.52 ± 0.13</td><td>0.84 ± 0.11</td><td>84.25 ± 9.08</td></tr><tr><td>CPC</td><td>0.49 ± 0.02</td><td>51.85 ± 1.81</td><td>0.69 ± 0.06</td><td>70.26 ± 6.48</td><td>0.49 ± 0.05</td><td>56.65 ± 0.81</td></tr><tr><td>T-Loss</td><td>0.48 ± 0.06</td><td>56.70 ± 1.07</td><td>0.78 ± 0.01</td><td>76.66 ± 1.14</td><td>0.73 ± 0.008</td><td>73.29 ± 1.58</td></tr></table>",
1449
+ "bbox": [
1450
+ 173,
1451
+ 99,
1452
+ 857,
1453
+ 186
1454
+ ],
1455
+ "page_idx": 14
1456
+ },
1457
+ {
1458
+ "type": "text",
1459
+ "text": "Table 4: Downstream classification performance for different window size $\\delta$ on the simulated dataset ",
1460
+ "bbox": [
1461
+ 171,
1462
+ 196,
1463
+ 825,
1464
+ 210
1465
+ ],
1466
+ "page_idx": 14
1467
+ },
1468
+ {
1469
+ "type": "text",
1470
+ "text": "A.5 CLUSTERING METRICS ",
1471
+ "text_level": 1,
1472
+ "bbox": [
1473
+ 176,
1474
+ 239,
1475
+ 377,
1476
+ 253
1477
+ ],
1478
+ "page_idx": 14
1479
+ },
1480
+ {
1481
+ "type": "text",
1482
+ "text": "Most cluster validity measures assess certain structural properties of a clustering result. In our evaluation, we have used two measures, namely the Silhouette score and Davies-Bouldin index, to evaluate the representations’ clustering quality. Davies-Bouldin measures intra-cluster similarity (coherence) and inter-cluster differences (separation). Let $\\mathcal { C } = \\{ \\mathcal { C } _ { 1 } , . . . , \\mathcal { C } _ { k } \\}$ be a clustering of a set $D$ of objects. The Davies-Bouldin score is evaluated as follows: ",
1483
+ "bbox": [
1484
+ 173,
1485
+ 265,
1486
+ 825,
1487
+ 335
1488
+ ],
1489
+ "page_idx": 14
1490
+ },
1491
+ {
1492
+ "type": "equation",
1493
+ "img_path": "images/54f39cde268a76b58bff765dfe357864b5d7e670c7dd92b4de468290b613a122.jpg",
1494
+ "text": "$$\nD B = \\frac { 1 } { k } \\sum _ { i = 1 } ^ { k } m a x _ { j } \\frac { s ( \\mathcal { C } ) + s ( \\mathcal { C } ) } { \\delta \\mathcal { C } _ { i } \\mathcal { C } _ { j } }\n$$",
1495
+ "text_format": "latex",
1496
+ "bbox": [
1497
+ 390,
1498
+ 352,
1499
+ 607,
1500
+ 396
1501
+ ],
1502
+ "page_idx": 14
1503
+ },
1504
+ {
1505
+ "type": "text",
1506
+ "text": "Where $s ( \\mathcal { C } )$ measures the scatter within a cluster, and $\\delta$ is a cluster to cluster distance measure. On the other hand, the silhouette score measures how similar an object is to its cluster compared to other clusters. Both measures are commonly used for the evaluation of clustering algorithms. A comparison of 2 metrics has shown that the Silhouette index produces slightly more accurate results in some cases. However, the Davies-Bouldin index is generally much less complex to compute Petrovic (2006). ",
1507
+ "bbox": [
1508
+ 173,
1509
+ 406,
1510
+ 826,
1511
+ 478
1512
+ ],
1513
+ "page_idx": 14
1514
+ },
1515
+ {
1516
+ "type": "text",
1517
+ "text": "A.6 SETTING THE WEIGHTS FOR PU LEARNING",
1518
+ "text_level": 1,
1519
+ "bbox": [
1520
+ 174,
1521
+ 494,
1522
+ 516,
1523
+ 508
1524
+ ],
1525
+ "page_idx": 14
1526
+ },
1527
+ {
1528
+ "type": "text",
1529
+ "text": "As mentioned in the Experiment section, the weight parameter in the loss is the probability of sampling a positive window from the non-neighboring region. One way to set this parameter is using prior knowledge of the number and the distribution of underlying states. Another way is to learn it as a hyperparameter. Table 5 shows the TNC loss for different weight parameters. The loss column reports the value measured in Equation 1, and the accuracy shows how well the discriminator identifies the neighboring samples from non-neighboring ones for settings with different weight parameters. To also assess the impact of re-weighting the loss on downstream classification performance, we compared these performance measures for weighted and non-weighted settings. Table 6 demonstrates these results and confirms that weight adjusting the loss for non-neighboring samples improves the quality of learned representations. ",
1530
+ "bbox": [
1531
+ 173,
1532
+ 520,
1533
+ 826,
1534
+ 660
1535
+ ],
1536
+ "page_idx": 14
1537
+ },
1538
+ {
1539
+ "type": "table",
1540
+ "img_path": "images/fba7183a0bf8aecdaebedd195051259d3e080453d5c4e43c164c0f596f99d775.jpg",
1541
+ "table_caption": [
1542
+ "Table 5: Training the TNC framework using different weight parameters. The loss is the measured value determined in Equation 1, and the Accuracy is the accuracy of the discriminator. "
1543
+ ],
1544
+ "table_footnote": [],
1545
+ "table_body": "<table><tr><td></td><td colspan=\"2\">Simulation</td><td colspan=\"2\">ECG Waveform</td><td colspan=\"2\">HAR</td></tr><tr><td>Weight</td><td>Loss</td><td>Accuracy</td><td>Loss</td><td>Accuracy</td><td>Loss</td><td>Accuracy</td></tr><tr><td>0.2</td><td>0.582±0.002</td><td>74.29±0.61</td><td>0.631±0.011</td><td>60.44±2.56</td><td>0.475±0.004</td><td>85.75±0.5</td></tr><tr><td>0.1</td><td>0.571±0.011</td><td>75.41±0.37</td><td>0.637±0.011</td><td>63.67±1.29</td><td>0.413±0.003</td><td>88.21±1.29</td></tr><tr><td>0.05</td><td>0.576±0.002</td><td>75.73±0.24</td><td>0.622±0.023</td><td>66.04±3.46</td><td>0.383±0.001</td><td>87.33±0.17</td></tr></table>",
1546
+ "bbox": [
1547
+ 173,
1548
+ 672,
1549
+ 851,
1550
+ 760
1551
+ ],
1552
+ "page_idx": 14
1553
+ },
1554
+ {
1555
+ "type": "table",
1556
+ "img_path": "images/68fdba447c84b79bbe4ba3a5d794bbc19c348fdae5e651fc5967bf75afc110a9.jpg",
1557
+ "table_caption": [],
1558
+ "table_footnote": [],
1559
+ "table_body": "<table><tr><td>Weighting?</td><td>Simulation</td><td>ECG Waveform</td><td>HAR</td></tr><tr><td>True</td><td>97.52±0.13</td><td>77.79±0.84</td><td>88.32±0.12</td></tr><tr><td>False</td><td>97.17±0.44</td><td>75.26±1.48</td><td>75.25±13.6</td></tr></table>",
1560
+ "bbox": [
1561
+ 289,
1562
+ 820,
1563
+ 705,
1564
+ 873
1565
+ ],
1566
+ "page_idx": 14
1567
+ },
1568
+ {
1569
+ "type": "text",
1570
+ "text": "Table 6: Downstream classification accuracy on simulated data with the TNC frameworks, using 2 different weighting strategies: 1)Trained with weight adjustment, 2)Trained with $w = 0$ . ",
1571
+ "bbox": [
1572
+ 171,
1573
+ 883,
1574
+ 825,
1575
+ 912
1576
+ ],
1577
+ "page_idx": 14
1578
+ },
1579
+ {
1580
+ "type": "text",
1581
+ "text": "A.7 SUPPLEMENTARY FIGURE ",
1582
+ "text_level": 1,
1583
+ "bbox": [
1584
+ 176,
1585
+ 103,
1586
+ 398,
1587
+ 117
1588
+ ],
1589
+ "page_idx": 15
1590
+ },
1591
+ {
1592
+ "type": "text",
1593
+ "text": "A.7.1 CLINICAL WAVEFORM DATA ",
1594
+ "text_level": 1,
1595
+ "bbox": [
1596
+ 176,
1597
+ 132,
1598
+ 424,
1599
+ 146
1600
+ ],
1601
+ "page_idx": 15
1602
+ },
1603
+ {
1604
+ "type": "text",
1605
+ "text": "In order to understand what TNC framework encodes from the high dimensional ECG signals, we visualize the trajectory of the representations of an individual sample over time. Figure A.3 demonstrates this example, where the top 2 rows are ECG signals from two recording leads and the bottom row demonstrates the representation vectors. We see that around second 40 the pattern in the representations change as a result of an artifact happening in one of the signals. With the help from our clinical expert, we also tried to interpret different patterns in the encoding space. For instance, between time 80 and 130, where features 0-10 become more activated, the heart rate (HR) has increased. Increase in HR can be seen as increased frequency in the ECG signals and is one of the indicators of arrhythmia that we believe TNC has captured. Figure A.2 shows the distribution of the latent encoding of ECG signals for different baselines, with colors indicating the arrhythmia class. ",
1606
+ "bbox": [
1607
+ 173,
1608
+ 159,
1609
+ 825,
1610
+ 299
1611
+ ],
1612
+ "page_idx": 15
1613
+ },
1614
+ {
1615
+ "type": "image",
1616
+ "img_path": "images/f8c41d9d7b7f3ca2dee2d86c0465aaca53aa42304010fcb04480e641114baedc.jpg",
1617
+ "image_caption": [
1618
+ "Figure A.2: T-SNE visualization of waveform signal representations for unsupervised representation learning baselines. Each point in the plot is a 64 dimensional representation of a window of time series, with the color indicating the latent state. "
1619
+ ],
1620
+ "image_footnote": [],
1621
+ "bbox": [
1622
+ 186,
1623
+ 323,
1624
+ 825,
1625
+ 458
1626
+ ],
1627
+ "page_idx": 15
1628
+ },
1629
+ {
1630
+ "type": "image",
1631
+ "img_path": "images/cdf9f7bf417aea0ea9e860d91a4ecbbb954f0389143f6bc77bc251a49ec2386f.jpg",
1632
+ "image_caption": [
1633
+ "Figure A.3: Trajectory of a waveform signal encoding. The top two plots show the ECG recordings from 2 ECG leads. The bottom plot shows the 64 dimensional encoding of the sliding windows $W _ { t }$ where $\\delta = 2 5 0 0$ . "
1634
+ ],
1635
+ "image_footnote": [],
1636
+ "bbox": [
1637
+ 184,
1638
+ 545,
1639
+ 815,
1640
+ 763
1641
+ ],
1642
+ "page_idx": 15
1643
+ },
1644
+ {
1645
+ "type": "text",
1646
+ "text": "A.7.2 HAR DATA ",
1647
+ "text_level": 1,
1648
+ "bbox": [
1649
+ 176,
1650
+ 854,
1651
+ 308,
1652
+ 868
1653
+ ],
1654
+ "page_idx": 15
1655
+ },
1656
+ {
1657
+ "type": "text",
1658
+ "text": "Figure A.4 and A.5 are similar plots to the ones demonstrated in the previous section, but for the HAR dataset. As shown in Figure A.5, the underlying states of the signal are clearly captured by the TNC framework as different patterns in the latent representations. ",
1659
+ "bbox": [
1660
+ 174,
1661
+ 882,
1662
+ 823,
1663
+ 924
1664
+ ],
1665
+ "page_idx": 15
1666
+ },
1667
+ {
1668
+ "type": "image",
1669
+ "img_path": "images/38c699134734a4f23855bab05972dcbbbc7b6990cd92481f77433a1d84d0a101.jpg",
1670
+ "image_caption": [
1671
+ "Figure A.4: T-SNE visualization of HAR signal representations for all baselines. Each point in the plot is a 10 dimensional representation of a window of $\\delta = 4$ , with colors indicating latent states. "
1672
+ ],
1673
+ "image_footnote": [],
1674
+ "bbox": [
1675
+ 186,
1676
+ 106,
1677
+ 825,
1678
+ 242
1679
+ ],
1680
+ "page_idx": 16
1681
+ },
1682
+ {
1683
+ "type": "image",
1684
+ "img_path": "images/216f68c8384e1cbf4eb2ddfdaf789ae535631b1609fb8be97fdb673ecdfa0ab9.jpg",
1685
+ "image_caption": [
1686
+ "Figure A.5: Trajectory of a HAR signal encoding. The top plot shows the original time series with shaded regions indicating the underlying state. The bottom plot shows the 10 dimensional encoding of the sliding windows $W _ { t }$ where $\\delta = 4$ . "
1687
+ ],
1688
+ "image_footnote": [],
1689
+ "bbox": [
1690
+ 184,
1691
+ 297,
1692
+ 815,
1693
+ 477
1694
+ ],
1695
+ "page_idx": 16
1696
+ },
1697
+ {
1698
+ "type": "text",
1699
+ "text": "A.7.3 SIMULATION DATA ",
1700
+ "text_level": 1,
1701
+ "bbox": [
1702
+ 176,
1703
+ 559,
1704
+ 361,
1705
+ 574
1706
+ ],
1707
+ "page_idx": 16
1708
+ },
1709
+ {
1710
+ "type": "text",
1711
+ "text": "In addition to the initial experiment, we also show the trajectory of the encoding for a smaller encoding size (3). In this setting, we have 4 underlying states in the signal, and only 3 dimensions for the encoding. ",
1712
+ "bbox": [
1713
+ 174,
1714
+ 583,
1715
+ 825,
1716
+ 627
1717
+ ],
1718
+ "page_idx": 16
1719
+ },
1720
+ {
1721
+ "type": "image",
1722
+ "img_path": "images/074b1e338add5ee800eec2f4ff2c88c3d91ab6b2e3e944728fe7d5a7e281b535.jpg",
1723
+ "image_caption": [
1724
+ "Figure A.6: Trajectory of a simulation signal encoding. The top plots shows the signals and the bottom plot shows the 3 dimensional encoding of the sliding windows $W _ { t }$ where $\\delta = 5 0$ . "
1725
+ ],
1726
+ "image_footnote": [],
1727
+ "bbox": [
1728
+ 184,
1729
+ 641,
1730
+ 815,
1731
+ 820
1732
+ ],
1733
+ "page_idx": 16
1734
+ }
1735
+ ]
parse/train/8qDwejCuCN/8qDwejCuCN_middle.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/8qDwejCuCN/8qDwejCuCN_model.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/B1gd0nEFwS/B1gd0nEFwS.md ADDED
@@ -0,0 +1,433 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # UNIVERSAL SOURCE-FREE DOMAIN ADAPTATION
2
+
3
+ Anonymous authors Paper under double-blind review
4
+
5
+ # ABSTRACT
6
+
7
+ There is a strong incentive to develop versatile learning techniques that can transfer the knowledge of class-separability from a labeled source domain to an unlabeled target domain in the presence of a domain-shift. Existing domain adaptation (DA) approaches are not equipped for practical DA scenarios as a result of their reliance on the knowledge of source-target label-set relationship (e.g. Closed-set, Open-set or Partial DA). Furthermore, almost all the prior unsupervised DA works require coexistence of source and target samples even during deployment, making them unsuitable for incremental, real-time adaptation. Devoid of such highly impractical assumptions, we propose a novel two-stage learning process. Initially, in the procurement-stage, the objective is to equip the model for future sourcefree deployment, assuming no prior knowledge of the upcoming category-gap and domain-shift. To achieve this, we enhance the model’s ability to reject out-of-source distribution samples by leveraging the available source data, in a novel generative classifier framework. Subsequently, in the deployment-stage, the objective is to design a unified adaptation algorithm capable of operating across a wide range of category-gaps, with no access to the previously seen source samples. To achieve this, in contrast to the usage of complex adversarial training regimes, we define a simple yet effective source-free adaptation objective by utilizing a novel instancelevel weighing mechanism, named as Source Similarity Metric (SSM). A thorough evaluation shows the practical usability of the proposed learning framework with superior DA performance even over state-of-the-art source-dependent approaches.
8
+
9
+ # 1 INTRODUCTION
10
+
11
+ Deep learning models have proven to be highly successful over a wide variety of tasks (Krizhevsky et al., 2012; Ren et al., 2015). However, a majority of these remain heavily dependent on access to a huge amount of labeled samples to achieve a reliable level of generalization. A recognition model trained on a certain distribution of labeled samples (source domain) often fails to generalize (Chen et al., 2017) when deployed in a new environment (target domain) in the presence a discrepancy in the input distribution (Shimodaira, 2000). Domain adaptation (DA) algorithms seek to minimize this discrepancy either by learning a domain invariant feature representation (Long et al., 2015; Kumar et al., 2018; Ganin et al., 2016; Tzeng et al., 2015), or by learning independent domain transformations (Long et al., 2016) to a common latent representation through adversarial distribution matching (Tzeng et al., 2017; Nath Kundu et al., 2018), in the absence of target label information.
12
+
13
+ Most of the existing approaches (Zhang et al., 2018c; Tzeng et al., 2017) assume a common label-set shared between the source and target domains (i.e. $\mathcal { C } _ { s } = \mathcal { C } _ { t }$ ), which is often regarded as Closed-Set $D A$ (see Fig. 1). Though this assumption helps to analyze various insights of DA algorithms, such an assumption rarely holds true in real-world scenarios. Recently researchers have independently explored two broad adaptation settings by partly relaxing the above assumption. In the first kind, Partial DA (Zhang et al., 2018b; Cao et al., 2018a;b), the target label space is considered as a subset of the source label space (i.e. $\mathcal { C } _ { t } \subset \mathcal { C } _ { s } .$ ). This setting is more suited for large-scale universal source datasets, which will almost always subsume the label-set of a wide range of target domains. However, the availability of such a universal source is highly questionable for a wide range of input domains and tasks. In the second kind, regarded as Open-set $D A$ (Baktashmotlagh et al., 2019; Ge et al., 2017), the target label space is considered as a superset of the source label space (i.e. $\mathcal { C } _ { t } \supset \mathcal { C } _ { s }$ ). The major challenge in this setting is attributed to detection of target samples from the unobserved categories in a fully-unsupervised scenario. Apart from the above two extremes, certain works define a partly mixed scenario by allowing “private” label-set for both source and target domains (i.e. $\mathcal { C } _ { s } \setminus \bar { \mathcal { C } } _ { t } \ne \bar { \varnothing }$ and $\mathcal { C } _ { t } \setminus \mathcal { C } _ { s } \ne \emptyset )$ ) but with extra supervision such as few-shot labeled data (Luo et al., 2017) or accessA to the knowledge of common categories (Panareda Busto & Gall, 2017).
14
+
15
+ Most of the prior approaches consider each scenario in isolation and propose independent solutions. Thus, they require access to the knowledge of label-set relationship (or category-gap) to carefully choose a DA algorithm, which would be suitable for the problem in hand. Furthermore, all the prior unsupervised DA works require coexistence of source and
16
+
17
+ ![](images/d304e18939ede2d27e8f3a9e6132a58f7ef5985335bb5e554b9007baf8d121e6.jpg)
18
+ Figure 1: Various label-set relationships (category-gap).
19
+
20
+ target samples even during deployment, hence not source-free. This is highly impractical, as labeled source data may not be accessible after deployment due to several reasons such as, privacy concerns, restricted access to proprietary data, accidental loss of source data or other computational limitations in real-time deployment scenarios.
21
+
22
+ Acknowledging the aforementioned shortcomings, we propose one of the most convenient DA frameworks which is ingeniously equipped to address source-free DA for all kinds of label-set relationships, without any prior knowledge of the associated category-gap (i.e. universal- $D A$ ). We not only focus on identifying the key complications associated with the challenging problem setting, but also devise insightful ideas to tackle such complications by adopting learning techniques much different from the available DA literature. This leads us to realize a holistic solution which achieves superior DA performance even over prior source-dependent approaches.
23
+
24
+ # 2 RELATED WORK
25
+
26
+ We briefly review the available domain adaptation methods under the three major divisions according to the assumption on label-set relationship. a) Closed-set DA. The cluster of previous works under this setting focuses on minimizing the domain gap at some intermediate feature level either by minimizing well-defined statistical distance functions (Wang & Schneider, 2014; Duan et al., 2012; Zhang et al., 2013; Saenko et al., 2010) or by formalizing it as an adversarial distribution matching problem (Tzeng et al., 2017; Kang et al., 2018; Long et al., 2018; Hu et al., 2018; Hoffman et al., 2018) inspired from the Generative Adversarial Nets (Goodfellow et al., 2014). Certain prior works (Sankaranarayanan et al., 2018; Zhu et al., 2017; Hoffman et al., 2018) use GAN framework to explicitly generate target-like images translated from the source image samples, which is also regarded as pixel-level adaptation (Bousmalis et al., 2017) in contrast to other feature level adaptation works (Nath Kundu et al., 2018; Tzeng et al., 2017; Long et al., 2015; 2016). b) Partial DA. Focusing on Partial $D A$ , Cao et al. (2018a) proposed to achieve adversarial class-level matching by utilizing multiple domain discriminators furnishing class-level and instance-level weighting for individual data samples. Zhang et al. (2018b) proposed to utilize importance weights for source samples depending on their similarity to the target domain data using an auxilliary discriminator. To effectively address the problem of negative-transfer (Wang et al., 2019), Cao et al. (2018b) employed a single discriminator to achieve both adversarial adaptation and class-level weighting of source samples. c) Open-set DA. Saito et al. (2018b) proposed a more general open-set adaptation setting without accessing the knowledge of source private labels set in contrast to the prior work (Panareda Busto & Gall, 2017). They extended the source classifier to accommodate an additional “unknown” class, which is trained adversarially against the other source classes. Universal DA. You et al. (2019) proposed Universal DA, which requires no prior knowledge of label-set relationship similar to the proposed setting, but considers access to both source and target samples during adaptation.
27
+
28
+ # 3 PROPOSED APPROACH
29
+
30
+ The problem setting for source-free domain adaptation is broadly divided into a two stage process.
31
+
32
+ a) Procurement stage. In this stage, we are given full access to the labeled samples of source domain, $\mathcal { D } _ { s } = \{ ( x _ { s } , y _ { s } ) : x _ { s } \sim p , y _ { s } \in$ $y _ { s } \in \mathcal { C } _ { s } \}$ , where $p$ is the distribution of source samples and $\mathcal { C } _ { s }$ denotes the label-set of the source domain. Here, the objective is to equip the model for the second stage, i.e. the Deployment stage, in the presence of a discrepancy in the distribution of input target samples. To achieve this we rely on an artificially generated negative dataset, $\mathcal { D } _ { n } = \{ ( x _ { n } ^ { - } , y _ { n } ) : \overset { \vartriangle } { x } _ { n } \sim \overset { \vartriangle } { p _ { n } }$ , $y _ { n } \in$ $\mathcal { C } _ { n } \}$ , where $p _ { n }$ is the distribution of negative source samples such that $\mathcal { C } _ { n } \cap \mathcal { C } _ { s } = \emptyset$ .
33
+
34
+ ![](images/7c4a1bdabcf61ef03c45f30f424abd934dfda6906789971b4cd39ea0c0cf2575.jpg)
35
+ Figure 2: Latent space cluster arrangement during adaptation (see Section 3.1.1).
36
+
37
+ inter-class separability negative classesb) Deployment stage. After obtaining a trained model from the Procurement stage, the model will have its first encounter with the unlabeled target domain samples from the deployed environment. We denote the unlabeled target data by $\mathcal { D } _ { t } = \{ x _ { t } : x _ { t } \sim q \}$ , where $q$ is the distribution of target samples. Note that, access to the source dataset $\mathcal { D } _ { s }$ from the previous stage is fully restricted during adaptation in the Deployment stage. Suppose that, $\mathcal { C } _ { t }$ is the "unknown" label-set of the target domain. We define the common label space between the source and target domain as ${ \mathcal { C } } = { \mathcal { C } } _ { s } \cap { \mathcal { C } } _ { t }$ . The private label-set for the source and the target domains is represented as $\overline { { \mathcal { C } } } _ { s } = \mathcal { C } _ { s } \setminus \mathcal { C } _ { t }$ and $\overline { { \mathcal { C } } } _ { t } = \mathcal { C } _ { t } \setminus \bar { \mathcal { C } } _ { s }$ respectively.
38
+
39
+ # 3.1 LEARNING IN THE PROCUREMENT STAGE
40
+
41
+ 3.1.1 Challenges. The available DA techniques heavily rely on the adversarial discriminative (Tzeng et al., 2017; Saito et al., 2018a) strategy. Thus, they require access to the source samples to reliably characterize the source domain distribution. Moreover, these approaches are not equipped to operate in a source-free setting. Though a generative model can be used as a memory-network (Sankaranarayanan et al., 2018; Bousmalis et al., 2017) to realize source-free adaptation, such a solution is not scalable for large-scale source datasets (e.g. ImageNet (Russakovsky et al., 2015)), as it introduces unnecessary extra parameters in addition to the associated training difficulties (Salimans et al., 2016). This calls for a fresh analysis of the requirements beyond the solutions found in literature.
42
+
43
+ In a general DA scenario, with access to source samples in the Deployment stage (specifically for Open-set or Partial DA), a widely adopted approach is to learn domain invariant features. In such approaches the placement of source category clusters is learned in the presence of unlabeled target samples which obliquely provides a supervision regarding the relationship between $\mathcal { C } _ { s }$ and $\mathcal { C } _ { t }$ . For instance, in case of Open-set DA, the source clusters may have to disperse to make space for the clusters from target private $\overline { { \mathcal { C } } } _ { t }$ (see Fig. 2a to 2b). Similarly, in partial DA, the source clusters may have to rearrange themselves to keep all the target shared clusters $\boldsymbol { \mathcal { C } } = \boldsymbol { \mathcal { C } } _ { t } ,$ separated from the source private $\overline { { \mathcal { C } } } _ { s } $ (see Fig. 2a to 2c). However in a complete source-free framework, we do not have the liberty to leverage such information as source and target samples never coexist together during training. Motivated by the adversarial discriminative DA technique (Tzeng et al., 2017), we hypothesize that, inculcating the ability to reject samples that are out of the source data distribution can facilitate future source-free domain alignment using this discriminatory knowledge. Therefore, in the Procurement stage the overarching objective is two-fold.
44
+
45
+ • Firstly, we must aim to learn a certain placement of source clusters best suited for all kinds of category-gap scenarios acknowledging the fact that, a source-free scenario does not allow us to modify the placement in the presence of target samples during adaptation (see Fig. 2d). • Secondly, the learned embedding must have the ability to reject out-of-distribution samples, which is an essential requirement for unsupervised adaptation in the presence of domain-shift.
46
+
47
+ 3.1.2 Solution. In the presence of source data, we aim to restrain the model’s domain and category bias which is generally inculcated as a result of the over-confident supervised learning paradigms (see Fig. 4A). To achieve this goal, we adopt two regularization strategies viz. i) regularization via generative modeling and ii) utilization of a labeled simulated negative source dataset to generalize for the latent regions not covered by the given positive source samples (see Fig. 4C).
48
+
49
+ How to configure the negative source dataset? While configuring $\mathcal { D } _ { n }$ , the following key properties have to be met. Firstly, latent clusters formed by the negative categories must lie in-between the latent clusters of positive source categories to enable a higher degree of intra-class compactness with interclass separability (Fig. 4C). Secondly, the negative source samples must enrich the source domain
50
+
51
+ ![](images/704a21e07f4382cc2d80b5105996d00e128de13f323c4b080a74999171e1d46f.jpg)
52
+ Figure 3: A) Simulated labeled negative samples using randomly created spline segments (in pink), p=20 B) Proposed architecture, C) Procurement stage yields compact source clusters on experimental data.
53
+
54
+ ![](images/cb8ad709403d5e32b81ce7f1e55a5fa54ee65a3c7e19fe5852dae47456b5f4c9.jpg)
55
+ Figure 4: Achieving intra-class compactness and inter-class separability using negative dataset $\mathcal { D } _ { n }$
56
+
57
+ A distribution without forming a new domain by themselves. This rules out the use of Mixup (Zhang et al., 2018a) or adversarial noise (Shu et al., 2018) as negative samples in this scenario. Thus, we propose the following two ways to synthesize the desired negative source dataset.
58
+
59
+ a) Image-composition as negative dataset $\mathcal { D } _ { n } ^ { ( a ) }$ . One of the key characteristics shared between the samples from source and unknown target domain is the semantics of the local part-related features specifically for image-based object recognition tasks. Relying on this assumption, we propose a systematic procedure to simulate the samples of $\mathcal { D } _ { n } ^ { ( a ) }$ by randomly compositing local regions between a pair of images drawn from the positive source dataset $\mathcal { D } _ { s }$ (see Fig. 3A and appendix, Algo. 2). Intuitively, composite samples $x _ { n }$ created on image pairs from different source categories are expected to lie in-between the two positive source clusters in the latent space, thereby introducing a combinatorial amount of new class labels i.e. $| \mathcal { C } _ { n } | = | \mathcal { C } _ { s } | _ { C _ { 2 } }$ .
60
+
61
+ b) Latent-simulated negative dataset $\mathcal { D } _ { n } ^ { ( b ) }$ . As an alternative approach, in the absence of domain knowledge (e.g. non-image datasets, or for tasks beyond image-recognition such as pose estimation), we propose to sample virtual negative instances, $u _ { n }$ from the latent space which are away from the high confidence regions (3-sigma) of positive source clusters (Fig. 4B). For each negative sample, we assign a negative class label (one of $| { \mathcal { C } } _ { n } | = | { \mathcal { C } } _ { s } | _ { C _ { 2 } ) }$ corresponding to the pair of most confident source classes predicted by the classifier. Thus, we obtain $\mathcal { D } _ { n } ^ { ( b ) } = \{ ( u _ { n } , y _ { n } ) : u _ { n } \sim p _ { n } ^ { u }$ , $y _ { n } \in { \mathcal { C } } _ { n } \}$ where $p _ { n } ^ { u }$ is the distribution of negative samples in the latent $u$ -space (more details in appendix Algo. 3).
62
+
63
+ Training procedure. The generative source classifier is divided into three stages; i) backbone-model $M$ , ii) feature extractor $F _ { s }$ , and iii) classifier $D$ (see Fig. 3B). Output of the backbone-model is denoted as $v = M ( x )$ , where $x$ is drawn from either $\mathcal { D } _ { s }$ or $\mathcal { D } _ { n }$ . Following this, the output of $F _ { s }$ and $D$ are represented as $u$ and $d$ respectively. $D$ outputs a $K$ -dimensional logit denoted as $d ^ { ( k ) }$ for $k = 1 , 2 . . . K$ ; $K = | \mathcal C _ { s } | + | \mathcal C _ { n } |$ . The individual class probabilities, $\hat { y } ^ { ( k ) }$ are obtained by applying softmax over the logits i.e. $\begin{array} { r } { \hat { y } ^ { ( k ) } = e x p ( d ^ { ( k ) } ) / \sum _ { k = 1 } ^ { K } e x p ( d ^ { ( k ) } ) = \sigma ^ { ( k ) } ( D \circ F _ { s } \circ M ( x ) ) . } \end{array}$ Additionally, we define priors of only positive source classes as $P ( u _ { s } | c _ { i } ) = \mathcal { N } ( u _ { s } | \mu _ { c _ { i } } , \Sigma _ { c _ { i } } )$ for $i = 1 , 2 . . . | \mathcal { C } _ { s } |$ at
64
+
65
+ # Algorithm 1 Training algorithm in the Procurement stage
66
+
67
+ 1: input: $( x _ { s } , y _ { s } ) \in \mathcal { D } _ { s }$ , $( x _ { n } , y _ { n } ) \in \mathcal { D } _ { n }$ ; $\theta _ { F _ { s } }$ , $\theta _ { D }$ , $\theta _ { G }$ : Parameters of $F _ { s }$ , $D$ and $G$ respectively.
68
+ 2: initialization: pretrain $\{ \theta _ { F _ { s } } , \theta _ { D } \}$ using cross-entropy loss on $( x _ { s } , y _ { s } )$ followed by initialization of the sample mean $\mu _ { c _ { i } }$ and covariance $\Sigma _ { c _ { i } }$ (at $u$ -space) of $F _ { s } \circ M ( x _ { s } )$ for $x _ { s }$ from class $c _ { i }$ ; $i = 1 , 2 , \dots | \mathcal { C } _ { s } |$
69
+ 3: for iter < MaxIter do
70
+ 4: $v _ { s } = M ( x _ { s } )$ ; $u _ { s } = F _ { s } ( v _ { s } )$ ; $\hat { v } _ { s } = G ( u _ { s } )$ ; $\boldsymbol { u } _ { r } \sim \mathcal { N } ( \mu _ { c _ { i } } , \Sigma _ { c _ { i } } )$ for $i = 1 , 2 , \dots | \mathcal { C } _ { s } |$ ; $\hat { u } _ { r } = F _ { s } \circ G ( u _ { r } )$
71
+ 5: $\hat { y } _ { s } ^ { ( k _ { s } ) } = \sigma ^ { ( k _ { s } ) } ( D \circ F _ { s } \circ M ( x _ { s } ) )$ , and $\hat { y } _ { n } ^ { ( k _ { n } ) } = \sigma ^ { ( k _ { n } ) } ( D \circ F _ { s } \circ M ( x _ { n } ) )$ where $k _ { s }$ and $k _ { n }$ are the index of ground-truth label $y _ { s }$ and $y _ { n }$ respectively.
72
+ 6: $\mathcal { L } _ { C E } = - \log \hat { y } _ { s } ^ { ( k _ { s } ) } - \alpha \log \hat { y } _ { n } ^ { ( k _ { n } ) }$ ; $\mathcal { L } _ { v } = | v _ { s } - \hat { v } _ { s } |$ ; $\mathcal { L } _ { u } = | u _ { r } - \hat { u } _ { r } |$
73
+ 7: $\begin{array} { r } { \mathcal { L } _ { p } = - \log ( \exp ( P ( u _ { s } | c _ { k _ { s } } ) ) / \sum _ { i = 1 } ^ { | \mathcal { C } _ { s } | } \exp ( P ( u _ { s } | c _ { i } ) ) ) } \end{array}$ , where $P ( u _ { s } | c _ { i } ) = \mathcal { N } ( u _ { s } | \mu _ { c _ { i } } , \Sigma _ { c _ { i } } )$
74
+ 8: Update $\theta _ { F _ { s } }$ , $\theta _ { D }$ , $\theta _ { G }$ by minimizing $\mathcal { L } _ { C E } , \mathcal { L } _ { v } , $ $\mathcal { L } _ { u }$ , and $\mathcal { L } _ { p }$ alternatively using separate optimizers.
75
+ 9: if $( i t e r \ \% \ U p d a t e I t e r = = 0 )$ ) then
76
+ 10: Recompute the sample mean $( \mu _ { c _ { i } } )$ and covariance $\left( \Sigma _ { c _ { i } } \right)$ of $F _ { s } \circ M ( x _ { s } )$ for $x _ { s }$ from class $c _ { i }$ ; $i = 1 , 2 . . . | \mathcal { C } _ { s } |$ (For $\mathcal { D } _ { n } ^ { ( b ) }$ : generate fresh latent-simulated negative samples using the updated priors)
77
+
78
+ the intermediate embedding $u _ { s } = F _ { s } \circ M ( x _ { s } )$ . Here, parameters of the normal distributions are computed during training as shown in line-10 of Algo. 1. A cross-entropy loss over these prior distributions is defined as ${ \mathcal { L } } _ { p }$ (line-7 in Algo. 1), to effectively enforce intra-class compactness with inter-class separability (progression from Fig. 4B to 4C). Motivated by generative variational auto-encoder (VAE) setup (Kingma & Welling, 2013), we introduce a feature decoder $G$ , which aims to minimize the cyclic reconstruction loss selectively for the samples from positive source categories $v _ { s }$ and randomly drawn samples $u _ { r }$ from the corresponding class priors (i.e. $\mathcal { L } _ { v }$ and $\mathcal { L } _ { u }$ , line-6 in Algo. 1). This along with a lower weightage $\alpha$ for the negative source categories (i.e. at the cross-entropy loss $\mathcal { L } _ { C E }$ , line-6 in Algo. 1) is incorporated to deliberately bias $F _ { s }$ towards the positive source samples, considering the level of unreliability of the generated negative dataset.
79
+
80
+ # 3.2 LEARNING IN THE DEPLOYMENT STAGE
81
+
82
+ 3.2.1 Challenges. We hypothesize that, the large number of negative source categories along with the positive source classes i.e. $\mathcal { C } _ { s } \cup \mathcal { C } _ { n }$ can be interpreted as a universal source dataset, which can subsume label-set $\mathcal { C } _ { t }$ of a wide range of target domains. Moreover, we seek to realize a unified adaptation algorithm, which can work for a wide range of category-gaps. However, a forceful adaptation of target samples to positive source categories will cause target private samples to be classified as an instance of the source private or the common label-set, instead of being classified as "unknown", i.e. one of the negative categories in ${ \mathcal { C } } _ { n }$ .
83
+
84
+ 3.2.2 Solution. In contrast to domain agnostic architectures (You et al., 2019; Cao et al., 2018a; Saito et al., 2018a), we resort to an architecture supporting domain specific features (Tzeng et al., 2017), as we must avoid disturbing the placement of source clusters obtained from the Procurement stage. This is an essential requirement to retain the task-dependent knowledge gathered from the source dataset. Thus, we introduce a domain specific feature extractor denoted as $F _ { t }$ , whose parameters are initialized from the fully trained $F _ { s }$ (see Fig. 3B). Further, we aim to exploit the learned generative classifier from the Procurement stage to complement for the purpose of separate ad-hoc networks (critic or discriminator) as utilized by the prior works (You et al., 2019; Cao et al., 2018b).
85
+
86
+ a) Source Similarity Metric (SSM). We define a weighting factor (SSM) for each target sample $x _ { t }$ , as $w ( x _ { t } )$ . A higher value of this metric indicates $x _ { t }$ ’s similarity towards the positive source categories, specifically inclined towards the common label space $\mathcal { C }$ . Similarly, a lower value of this metric indicates $x _ { t }$ ’s similarity towards the negative source categories ${ \mathcal { C } } _ { n }$ , showing its inclination towards the private target labels $\overline { { \mathcal { C } } } _ { t }$ . Let, $p _ { \bar { s } } , q _ { \bar { t } }$ be the distribution of source and target samples with labels in $\overline { { \mathcal { C } } } _ { s }$ and $\overline { { \mathcal { C } } } _ { t }$ respectively. We define, $p _ { c }$ and $q _ { c }$ to denote the distribution of samples from source and target domains belonging to the shared label-set $\mathcal { C }$ . Then, the SSM for the positive and negative source samples should lie on the two extremes, forming the following inequality:
87
+
88
+ $$
89
+ \begin{array} { r } { \mathbb { E } _ { x _ { n } \sim p _ { n } } w ( x _ { n } ) \approx \mathbb { E } _ { x _ { t } \sim q _ { \bar { t } } } w ( x _ { t } ) \ < \mathbb { E } _ { x _ { t } \sim q _ { c } } w ( x _ { t } ) < \mathbb { E } _ { x _ { s } \sim p _ { c } } w ( x _ { s } ) \approx \mathbb { E } _ { x _ { s } \sim p _ { \bar { s } } } w ( x _ { s } ) } \end{array}
90
+ $$
91
+
92
+ To formalize the SSM criterion we rely on the class probabilities defined at the output of source model only for the positive class labels, i.e. $\hat { y } ^ { ( k ) }$ for $k = 1 , 2 . . . | \mathcal { C } _ { s } |$ . Note that, $\hat { y } ^ { ( k ) }$ is obtained by performing softmax over $\left| \mathcal { C } _ { s } \right| + \left| \mathcal { C } _ { n } \right|$ categories as discussed in the Procurement stage. Finally, the
93
+
94
+ SSM and its complement are defined as,
95
+
96
+ $$
97
+ w ( x _ { t } ) = \operatorname* { m a x } _ { i = 1 , 2 \ldots | \mathcal { C } _ { s } | } \exp ( \hat { y } ^ { ( i ) } ) , \ a n d \ w ^ { \prime } ( x _ { t } ) = \operatorname* { m a x } _ { i = 1 , 2 \ldots | \mathcal { C } _ { s } | } \exp ( 1 - \hat { y } ^ { ( i ) } )
98
+ $$
99
+
100
+ We hypothesize that, the above definition will satisfy Eq. 1, as a result of the generative learning strategy adopted in the Procurement stage. In Eq. 2 the exponent is used to further amplify separation between target samples from the shared $\mathcal { C }$ and those from the private $\overline { { \mathcal { C } } } _ { t }$ label-set (see Fig. 5A).
101
+
102
+ b) Source-free domain adaptation. To perform domain adaptation, the objective function aims to move the target samples with higher SSM value towards the clusters of positive source categories and vice-versa at the frozen source embedding, $u$ -space (from the Procurement stage). To achieve this, parameters of only $F _ { t }$ network are allowed to be trained in the Deployment stage. However, the decision of weighting the loss on target samples towards the positive or negative source clusters is computed using the source feature extractor $F _ { s }$ i.e. the SSM in Eq. 2. We define, the deployment model as $h = \bar { D } \circ F _ { t } \circ M ( x _ { t } )$ using the target feature extractor, with softmax predictions over $K$ categories obtained as $\hat { z } ^ { ( k ) } = \sigma ( h ^ { ( k ) } )$ . Thus, the primary loss function for adaptation is defined as,
103
+
104
+ $$
105
+ \begin{array} { r } { \mathcal { L } _ { d 1 } = - w ( x _ { t } ) \log ( \sum _ { k = 1 } ^ { | \mathcal { C } _ { s } | } \hat { z } ^ { ( k ) } ) - w ^ { \prime } ( x _ { t } ) \log ( \sum _ { k = 1 + | \mathcal { C } _ { s } | } ^ { | \mathcal { C } _ { s } | + | \mathcal { C } _ { n } | } \hat { z } ^ { ( k ) } ) } \end{array}
106
+ $$
107
+
108
+ Additionally, in the absence of label information, there would be uncertainty in the predictions $\hat { z } ^ { ( k ) }$ as a result of distributed class probabilities. This leads to a higher entropy for such samples. Entropy minimization (Grandvalet & Bengio, 2005; Long et al., 2016) is adopted in such scenarios to move the target samples close to the highly confident regions (i.e. positive and negative cluster centers from the Procurement stage) of the classifier’s feature space. However, it has to be done separately for positive and negative source categories based on the SSM values of individual target samples to effectively distinguish the target-private set from the full target dataset. To achieve this, we define two different class probability vectors separately for the positive and negative source classes denoted as, $\begin{array} { r } { \tilde { z } _ { s } ^ { ( i ) } = \exp ( h ^ { ( i ) } ) / { \sum _ { j = 1 } ^ { | \mathcal { C } _ { s } | } } \exp ( h ^ { ( j ) } ) } \end{array}$ and $\tilde { z } _ { n } ^ { ( i ) } = \exp ( h ^ { ( i + | \mathcal { C } _ { s } | ) } ) / { \sum _ { j = 1 } ^ { | \mathcal { C } _ { n } | } \exp ( h ^ { ( j + | \mathcal { C } _ { s } | ) } ) }$ respectively (see Fig. 3B). Entropy of the target samples in the positive and negative regimes of the source classifier is obtained as $\begin{array} { r } { \bar { H } _ { s } ( x _ { t } ) = - \sum _ { i = 1 } ^ { | \mathcal { C } _ { s } | } \tilde { z } _ { s } ^ { ( i ) } \log \tilde { z } _ { s } ^ { ( i ) } } \end{array}$ and $\begin{array} { r } { H _ { n } ( x _ { t } ) { \bf \bar { \Psi } } = - \sum _ { i = 1 } ^ { \vert { \mathcal C } _ { n } \vert } \tilde { z } _ { n } ^ { ( i ) } \log \tilde { z } _ { n } ^ { ( i ) } } \end{array}$ respectively. Consequently, the entropy minimization loss is formalized as,
109
+
110
+ $$
111
+ \begin{array} { r } { \mathcal { L } _ { d 2 } = w ( x _ { t } ) H _ { s } ( x _ { t } ) + w ^ { \prime } ( x _ { t } ) H _ { n } ( x _ { t } ) } \end{array}
112
+ $$
113
+
114
+ Thus, the final loss function for adapting the parameters of $F _ { t }$ is presented as $\mathcal { L } _ { d } = \mathcal { L } _ { d 1 } + \beta \mathcal { L } _ { d 2 }$ .
115
+ Here $\beta$ is a hyper-parameter controlling the importance of entropy minimization during adaptation.
116
+
117
+ # 4 EXPERIMENTS
118
+
119
+ We perform a thorough evaluation of the proposed source-free, universal domain adaptation framework against prior state-of-the-art models across multiple datasets. We also provide a comprehensive ablation study to establish generalizability of the approach across a variety of label-set relationships and justification of the various model components.
120
+
121
+ # 4.1 EXPERIMENTAL SETUP
122
+
123
+ Datasets. For all the following datasets, we resort to the experimental settings inline with the recent work by You et al. (2019) (UAN). Office-Home (Venkateswara et al., 2017) dataset consists of images from 4 different domains - Artistic $( \mathbf { A r } )$ , Clip-art (Cl), Product $( \mathbf { P r } )$ and Real-world $( \mathbf { R } \mathbf { w } )$ . Alphabetically, the first 10 classes are selected as $\mathcal { C }$ , the next 5 classes as $\overline { { \mathcal { C } } } _ { s }$ , and the rest 50 as $\overline { { \mathcal { C } } } _ { t }$ . VisDA2017 (Peng et al., 2018) dataset comprises of 12 categories with synthetic images as the source domain and natural images as the target domain, out of which, the first 6 are chosen as $\mathcal { C }$ , the next 3 as $\overline { { \mathcal { C } } } _ { s }$ and the rest as $\overrightarrow { \mathcal { C } } _ { t }$ . Office-31 (Saenko et al., 2010) dataset contains images from 3 distinct domains - Amazon (A), DSLR $\mathbf { \eta } ^ { ( \mathbf { D } ) }$ and Webcam $( \mathbf { W } )$ . We use the 10 classes shared by Office-31 and Caltech-256 (Gong et al., 2012) to construct the shared label-set $\mathcal { C }$ and alphabetically select the next 10 as $\overline { { \mathcal { C } } } _ { s }$ , with the remaining 11 classes contributing to $\overline { { \mathcal { C } } } _ { t }$ . To evaluate scalability, ImageNet-Caltech is also considered with 84 common classes inline with the setting in You et al. (2019).
124
+
125
+ Simulation of labeled negative samples. To simulate negative labeled samples for training in the Procurement stage, we first sample a pair of images, each from different categories of $\mathcal { C } _ { s }$ , to create unique negative classes in ${ \mathcal { C } } _ { n }$ . Note that, we impose no restriction on how the hypothetical classes are created (e.g. one can composite non-animal with animal). A random mask is defined which splits the images into two complementary regions using a quadratic spline passing through a central image region (see Appendix Algo. 2). Then, the negative image is created by merging alternate mask regions as shown in Fig. 3A. For the $\mathbf { I } { \xrightarrow { } } \mathbf { C }$ task of ImageNet-Caltech, the source domain (ImageNet), consisting of 1000 classes, results in a large number of possible negative classes (i.e. $| \mathcal { C } _ { n } | = | \bar { C } _ { s } | _ { C _ { 2 } } )$ . We address this by randomly selecting only 600 of these negative classes for ImageNet(I), and 200 negative classes for Caltech(C) in the task $\mathbf { C } { \xrightarrow { } } \mathbf { I } .$ In a similar fashion, we generate latent-simulated negative samples only for the selected negative classes in these datasets. Consequently, we compare two models with different Procurement stage training - (i) USFDA-a: using image-composition as negative dataset , and (ii) USFDA-b: using latent-simulated negative samples as the negative dataset. We use USFDA- $a$ for most of our ablation experiments unless mentioned explicitly.
126
+
127
+ Table 1: Average per-class accuracy $( \mathcal { T } _ { a v g } )$ for universal-DA tasks on Office-Home dataset (with $| C | / | \mathcal { C } _ { s } \cup \mathcal { C } _ { t } | = 0 . \bar { 1 } 5 )$ . Scores for the prior works are directly taken from UAN (You et al., 2019).
128
+
129
+ <table><tr><td rowspan="2">Method</td><td colspan="10">Office-Home Ar→Cl Ar-→Pr Ar-→RwCl→Ar Cl-→PrCl-→RwPr-→Ar Pr-→Cl Pr-→RwRw→Ar Rw-→Cl Rw-→Pr</td><td colspan="4"></td></tr><tr><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td>Avg</td></tr><tr><td>ResNet (He etal.,2016)</td><td>59.37</td><td>76.58</td><td>87.48</td><td>69.86</td><td>71.11</td><td>81.66</td><td></td><td>73.72</td><td>56.30</td><td>86.07</td><td>78.68</td><td>59.22</td><td>78.59</td><td>73.22</td></tr><tr><td>IWAN (Zhang et al., 2018b)</td><td>52.55</td><td>81.40</td><td>86.51</td><td>70.58</td><td>70.99</td><td>85.29</td><td>74.88</td><td></td><td>57.33</td><td>85.07</td><td>77.48</td><td>59.65</td><td>78.91</td><td>73.39</td></tr><tr><td>PADA (Zhang et al., 2018b)</td><td>39.58</td><td>69.37</td><td>76.26</td><td>62.57</td><td>67.39</td><td>77.47</td><td>48.39</td><td></td><td>35.79</td><td>79.60</td><td>75.94</td><td>44.50</td><td>78.10</td><td>62.91</td></tr><tr><td>ATI (Busto et al., 2017)</td><td>52.90</td><td>80.37</td><td>85.91</td><td>71.08</td><td>72.41</td><td>84.39</td><td>74.28</td><td></td><td>57.84</td><td>85.61</td><td>76.06</td><td>60.17</td><td>78.42</td><td>73.29</td></tr><tr><td>OSBP (Saito et al.,2018b)</td><td>47.75</td><td>60.90</td><td>76.78</td><td>59.23</td><td>61.58</td><td>74.33</td><td>61.67</td><td></td><td>44.50</td><td>79.31</td><td>70.59</td><td>54.95</td><td>75.18</td><td>63.90</td></tr><tr><td>UAN (You et al., 2019)</td><td>63.00</td><td>82.83</td><td>87.85</td><td>76.88</td><td>78.70</td><td>85.36</td><td>78.22</td><td></td><td>58.59</td><td>86.80</td><td>83.37</td><td>63.17</td><td>79.43</td><td>77.02</td></tr><tr><td colspan="10">Source-free adaptation</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>Ours USFDA-a</td><td>63.35</td><td>83.30</td><td>89.35</td><td>70.96</td><td>72.34</td><td>86.09</td><td>78.53</td><td>60.15</td><td></td><td>87.35</td><td>81.56</td><td>63.17</td><td>88.23</td><td>77.03</td></tr><tr><td>Ours USFDA-b</td><td>62.46</td><td>82.71</td><td>88.26</td><td>71.10</td><td>70.88</td><td>85.75</td><td>78.21</td><td>59.18</td><td></td><td>86.05</td><td>82.17</td><td>63.22</td><td>87.68</td><td>76.47</td></tr></table>
130
+
131
+ # 4.2 EVALUATION METHODOLOGY
132
+
133
+ Average accuracy on Target dataset, $\mathcal { T } _ { a v g }$ . We resort to the evaluation protocol proposed in the VisDA2018 Open-Set Classification challenge. Accordingly, all the target private classes are grouped into a single "unknown" class and the metric reports the average of per-class accuracy over $| \mathcal { C } _ { s } | + 1$ classes. In the proposed framework a target sample is marked as "unknown", if it is classified $( a r g m a x _ { k } \hat { z } ^ { ( k ) } )$ into any of the negative $| { \mathcal { C } } _ { n } |$ classes out of total $\left| \mathcal { C } _ { s } \right| + \left| \mathcal { C } _ { n } \right|$ categories. In contrast, UAN (You et al., 2019) relies on a sensitive hyperparameter, as a threshold on the sample-level weighting, to mark a target sample as "unknown". Also note that, our method is completely source-free during the Deployment stage, while all other methods have access to the full source-data.
134
+
135
+ Accuracy on Target-Unknown data, $\mathcal { T } _ { u n k }$ . We evaluate the target unknown accuracy, $\mathcal { T } _ { u n k }$ , as the proportion of actual target private samples (i.e. $\{ ( x _ { t } , y _ { t } ) : y _ { t } \in \hat { \mathcal { C } } _ { t } \}$ ) being classified as "unknown" after adaptation. Note that, UAN (You et al., 2019) does not report $\mathcal { T } _ { u n k }$ which is a crucial metric to evaluate the vulnerability of the model after its deployment in the target environment. The $\tau _ { a v g }$ metric fails to capture this as a result of class-imbalance in the Open-set scenario (Saito et al., 2018b). Hence, to realize a common evaluation ground, we train the UAN implementation provided by the authors (You et al., 2019) and denote it as $\mathrm { U A N ^ { \ast } }$ in further sections of this paper. We observe that, the UAN(You et al., 2019) training algorithm is often unstable with a decreasing trend of $\mathcal { T } _ { u n k }$ and $\mathcal { T } _ { a v g }$ over increasing training iterations. We thus report the mean and standard deviation of the peak values of $\mathcal { T } _ { u n k }$ and $\mathcal { T } _ { a v g }$ achieved by $\mathrm { U A N ^ { * } }$ , over 5 separate runs on Office-31 dataset (see Table 7).
136
+
137
+ Implementation Details. We implement our network in PyTorch and use ResNet-50 (He et al., 2016) as the backbone-model $M$ , pre-trained on ImageNet (Russakovsky et al., 2015) inline with UAN (You et al., 2019). The complete architecture of other components with fully-connected layers is provided in the Supplementary. A sensitivity analysis of the major hyper-parameters used in the proposed framework is provided in Fig. 5B-C, and Appendix Fig. 8B. In all our ablations across the datasets, we fix the hyperparameters values as $\alpha = 0 . 2$ and $\beta = 0 . 1$ . We utilize Adam optimizer (Kingma & Ba, 2014) with a fixed learning rate of 0.0001 for training in both Procurement and Deployment stage (see Appendix for the code). For the implementation of $\mathrm { U A N ^ { \ast } }$ , we use the hyper-parameter value $w _ { 0 } = - 0 . 5$ , as specified by the authors for the task $\mathbf { A } { \to } \mathbf { D }$ in Office-31 dataset.
138
+
139
+ # 4.3 DISCUSSION
140
+
141
+ a) Comparison with prior arts. We compare our approach with UAN You et al. (2019), and other prior methods. The results are presented in Table 1 and Table 2. Clearly, our framework achieves stateof-the-art results even in a source-free setting on several tasks. Particularly in Table 2, we present the target-unknown accuracy $\mathcal { T } _ { u n k }$ on various dataset. It also holds the mean and standard-deviation for both the accuracy metrics computed over 5 random initializations in the Office-31 dataset (the last six rows). Our method is able to achieve much higher $\mathcal { T } _ { u n k }$ than $\mathrm { U A N ^ { * } }$ (You et al., 2019), highlighting our superiority as a result of the novel learning approach incorporated in both Procurement and Deployment stages. Note that, both USFDA-a and USFDA- $^ { b }$ yield similar performance across a wide range of standard benchmarks. We also perform a characteristic comparison of algorithm complexity in terms of the amount of learnable parameters and training time. In contrast to UAN, the proposed framework offers a much simpler adaptation algorithm devoid of utilization of ad-hoc networks like adversarial discriminator and additional finetuning of the ResNet-50 backbone. Parameter size and training time; a) Ours procurement (USFDA-a): [11.1M, 380s], b) Ours deployment: [3.5M, 44s], c) UAN (You et al., 2019): [26.7M, 450s] (in a consistent setting). The significant computational advantage in the Deployment stage makes USFDA highly suitable for real-time adaptation.
142
+
143
+ ![](images/beecc87547cd422310b3d251911a670d1c49a0a8a38fc3e10aa458c4f0157c5b.jpg)
144
+ Figure 5: Ablative analysis on the task $\mathbf { A } { \to } \mathbf { D }$ in Office-31 dataset. A) Histogram of SSM values of $x _ { t }$ separately for target-private and target-shared samples at the Procurement iteration 100 (top) and 500 (bottom). B) The sensitivity curve for $\beta$ shows marginally stable adaptation accuracy for a wide-range of values. C) A marginal increase in $\mathcal { T } _ { a v g }$ is observed with increase in $| { \mathcal { C } } _ { n } |$ .
145
+
146
+ b) Does SSM satisfy the expected inequality? Effectiveness of the proposed learning algorithm, in case of source-free deployment, relies on the formulation of SSM, which is expected to satisfy Eq. 1. Fig. 5A shows a histogram of the SSM separately for samples from target-shared (blue) and target-private (red) label space. The success of this metric is attributed to the generative nature of Procurement stage, which enables the source model to distinguish between the marginally more negative target-private samples as compared to the samples from the shared label space.
147
+
148
+ c) Sensitivity to hyper-parameters. As we tackle DA in a source-free setting simultaneously intending to generalize across varied category-gaps, a low sensitivity to hyperparameters would further enhance our practical usability. To this end, we fix certain hyperparameters for all our ablations (also in Fig. 6C) even across datasets (i.e. $\alpha = 0 . 2$ , $\beta = 0 . 1$ ). Thus, one can treat them as global-constants with $| { \mathcal { C } } _ { n } |$ being the only hyperparameter, as variations in one by fixing the others yield complementary effect on regularization in the Procurement stage. A thorough analysis reported in the appendix Fig. 8, clearly demonstrates the low-sensitivity of our model to these hyperparameters.
149
+
150
+ d) Generalization across category-gap. One of the key objectives of the proposed framework is to effectively operate in the absence of the knowledge of label-set relationships. To evaluate it in the most compelling manner, we propose a tabular form shown in Fig. 6A. We vary the number of private classes for target and source along $\mathbf { X }$ and y axis respectively, with a fixed $| \mathcal { C } _ { s } \cup \mathcal { C } _ { t } | = 3 1$ . We compare the $\mathcal { T } _ { a v g }$ metric at the corresponding table instances, shown in Fig. 6B-C. The results clearly highlight superiority of the proposed framework specifically for the more practical scenarios (close to the diagonal instances) as compared to the unrealistic Closed-set setting ( $| \overline { { \mathcal { C } } } _ { s } | = | \overline { { \mathcal { C } } } _ { t } | = 0 .$ ).
151
+
152
+ ![](images/c0d5f77419e52fd8de8bdb652c3c37e39299a765ec012e49797737c7cdcca0fd.jpg)
153
+ Figure 6: Comparison across varied label-set relationships for the task $\mathbf { A } { \to } \mathbf { D }$ in Office-31 dataset. A) Visual representation of label-set relationships and $\mathcal { T } _ { a v g }$ at the corresponding instances for $\mathbf { B }$ ) $\mathrm { U A N ^ { * } }$ (You et al., 2019) and C) ours source-free model. Effectively, the direction along $\mathbf { X }$ -axis (blue horizontal arrow) characterizes increasing Open-set complexity. The direction along y-axis (red vertical arrow) shows increasing complexity of Partial DA scenario. The pink diagonal arrow denotes the effect of decreasing shared label space.
154
+
155
+ e) DA in absence of shared categories. In universal adaptation, we seek to transfer the knowledge of "class-separability criterion" obtained from the source domain to the deployed target environment. More concretely, it is attributed to the segregation of data samples based on some expected characteristics, such as classification of objects according to their pose, color, or shape etc. To quantify this, we consider an extreme case where $\mathcal { C } _ { s } \cap \mathcal { C } _ { t } = \emptyset$ $\mathbf { A } { } \mathbf { D }$ in Office-31 with $| \mathcal { C } _ { s } | = 1 5$ , $| \mathcal { C } _ { t } | = 1 6 $ ). Allowing access to a single labeled target sample from each category in $\overline { { \mathcal { C } } } _ { t } = \mathcal { C } _ { t }$ , we aim to obtain a one-shot recognition accuracy (assignment of cluster index or class label using the one-shot samples as the cluster center at $F _ { t } \circ M ( x _ { t } ) ) ,$ ) to quantify the above metric. We obtain $6 4 . 7 2 \%$ accuracy for the proposed framework as compared to $1 3 . 4 3 \%$ for $\mathrm { U A N ^ { * } }$ (You et al., 2019). This strongly validates our superior knowledge transfer capability as a result of the generative classifier with labeled negative samples complementing for the target-private categories.
156
+
157
+ f) Dependency on the simulated negative dataset. Conceding that a combinatorial amount of negative labels can be created, we evaluate the scalability of the proposed approach, by varying the number of negative classes in the Procurement stage by selecting 0, 4, 8, 64, 150 and 190 negative classes as reported in the $\mathrm { X }$ -axis of Fig. 5C. For the case of 0 negative classes, denoted as $| \mathcal { C } _ { n } | ^ { * } = 0$ in Fig. 5C, we synthetically generate random negative features at the intermediate level $u$ , which are at least 3-sigma away from each of the positive source priors $P ( u _ { s } | c _ { i } )$ . We then make use of these feature samples along with positive image samples, to train a $\left( | \mathcal { C } _ { s } | + 1 \right)$ class Procurement model with a single negative class. The results are reported in Fig. 5C on the $\mathbf { A } { \to } \mathbf { D }$ task of Office-31 dataset with category relationship inline with the setting in Table 7. We observe an acceptable drop in accuracy with decrease in number of negative classes, hence validating scalability of the approach for large-scale classification datasets (such as ImageNet). Similarly, we also evaluated our framework by combining three or more images to form such negative classes. An increasing number of negative classes $( { \breve { c _ { s } } } | _ { C _ { 3 } } > { | \mathscr { C } _ { s } | } _ { C _ { 2 } } )$ attains under-fitting on positive source categories (similar to Fig. 5C, where accuracy reduces beyond a certain limit because of over regularization).
158
+
159
+ # 5 CONCLUSION
160
+
161
+ We have introduced a novel source-free, universal domain adaptation framework, acknowledging practical domain adaptation scenarios devoid of any assumption on the source-target label-set relationship. In the proposed two-stage framework, learning in the Procurement stage is found to be highly crucial, as it aims to exploit the knowledge of class-separability in the most general form with enhanced robustness to out-of-distribution samples. Besides this, success in the Deployment stage is attributed to the well-designed learning objectives effectively utilizing the source similarity criterion. This work can be served as a pilot study towards learning efficient inheritable models in future.
162
+
163
+ # REFERENCES
164
+
165
+ Mahsa Baktashmotlagh, Masoud Faraki, Tom Drummond, and Mathieu Salzmann. Learning factorized representations for open-set domain adaptation. In International Conference on Learning Representations, 2019. 1
166
+
167
+ Konstantinos Bousmalis, Nathan Silberman, David Dohan, Dumitru Erhan, and Dilip Krishnan. Unsupervised pixel-level domain adaptation with generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2017. 2, 3
168
+
169
+ Zhangjie Cao, Mingsheng Long, Jianmin Wang, and Michael I Jordan. Partial transfer learning with selective adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2018a. 1, 2, 5
170
+
171
+ Zhangjie Cao, Lijia Ma, Mingsheng Long, and Jianmin Wang. Partial adversarial domain adaptation In Proceedings of the European Conference on Computer Vision, 2018b. 1, 2, 5
172
+
173
+ Yi-Hsin Chen, Wei-Yu Chen, Yu-Ting Chen, Bo-Cheng Tsai, Yu-Chiang Frank Wang, and Min Sun. No more discrimination: Cross city adaptation of road scene segmenters. In Proceedings of the IEEE International Conference on Computer Vision, 2017. 1
174
+
175
+ Lixin Duan, Ivor W Tsang, and Dong Xu. Domain transfer multiple kernel learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(3):465–479, 2012. 2
176
+
177
+ Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The Journal of Machine Learning Research, 17(1):2096–2030, 2016. 1
178
+
179
+ ZongYuan Ge, Sergey Demyanov, Zetao Chen, and Rahil Garnavi. Generative openmax for multiclass open set classification. arXiv preprint arXiv:1707.07418, 2017. 1
180
+
181
+ Boqing Gong, Yuan Shi, Fei Sha, and Kristen Grauman. Geodesic flow kernel for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2012. 6
182
+
183
+ Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in neural information processing systems, 2014. 2
184
+
185
+ Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. In Advances in neural information processing systems, 2005. 6
186
+
187
+ Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. 7, 8
188
+
189
+ Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei A Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In International Conference on Learning Representations, 2018. 2
190
+
191
+ Lanqing Hu, Meina Kan, Shiguang Shan, and Xilin Chen. Duplex generative adversarial network for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2018. 2
192
+
193
+ Guoliang Kang, Liang Zheng, Yan Yan, and Yi Yang. Deep adversarial attention alignment for unsupervised domain adaptation: the benefit of target expectation maximization. In Proceedings of the European Conference on Computer Vision, 2018. 2
194
+
195
+ Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7
196
+
197
+ Diederik $\mathrm { \bf P }$ Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 5
198
+
199
+ Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, 2012. 1
200
+
201
+ Abhishek Kumar, Prasanna Sattigeri, Kahini Wadhawan, Leonid Karlinsky, Rogerio Feris, Bill Freeman, and Gregory Wornell. Co-regularized alignment for unsupervised domain adaptation. In Advances in neural information processing systems, 2018. 1
202
+
203
+ Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In International Conference on Machine Learning, 2015. 1, 2, 17
204
+
205
+ Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Unsupervised domain adaptation with residual transfer networks. In Advances in neural information processing systems, 2016. 1, 2, 6
206
+
207
+ Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. In Advances in neural information processing systems, 2018. 2, 17
208
+
209
+ Zelun Luo, Yuliang Zou, Judy Hoffman, and Li F Fei-Fei. Label efficient learning of transferable representations acrosss domains and tasks. In Advances in neural information processing systems, 2017. 2
210
+
211
+ Jogendra Nath Kundu, Phani Krishna Uppala, Anuj Pahuja, and R Venkatesh Babu. Adadepth: Unsupervised content congruent adaptation for depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2018. 1, 2
212
+
213
+ Pau Panareda Busto and Juergen Gall. Open set domain adaptation. In Proceedings of the IEEE International Conference on Computer Vision, 2017. 2
214
+
215
+ Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge. 2018. 6
216
+
217
+ Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in neural information processing systems, 2015. 1
218
+
219
+ Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211–252, 2015. 3, 7
220
+
221
+ Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell. Adapting visual category models to new domains. In Proceedings of the European Conference on Computer Vision, 2010. 2, 6, 16, 20
222
+
223
+ Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2018a. 3, 5
224
+
225
+ Kuniaki Saito, Shohei Yamamoto, Yoshitaka Ushiku, and Tatsuya Harada. Open set domain adaptation by backpropagation. In Proceedings of the European Conference on Computer Vision, 2018b. 2, 7, 8
226
+
227
+ Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. In Advances in neural information processing systems, 2016. 3
228
+
229
+ Swami Sankaranarayanan, Yogesh Balaji, Carlos D Castillo, and Rama Chellappa. Generate to adapt: Aligning domains using generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2018. 2, 3
230
+
231
+ Hidetoshi Shimodaira. Improving predictive inference under covariate shift by weighting the loglikelihood function. Journal of statistical planning and inference, 90(2):227–244, 2000. 1
232
+
233
+ Rui Shu, Hung Bui, Hirokazu Narui, and Stefano Ermon. A DIRT-t approach to unsupervised domain adaptation. In International Conference on Learning Representations, 2018. 4
234
+
235
+ Eric Tzeng, Judy Hoffman, Trevor Darrell, and Kate Saenko. Simultaneous deep transfer across domains and tasks. In Proceedings of the IEEE International Conference on Computer Vision, 2015. 1
236
+
237
+ Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2017. 1, 2, 3, 5, 17
238
+
239
+ Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2017. 6
240
+
241
+ Xuezhi Wang and Jeff Schneider. Flexible transfer learning under support and model shift. In Advances in neural information processing systems, 2014. 2
242
+
243
+ Zirui Wang, Zihang Dai, Barnabás Póczos, and Jaime Carbonell. Characterizing and avoiding negative transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2019. 2
244
+
245
+ Kaichao You, Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I. Jordan. Universal domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, June 2019. 2, 5, 6, 7, 8, 9, 16, 17, 18, 19, 20
246
+
247
+ Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018a. 4
248
+
249
+ Jing Zhang, Zewei Ding, Wanqing Li, and Philip Ogunbona. Importance weighted adversarial nets for partial domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2018b. 1, 2, 7, 8
250
+
251
+ Kun Zhang, Bernhard Schölkopf, Krikamol Muandet, and Zhikun Wang. Domain adaptation under target and conditional shift. In International Conference on Machine Learning, 2013. 2
252
+
253
+ Weichen Zhang, Wanli Ouyang, Wen Li, and Dong Xu. Collaborative and adversarial network for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2018c. 1
254
+
255
+ Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 16
256
+
257
+ Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, 2017. 2
258
+
259
+ # A APPENDIX
260
+
261
+ This appendix is organized as follows, • Implementation details – Procurement Stage. – Deployment Stage.
262
+
263
+ • Ablation Studies and Additional Results
264
+
265
+ – Pretraining the backbone network on Places instead of ImageNet.
266
+ – Space and Time Complexity Analysis.
267
+ – Varying label-set relationship.
268
+ – Sensitivity analysis.
269
+ – Closed-set adaptation.
270
+ – Accuracy on source dataset post Procurement.
271
+ – Incremental one-shot classification.
272
+ – Feature Space Visualization.
273
+
274
+ • Miscellaneous
275
+
276
+ – Specification of Computing Resources. – References to code
277
+
278
+ # B IMPLEMENTATION DETAILS
279
+
280
+ In this section, we describe the architecture and the training process used for the Procurement and Deployment stages of our approach.
281
+
282
+ # B.1 PROCUREMENT STAGE
283
+
284
+ a) Design of classifier $D$ used in the Procurement stage. Keeping in mind the possibility of an additional domain shift after performing adaptation (e.g. encountering domain W after performing the adaptation $\mathbf { A } \mathbf { D }$ in Office-31 dataset), we design the classifier’s architecture in a manner which allows for dynamic modification in the number of negative classes post-procurement. We achieve this by maintaining two separate classifiers during Procurement - $D _ { s r c }$ , that operates on the positive source classes, and, $D _ { n e g }$ that operates on the negative source classes (see architecture in Table 5). The final classification score is obtained by computing softmax over the concatenation of logit vectors produced by $D _ { s r c }$ and $D _ { n e g }$ . Therefore, the model can be retrained on a different number of negative classes post deployment (using another negative class classifier $D _ { n e g . } ^ { \prime }$ ), thus preparing it for a subsequent adaptation step to another domain.
285
+
286
+ b) Negative dataset generation. We propose two methods to generate negative samples for the Procurement stage, and name the models trained subsequently as USFDA-a and USFDA-b. Here, we describe the two processes:
287
+
288
+ • Using image-composition for $\mathcal { D } _ { n } ^ { ( a ) }$ (USFDA-a). In the presence of domain knowledge (knowledge of the task at hand, i.e. object recognition using images), we generate the negative dataset $\mathcal { D } _ { n } ^ { ( a ) }$ by compositing images taken from different classes, as described in Algo. 2. We generate random masks using quadratic splines passing through a central image region (lines 3-9). Using these masks, we merge alternate regions of the images, both horizontally and vertically, resulting in 4 negative images for each pair of images (lines 10-13). To effectively cover the inter-class negative region, we randomly sample image pairs from $D _ { s }$ belonging to different classes, however we do not impose any constraint on how the classes are selected (for e.g. one can composite images from an animal and a non-animal class). We choose 5000 pairs for tasks on Office-31, Office-Home and VisDA datasets, and 12000 for ImageNet-Caltech. Since the input source distribution $( p )$ is fixed we first synthesize a negative dataset offline (instead of creating them on the fly) to ensure finiteness of the training set. The training algorithm for USFDA-a is given in Algo. 1.
289
+
290
+ # Algorithm 2 Image-composition algorithm
291
+
292
+ . horizontal splicing . vertical splicing
293
+
294
+ 1: input: Image pair $( I _ { 1 } , I _ { 2 } ) \in \mathcal { D } _ { s }$ . (image shape $H _ { \mathrm { X } } W _ { \mathrm { X } } 3 = 2 2 4 \mathrm { x } 2 2 4 \mathrm { x } 3 )$
295
+ 2: $k \gets 3 0$
296
+ 3: $x _ { 1 }$ , $x _ { 2 }$ , $y _ { 1 }$ , $y _ { 2 } \gets \mathrm { r a n d } ( 0 , W )$ , rand $( 0 , W )$ , rand $( 0 , H )$ , rand $( 0 , H )$
297
+ 4: $c _ { x }$ , $c _ { y } \gets$ rand $( W / 2 - k , W / 2 + k )$ , ran $\mathrm { { 1 } } ( H / 2 - k / 3 , H / 2 + k / 3 )$
298
+ 5: $d _ { x }$ , $d _ { y } \gets \mathrm { r a n d } ( W / 2 - k / 3 , W / 2 + k / 3 )$ , rand $( H / 2 - k , H / 2 + k )$
299
+ 6: $s _ { 1 } $ quadratic_interpolation $( [ ( 0 , y _ { 1 } )$ , $\left( c _ { x } , c _ { y } \right)$ , (223, y2)])
300
+ 7: $s _ { 2 } $ quadratic_interpolation $\iota ( [ ( x _ { 1 } , 0 ) , ( d _ { x } , d _ { y } ) , ( x _ { 2 } , 2 2 3 ) ] )$
301
+ 8: $m _ { 1 } \gets$ mask region below $s _ { 1 }$
302
+ 9: $m _ { 2 } \gets$ mask region to the left of $s _ { 2 }$
303
+ 10: $I _ { a } \gets m _ { 1 } * I _ { 1 } + \left( 1 - m _ { 1 } \right) * I _ { 2 }$
304
+ 11: $I _ { b } \gets m _ { 2 } * I _ { 1 } + ( 1 - m _ { 2 } ) * I _ { 2 }$
305
+ 12: $I _ { c } \gets m _ { 1 } * I _ { 2 } + ( 1 - m _ { 1 } ) * I _ { 1 }$
306
+ 13: $I _ { d } m _ { 2 } * I _ { 2 } + ( 1 - m _ { 2 } ) * I _ { 1 }$
307
+ 14: return ${ l _ { a } , I _ { b } , I _ { c } , I _ { d } }$
308
+
309
+ # Algorithm 3 Dataset generation using latent-simulated negative samples
310
+
311
+ 1: input: class-wise source priors $\mathcal { N } ( \mu _ { c _ { j } } , \Sigma _ { c _ { j } } )$ , global source prior $\mathcal { N } ( \boldsymbol { \mu } , \boldsymbol { \Sigma } )$ , number of required
312
+ samples $n$ , source classifier $D _ { s r c }$ . $| |$ signifies an Append Operation
313
+ 2: $\tilde { \mathcal { U } } \gets \{ \} ; \tilde { \mathcal { V } } \gets \{ \}$
314
+ 3: while $| { \tilde { \mathcal { U } } } | \leq n$ do
315
+ 4: Let $\lambda _ { c _ { j } }$ and $l _ { c _ { j } }$ be the maximum eigen value and the corresponding eigen vector
316
+ of $\Sigma _ { c _ { j } }$ , for each class $c _ { j }$
317
+ 5: $\tilde { u } _ { r } \sim \mathcal N ( \mu , \Sigma )$
318
+ 6: if $P ( \tilde { u } _ { r } | c _ { j } ) < P ( \mu _ { c _ { j } } + 3 * \sqrt { \lambda _ { c _ { j } } } * l _ { c _ { j } } \mid c _ { j } )$ for all class $c _ { j }$ then
319
+ 7: $\hat { y } \sigma ( D _ { s r c } ( \tilde { u } _ { r } ) )$
320
+ 8: $\tilde { y } _ { r } \gets$ assign the negative class based on the top-2 confident classes in $\hat { y }$
321
+ 9: $\mathcal { \tilde { U } } \mathcal { \tilde { U } } \| \mathcal { \bar { u } } _ { r } ; \mathcal { \tilde { y } } \mathcal { \tilde { y } } \| \tilde { y } _ { r }$
322
+ 10: else
323
+ 11: reject $\tilde { u } _ { r }$
324
+ 12: return U˜, Y˜
325
+
326
+ • Using latent-simulated negative samples for $\mathcal { D } _ { n } ^ { ( b ) }$ (USFDA-b): Here, we perform rejection sampling as given in Algorithm 3. Here, we obtain a sample from the global source prior $P ( u _ { s } ) = \mathcal { N } ( u _ { s } | \boldsymbol { \mu } , \Sigma )$ , where $\mu$ and $\Sigma$ are the mean and covariance computed at $u$ -space over all the positive source image samples.We reject the sample if it lies within the 3-sigma bound of any class (i.e. we keep the sample if it is far away from all source class-priors, $\mathcal { N } ( \mu _ { c _ { i } } , \Sigma _ { c _ { i } } ) )$ , as shown in lines 6 to 11 in Algo. 3. A sample selected in this fashion is expected to lie in an intermediate region between the source class priors. The two classes in the vicinity of the sample are then determined by obtaining the two most confident class predictions given by the classifier $D _ { s r c }$ (lines 7 and 8). Using this pair of classes, we assign a unique negative class label to the sample which corresponds to the intermediate region between the pair of classes. Note, to learn the arrangement of positive and negative clusters, the feature extractor $F _ { s }$ must be trained using negative samples. We do this by passing the sampled latent-simulated negative instance $( \tilde { u } _ { r } )$ through the decoder-encoder pair, (i.e. $D \circ F _ { s } \circ { \cal G } ( \tilde { u } _ { r } ) )$ , and enforcing the cross-entropy loss to classify them into the respective negative class. The training algorithm for USFDA- $^ { b }$ is given in Algo. 4.
327
+
328
+ c) Justification of ${ \mathcal { L } } _ { p }$ . The cross-entropy loss on the likelihoods (referred as ${ \mathcal { L } } _ { p }$ in the paper) not only enforces intra-class compactness but also ensures inter-class separability in the embedding space, $u$ . Since the negative samples are only an approximation of future target private classes expected to be encountered during deployment, we choose not to employ this loss for them. Such a training procedure, eventually results in a natural development of bias towards the confident positive source classes. This subsequently leads to the placement of source clusters in a manner which enables source-free adaptation (See Fig. 4).
329
+
330
+ # Algorithm 4 Training algorithm for $U S F D A { - } b$ in the Procurement stage
331
+
332
+ 1: input: $( x _ { s } , y _ { s } ) \in \mathcal { D } _ { s }$ ; $\theta _ { F _ { s } }$ , $\theta _ { D }$ , $\theta _ { G }$ : Parameters of $F _ { s }$ , $D$ and $G$ respectively.
333
+ 2: initialization: pretrain $\{ \theta _ { F _ { s } } , \theta _ { D } \}$ using cross-entropy loss on $( x _ { s } , y _ { s } )$ , then, compute the sample mean $\mu _ { c _ { i } }$
334
+ and covariance $\Sigma _ { c _ { i } }$ of $F _ { s } \circ M ( x _ { s } )$ for $x _ { s }$ from class $c _ { i }$ , for $i = 1 , 2 , \dots | \mathcal { C } _ { s } |$
335
+ 3: for iter < MaxIter do
336
+ 4: $v _ { s } = M ( x _ { s } )$ ; $u _ { s } = F _ { s } ( v _ { s } )$ $; ~ { \hat { v } } _ { s } = G ( u _ { s } ) ; ~ u _ { r } \sim \mathcal { N } ( \mu _ { c _ { i } } , \Sigma _ { c _ { i } } ) { \mathrm { ~ f o r ~ } } i = 1 , 2 , . . . | \mathcal { C } _ { s } | ; ~ { \hat { u } } _ { r } = F _ { s } \circ G ( u _ { r } )$
337
+ 5: (˜ur, y˜r) = sample latent-simulated negative instances from D(b)n
338
+ 6: $\hat { y } _ { s } ^ { ( k _ { s } ) } = \sigma ^ { ( k _ { s } ) } ( D \circ F _ { s } \circ M ( x _ { s } ) )$ , and $\hat { y } _ { n } ^ { ( k _ { n } ) } = \sigma ^ { ( k _ { n } ) } ( D \circ F _ { s } \circ G ( \tilde { u } _ { r } ) )$ where $k _ { s }$ and $k _ { n }$ are the index
339
+ of ground-truth label $y _ { s }$ and $y _ { n }$ respectively, and $\sigma$ is the softmax activation.
340
+ 7: $\mathcal { L } _ { C E } = - \log \hat { y } _ { s } ^ { ( k _ { s } ) } - \alpha \log \hat { y } _ { n } ^ { ( k _ { n } ) }$ g ˆy(kn)n ; Lv = |vs − vˆs|; Lu = |ur − uˆr|
341
+ 8: $\begin{array} { r } { \mathcal { L } _ { p } = - \log ( \exp ( P ( u _ { s } | c _ { k _ { s } } ) ) / \sum _ { i = 1 } ^ { | \mathcal { C } _ { s } | } \exp ( P ( u _ { s } | c _ { i } ) ) ) } \end{array}$ , where $P ( u _ { s } | c _ { i } ) = \mathcal { N } ( u _ { s } | \mu _ { c _ { i } } , \Sigma _ { c _ { i } } )$
342
+ 9: Update $\boldsymbol { \cdot } _ { F _ { s } } , \boldsymbol { \theta } _ { D } , \boldsymbol { \theta } _ { G }$ by minimizing $\mathcal { L } _ { C E } , \mathcal { L } _ { v } , \mathcal { L }$ $\mathcal { L } _ { u }$ , and $\mathcal { L } _ { p }$ alternatively using separate optimizers.
343
+ 10: if $( i t e r \ \% \ U p d a t e I t e r = = 0 )$ then
344
+ 11: Recompute $\mu _ { c _ { i } }$ , $\Sigma _ { c _ { i } }$ for each source class $c _ { i }$ ; Generate $\mathcal { D } _ { n } ^ { ( b ) }$ using the updated priors.
345
+
346
+ d) Minibatch negative sampling strategy. We create an unbiased batch of training samples for a training iteration by sampling equal number of positive and negative samples from the dataset. For USFDA-a we sample 32 positive images $( b _ { + v e } = 3 2 ^ { \cdot }$ ) and 32 negative images per training iteration $( b _ { - v e } = 3 2 )$ ). Similarly, for USFDA- $^ { b }$ we sample 32 positive images and 32 latent-simulated negative samples. This gives an effective batch size of $b _ { + v e } + b _ { - v e } = 6 4$ .
347
+
348
+ e) Use of multiple optimizers for training. In the presence of multiple loss terms, we subvert a time-consuming loss-weighting scheme search by making use of multiple Adam optimizers during training. Essentially, we define a separate optimizer for each loss term, and optimize only one of the losses (chosen in a round robin fashion) in each iteration of training. We use a learning rate of 0.0001 during training. Intuitively, the higher order moment parameters in the Adam optimizer adaptively scale the gradients as required by the loss landscape.
349
+
350
+ f) Label-Set Relationships. For Office-31 dataset in the UDA setting, we use the 10 classes shared by Office-31 and Caltech-256 as the shared label-set $\mathcal { C }$ . These classes are: back_pack, calculator, keyboard, monitor, mouse, mug, bike, laptop_computer, headphones, projector. From the remaining classes, in alphabetical order, we choose the first 10 classes as source-private $( \overline { { \mathcal { C } } } _ { s } )$ classes, and the rest 11 as target-private $( \overline { { \mathcal { C } } } _ { t } )$ classes. For VisDA, alphabetically, the first 6 classes are considered $\mathcal { C }$ , the next 3 as $\overline { { \boldsymbol { \mathcal { C } } } } _ { s }$ and the last 3 comprise $\overline { { \mathcal { C } } } _ { t }$ . The Office-Home dataset has 65 categories, of which we use the first 10 classes as $\mathcal { C }$ , the next 5 for $\overline { { \mathcal { C } } } _ { s }$ , and the rest 50 classes as $\overline { { \boldsymbol { \mathcal { C } } } } _ { t }$ .
351
+
352
+ # B.2 DEPLOYMENT STAGE
353
+
354
+ The details of the architecture used during the Deployment stage are given in Table 7. Note that the Feature Decoder $G$ used during the Procurement stage, is not available during the Deployment stage, restricting complete access to the source data.
355
+
356
+ Training during the Deployment stage. The only trainable component is the Feature Extractor $F _ { t }$ , which is initialized from $F _ { s }$ at Deployment. Here, the SSM is calculated by passing the target images through the network trained on source data (source model), i.e for each image $x _ { t }$ , we calculate $\hat { y } =$ $s o f t m a x ( D \circ F _ { s } \circ M ( x _ { t } ) )$ . Note that the softmax is calculated over all $\left| \mathcal { C } _ { s } \right| + \left| \mathcal { C } _ { n } \right|$ classes. This is done by concatenating the outputs of $D _ { s r c }$ and $D _ { n e g }$ , and then calculating softmax. Then, the SSM is determined by the exponential confidence of a target sample, where confidence is the highest softmax value in the categories in $| { \mathcal { C } } _ { s } |$ .
357
+
358
+ # C ABLATION STUDIES AND ADDITIONAL RESULTS
359
+
360
+ # C.1 PRETRAINING THE BACKBONE NETWORK ON PLACES INSTEAD OF IMAGENET.
361
+
362
+ We find that widely adopted standard domain adaptation datasets such as Office-31 and VisDA often share a part or all of their label-set with ImageNet. Therefore, to validate our method’s applicability when initialized from a network pretrained on an unrelated dataset, we attempt to solve the adaptation task $\mathbf { A } { \to } \mathbf { D }$ in Office-31 dataset by pretraining the ResNet-50 backbone on Places dataset (Zhou et al., 2017). In Table 3 it can be observed that our method outperforms even source-dependent methods (e.g. UAN (You et al., 2019), which is also initialized a ResNet-50 backbone pretrained on Places dataset). In contrast to our method, the algorithm in UAN involves ResNet-50 finetuning. Therefore, we also compare against a variant of UAN with a frozen backbone network, by inserting an additional feature extractor that operates on the features extracted from ResNet-50 (similar to $F _ { s }$ in the proposed method). The architecture of the feature extractor used for this variant of UAN is outlined in Table 6. We observe that our method significantly outperforms this variant of UAN with lesser number of trainable parameters (see Table 3).
363
+
364
+ ![](images/4452e48e11c448173aa7b055707427b7bc9a9285de0b4c31307f6534d73687da.jpg)
365
+
366
+ ![](images/19aed676b96ba390cd62dc8ad56e94707a72a089eb08c025ef1e12ea675fca96.jpg)
367
+ Figure 7: Comparison of target-unknown accuracy $\mathcal { T } _ { u n k }$ across varied label-set relationships for the task $\mathbf { A } { \to } \mathbf { D }$ in Office-31 dataset. A) Visual representation of label-set relationships and $\mathcal { T } _ { u n k }$ at the corresponding instances for B) $\mathrm { U A N ^ { \ast } }$ (You et al., 2019) and C) ours source-free model. Effectively, the direction along $\mathbf { X }$ -axis (blue horizontal arrow) characterizes increasing Open-set complexity. The direction along y-axis (red vertical arrow) shows increasing complexity of Partial DA scenario. And the pink diagonal arrow denotes the effect of decreasing shared label space.
368
+
369
+ # C.2 SPACE AND TIME COMPLEXITY ANALYSIS.
370
+
371
+ On account of keeping the weights of the backbone network frozen throughout the training process, and devoid of ad-hoc networks such as adversarial discriminator our method makes use of significantly lesser trainable parameters when compared to previous methods such as UAN (See Table 3). Devoid of adversarial training, the proposed method also has a significantly lesser total training time for adaptation: 44 sec versus 280 sec in UAN (for the $\mathbf { A } { \to } \mathbf { D }$ task of Office-31 and batch size of 32). Therefore, the proposed framework offers a much simpler adaptation pipeline, with a superior time and space complexity and at the same time achieves state-of-the-art domain adaptation performance across different datasets, even without accessing labeled source data at the time of adaptation (See Table 3). This corroborates the superiority of our method in real-time deployment scenarios.
372
+
373
+ # C.3 VARYING LABEL-SET RELATIONSHIP
374
+
375
+ In addition to the $\mathcal { T } _ { a v g }$ reported in Fig. 6 in the paper, we also compare the target-unknown accuracy $\mathcal { T } _ { u n k }$ for $\mathrm { U A N ^ { * } }$ and our pipeline. The results are presented in Figure 7. Refer the link to the code provided in the submission for details of the chosen class labels for each adaptation scenario shown in Figure 7. Clearly, our method achieves a statistically significant improvement on most of the label-set relationships over UAN. This demonstrates the capability of our algorithm to detect outlier classes more efficiently than UAN, which can be attributed to the ingeniously developed Procurement stage.
376
+
377
+ ![](images/0d55b9b21dda9cf8b21c4f6f04a30fa12fb65e77d81bffc6cbec69e7c294c4d3.jpg)
378
+ Office-31, D→A 1.0 A = 10 = 10 Figure 8: A. Sensitivity against $| { \mathcal { C } } _ { n } |$ 0.9, represented by $| { \mathcal { C } } _ { n } | / | ^ { | { \mathcal { C } } _ { s } | } C _ { 2 }$ for varying $| \overline { { \mathcal { C } } } _ { s } |$ or $| \overline { { \mathcal { C } } } _ { t } |$ (see fig. 0.9 0.8legend) by fixing the others (top cyan box), across varied datasets. B. Sensitivity against $\alpha$ and 0.8 batch-size ratio (fixed $b _ { + v e } + b _ { - v e } = 6 4 )$ .7. Note the scale of X and Y-axis.
379
+
380
+ 0.6 Table 4: Accuracy $( \% )$ 0.5 on unsupervised closed-set DA (all use ResNet50). Ours is w/o hyperparmeter 0.5 tuning. Refer Section C.5.
381
+
382
+ <table><tr><td rowspan="2">Closed-set DA methods</td><td rowspan="2">source- free</td><td rowspan="2">Universal- DA</td><td colspan="7">Office-31</td><td>VisDA</td></tr><tr><td>D→A</td><td>A→D</td><td>A→W</td><td>W→D</td><td>W→A</td><td>D→W</td><td>Avg.</td><td>S→R</td></tr><tr><td>DAN (ICML&#x27;15)</td><td>X</td><td>X</td><td>63.6</td><td>78.6</td><td>80.5</td><td>99.6</td><td>62.8</td><td>97.1</td><td>80.4</td><td>61.1</td></tr><tr><td>ADDA (CVPR&#x27;17)</td><td>X</td><td>X</td><td>69.5</td><td>77.8</td><td>86.2</td><td>98.4</td><td>68.9</td><td>96.2</td><td>82.8</td><td>-</td></tr><tr><td>CDAN (NeurIPS&#x27;18)</td><td>X</td><td>X</td><td>70.1</td><td>89.8</td><td>93.1</td><td>100</td><td>68.0</td><td>98.2</td><td>86.5</td><td>66.8</td></tr><tr><td>UAN (CVPR&#x27;19)</td><td></td><td></td><td>68.4</td><td>85.3</td><td>81.2</td><td>99.1</td><td>69.7</td><td>98.1</td><td>83.6</td><td>1</td></tr><tr><td>Ours USFDA-a (source-free)</td><td>X</td><td>V</td><td>70.4</td><td>85.4</td><td>81.6</td><td>98.0</td><td>69.4</td><td>98.4</td><td>83.9</td><td>59.8</td></tr></table>
383
+
384
+ # C.4 SENSITIVITY ANALYSIS
385
+
386
+ In all our experiments (across datasets as in Tables 1 and 2 and across varied label-set relationships as in Fig. 6), we fix the hyperparameters as, $\alpha = 0 . 2$ , $\beta = 0 . 1$ , $| \mathcal { C } _ { n } | = | \mathcal { C } _ { s } | _ { C _ { 2 } }$ and $b _ { + v e } / b _ { - v e } = 1$ As mentioned in Section 4.3, one can treat these hyperparameters as global constants. In Fig. 8 we demonstrate the sensitivity of the model to these hyperparameters. Specifically, in Fig. 8A we show the sensitivity of the adaptation performance, to the choice of $| { \mathcal { C } } _ { n } |$ during the Procurement stage, across a spectrum of label-set relationships. In Fig. 8B we show the sensitivity of the model to $\alpha$ and the batch-size ratio $b _ { + v e } / b _ { - v e }$ . Sensitivity to $\beta$ is shown in Fig. 5. Clearly, the model achieves a reasonably low sensitivity to the hyperparameters, even in the challenging source-free scenario.
387
+
388
+ # C.5 CLOSED-SET ADAPTATION
389
+
390
+ We additionally evaluate our method in the unsupervised closed set adaptation scenario. In Table 4 we compare with the closed set domain adaptation methods DAN (Long et al., 2015), ADDA (Tzeng et al., 2017), CDAN (Long et al., 2018) and the universal domain adaptation method UAN (You et al., 2019). Note that, DAN, ADDA and CDAN rely on the assumption of a shared label space between the source and the target, and hence are not suited for a universal setting. Furthermore, all other methods require an explicit retraining on the source data during adaptation to perform well, even in the closed-set scenario. This clearly establishes the superiority of our method in the source-free setting.
391
+
392
+ # C.6 ACCURACY ON SOURCE DATASET POST PROCUREMENT
393
+
394
+ We observe in our experiments that the accuracy on the source samples does not drop as a result of the partially generative framework. For the experiments conducted in Fig. 5C, we observe similar classification accuracy on the source validation set, on increasing the number of negative classes from 0 to 190. This effect can be attributed to a carefully chosen $\alpha = 0 . 2$ , which is deliberately biased towards positive source samples to help maintain the discriminative power of the model even in the presence of class imbalance (i.e. $| \mathcal { C } _ { n } | \overset { \_ } { \gg } | \mathcal { C } _ { s } | )$ . This enhances the model’s generative ability without compromising on the discriminative capacity on the positive source samples.
395
+
396
+ ![](images/a52f60b77224e3e7d0732b1f98008daffe6df1c74090c8b77b9989be5c4e4e03.jpg)
397
+ Figure 9: t-SNE plot showing placement of all the four clusters computed after adaptation for the task $\mathbf { A } { \to } \mathbf { D }$ in Office-31. It validates our hypothesis in both Procurement and Deployment stages as shown by the highlighted clusters and the corresponding inferences in the legend under "Category clusters".
398
+
399
+ # C.7 INCREMENTAL ONE-SHOT CLASSIFICATION
400
+
401
+ In universal adaptation, we seek to transfer the knowledge of "class separability" obtained from the source domain to the deployed target environment. More concretely, it is attributed to the segregation of data samples based on an expected characteristics, such as classification of objects according to their pose, color, or shape etc. To quantify this, we consider an extreme case where $\mathcal { C } _ { s } \cap \mathcal { C } _ { t } \overset { = } { = } \varnothing$ $( \mathbf { A } { } \mathbf { D }$ in Office-31 with $| { \mathcal { C } } _ { s } | = 1 5$ , $| \mathcal { C } _ { t } | = 1 6 $ ). Considering access to a single labeled target sample from each target category in $\overline { { \boldsymbol { { \mathcal { C } } } } } _ { t } = \boldsymbol { { \mathcal { C } } } _ { t }$ , which are denoted as $\boldsymbol { x } _ { t } ^ { c _ { j } }$ , where $j = 1 , 2 , . . , | \mathcal { C } _ { t } |$ we perform one-shot Nearest-Neighbour based classification by obtaining the predicted class label as $\begin{array} { r } { \hat { c } _ { t } = \mathrm { a r g m i n } _ { c _ { j } } | | F _ { t } \circ M ( x _ { t } ) - \mathbf { \bar { F } } _ { t } \circ M ( x _ { t } ^ { c _ { j } } ) | | _ { 2 } } \end{array}$ . Then, the classification accuracy for the entire target set is computed by comparing $\hat { c } _ { t }$ with the corresponding ground-truth category. We obtain $6 4 . 7 2 \%$ accuracy for the proposed framework as compared to $1 3 . 4 3 \%$ for $\mathrm { U A N ^ { * } }$ (You et al., 2019). A higher accuracy indicates that, the samples are inherently clustered in the intermediate feature level $M \circ F _ { t } ( x _ { t } )$ validating an efficient transfer of “class separability” in a fully unsupervised manner.
402
+
403
+ # C.8 FEATURE SPACE VISUALIZATION
404
+
405
+ We obtain a t-SNE plot at the intermediate feature level $u$ for both target and source samples (see Figure 9), where the embedding for the target samples is obtained as $u _ { t } = F _ { t } \circ M ( x _ { t } ) { \overline { { } } }$ and the same for the source samples is obtained as $u _ { s } = F _ { s } \circ M ( x _ { s } )$ . This is because we aim to learn domain-specific features in contrast to domain-agnostic features as a result of the restriction imposed by the source-free scenario ("cannot disturb placement of source clusters"). Firstly we obtain compact clusters for the source-categories as a result of the partially generative Procurement stage. Secondly, the target-private clusters are placed away from the source-shared and source-private as expected as a result of the carefully formalized SSM weighting scheme in the Deployment stage. This plot clearly validates our hypothesis.
406
+
407
+ # D MISCELLANEOUS
408
+
409
+ # D.1 SPECIFICATIONS OF COMPUTING RESOURCES
410
+
411
+ For both Procurement and Deployment stages, we make use of the machine with the specifications mentioned in Table 8. The architecture is developed and trained in Python 2.7 with PyTorch 1.0.0.
412
+
413
+ Table 5: Network architecture for Procurement stage. Hyperparameter $\alpha = 0 . 2$
414
+
415
+ <table><tr><td>Component</td><td>Trainable?</td><td>Operation</td><td>Notation</td><td>Features</td><td></td><td>Batch Norm? Non-Linearity</td></tr><tr><td>Resnet-50</td><td>X</td><td></td><td>M</td><td>2048</td><td></td><td></td></tr><tr><td>(Upto AvgPool layer) Feature Extractor</td><td>√</td><td></td><td>Fs</td><td>256</td><td></td><td></td></tr><tr><td></td><td></td><td>Input</td><td></td><td>2048</td><td>xx√</td><td></td></tr><tr><td></td><td></td><td>Fully connected</td><td></td><td>1024</td><td></td><td>ELU</td></tr><tr><td></td><td></td><td>Fully connected</td><td></td><td>1024</td><td></td><td>ELU</td></tr><tr><td></td><td></td><td>Fully connected</td><td></td><td>256</td><td>X</td><td>ELU</td></tr><tr><td></td><td></td><td>Fully connected</td><td></td><td>256</td><td>√</td><td>ELU</td></tr><tr><td>FeatureDecoder</td><td>√</td><td></td><td>G</td><td>2048</td><td></td><td></td></tr><tr><td></td><td></td><td>Input</td><td></td><td>256</td><td>X</td><td></td></tr><tr><td></td><td></td><td>Fully connected</td><td></td><td>1024</td><td>×</td><td>ELU</td></tr><tr><td></td><td></td><td></td><td></td><td>1024</td><td>√</td><td>ELU</td></tr><tr><td></td><td></td><td>Fully connected Fully connected</td><td></td><td>2048</td><td>X</td><td>ELU</td></tr><tr><td></td><td></td><td>Fully connected</td><td></td><td>2048</td><td>X</td><td>-</td></tr><tr><td>Classifier</td><td>√</td><td></td><td>D</td><td>|Cs|+|Cnl</td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td><td>256</td><td></td><td></td></tr><tr><td></td><td></td><td>Input</td><td></td><td></td><td>X ×</td><td></td></tr><tr><td></td><td></td><td>Fully connected</td><td>Dsrc</td><td>|Csl</td><td></td><td></td></tr><tr><td></td><td></td><td>Input Fully connected</td><td>Dneg</td><td>256 |Cn|</td><td>X ×</td><td></td></tr></table>
416
+
417
+ Table 6: Feature Extractor Architecture used for training UAN (You et al., 2019) under the "no ResNet-50 finetuning" case (Refer Table 3 and Section C.1)
418
+
419
+ <table><tr><td>Operation</td><td>Features</td><td>Non-Linearity</td></tr><tr><td>Input</td><td>2048</td><td></td></tr><tr><td>Fully connected</td><td>512</td><td>ReLU</td></tr><tr><td>Fully connected</td><td>256</td><td>ReLU</td></tr><tr><td>Fully connected</td><td>512</td><td>ReLU</td></tr><tr><td>Fully connected</td><td>2048</td><td>ReLU</td></tr></table>
420
+
421
+ Table 7: Network architecture for Deployment stage. Hyperparameter $\beta = 0 . 1$
422
+
423
+ <table><tr><td>Component</td><td>Trainable?</td><td>Operation</td><td>Notation</td><td>Features</td><td></td><td>Batch Norm? Non-Linearity</td></tr><tr><td>Resnet-50 (Upto AvgPool layer)</td><td>X</td><td></td><td>M</td><td>2048</td><td></td><td></td></tr><tr><td>Feature Extractor</td><td>√</td><td></td><td>Ft</td><td>256</td><td></td><td></td></tr><tr><td rowspan="5"></td><td></td><td>Input</td><td></td><td>2048</td><td></td><td></td></tr><tr><td></td><td>Fully connected</td><td></td><td>1024</td><td></td><td>ELU</td></tr><tr><td></td><td>Fully connected</td><td></td><td>1024</td><td>xx√</td><td>ELU</td></tr><tr><td></td><td>Fully connected</td><td></td><td>256</td><td>×</td><td>ELU</td></tr><tr><td></td><td>Fully connected</td><td></td><td>256</td><td>√</td><td>ELU</td></tr><tr><td>Classifier</td><td>X</td><td></td><td>D</td><td>|Cs|+|Cnl</td><td></td><td></td></tr><tr><td></td><td></td><td>Input</td><td></td><td>256</td><td>X</td><td></td></tr><tr><td></td><td></td><td>Fully connected</td><td>Dsrc</td><td>|Cs|</td><td>×</td><td></td></tr><tr><td></td><td></td><td>Input</td><td></td><td>256</td><td>X</td><td></td></tr><tr><td></td><td></td><td>Fully connected</td><td>Dneg</td><td>|Cnl</td><td>X</td><td></td></tr></table>
424
+
425
+ Table 8: Specifications of the machine used for both Procurement and Deployment stages
426
+
427
+ <table><tr><td>CPU</td><td>GPU</td><td>RAM</td><td>VRAM</td><td>CUDA</td></tr><tr><td>Intel i7-7700K</td><td colspan="4">NVIDIA GeForce GTX1080 Ti 32 GB 11GB V8.0.61</td></tr></table>
428
+
429
+ # D.2 REFERENCES TO CODE
430
+
431
+ Proposed Method. Our complete documented code (including data loaders, training pipeline etc.) used for running the experiments is available for reproducibility (refer to the private comment containing the link). Details of dataset splits can be found in Section B.1. For evaluating UAN (You et al., 2019), we execute the official implementation provided by the authors on github1.
432
+
433
+ Negative dataset creation. We have provided the complete dataset with augmentations and negative images for the task $\mathbf { A } { \xrightarrow { } } \mathbf { D }$ in Office-31 (Saenko et al., 2010), along with the negative dataset creation tool (refer to the code link).
parse/train/B1gd0nEFwS/B1gd0nEFwS_content_list.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/B1gd0nEFwS/B1gd0nEFwS_middle.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/B1gd0nEFwS/B1gd0nEFwS_model.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/BJYwwY9ll/BJYwwY9ll.md ADDED
@@ -0,0 +1,263 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # SNAPSHOT ENSEMBLES: TRAIN 1, GET M FOR FREE
2
+
3
+ Gao Huang∗, Yixuan Li∗, Geoff Pleiss
4
+ Cornell University
5
+ {gh349, yl2363}@cornell.edu, geoff@cs.cornell.edu
6
+ Zhuang Liu
7
+ Tsinghua University
8
+ liuzhuangthu@gmail.com
9
+
10
+ John E. Hopcroft, Kilian Q. Weinberger Cornell University jeh@cs.cornell.edu, kqw4@cornell.edu
11
+
12
+ # ABSTRACT
13
+
14
+ Ensembles of neural networks are known to be much more robust and accurate than individual networks. However, training multiple deep networks for model averaging is computationally expensive. In this paper, we propose a method to obtain the seemingly contradictory goal of ensembling multiple neural networks at no additional training cost. We achieve this goal by training a single neural network, converging to several local minima along its optimization path and saving the model parameters. To obtain repeated rapid convergence, we leverage recent work on cyclic learning rate schedules. The resulting technique, which we refer to as Snapshot Ensembling, is simple, yet surprisingly effective. We show in a series of experiments that our approach is compatible with diverse network architectures and learning tasks. It consistently yields lower error rates than state-of-the-art single models at no additional training cost, and compares favorably with traditional network ensembles. On CIFAR-10 and CIFAR-100 our DenseNet Snapshot Ensembles obtain error rates of $3 . 4 \%$ and $1 7 . 4 \%$ respectively.
15
+
16
+ # 1 INTRODUCTION
17
+
18
+ Stochastic Gradient Descent (SGD) (Bottou, 2010) and its accelerated variants (Kingma & Ba, 2014; Duchi et al., 2011) have become the de-facto approaches for optimizing deep neural networks. The popularity of SGD can be attributed to its ability to avoid and even escape spurious saddle-points and local minima (Dauphin et al., 2014). Although avoiding these spurious solutions is generally considered positive, in this paper we argue that these local minima contain useful information that may in fact improve model performance.
19
+
20
+ Although deep networks typically never converge to a global minimum, there is a notion of “good” and “bad” local minima with respect to generalization. Keskar et al. (2016) argue that local minima with flat basins tend to generalize better. SGD tends to avoid sharper local minima because gradients are computed from small mini-batches and are therefore inexact (Keskar et al., 2016). If the learningrate is sufficiently large, the intrinsic random motion across gradient steps prevents the optimizer from reaching any of the sharp basins along its optimization path. However, if the learning rate is small, the model tends to converge into the closest local minimum. These two very different behaviors of SGD are typically exploited in different phases of optimization (He et al., 2016a). Initially the learning rate is kept high to move into the general vicinity of a flat local minimum. Once this search has reached a stage in which no further progress is made, the learning rate is dropped (once or twice), triggering a descent, and ultimately convergence, to the final local minimum.
21
+
22
+ It is well established (Kawaguchi, 2016) that the number of possible local minima grows exponentially with the number of parameters—of which modern neural networks can have millions. It is therefore not surprising that two identical architectures optimized with different initializations or minibatch orderings will converge to different solutions. Although different local minima often have very similar error rates, the corresponding neural networks tend to make different mistakes. This diversity can be exploited through ensembling, in which multiple neural networks are trained from different initializations and then combined with majority voting or averaging (Caruana et al., 2004). Ensembling often leads to drastic reductions in error rates. In fact, most high profile competitions, e.g. Imagenet (Deng et al., 2009) or Kaggle1, are won by ensembles of deep learning architectures.
23
+
24
+ ![](images/1f88b86ec2fd97a2263949838ab06c8e2bf496e67dc1f5e93984e1f4e23ef598.jpg)
25
+ Figure 1: Left: Illustration of SGD optimization with a typical learning rate schedule. The model converges to a minimum at the end of training. Right: Illustration of Snapshot Ensembling. The model undergoes several learning rate annealing cycles, converging to and escaping from multiple local minima. We take a snapshot at each minimum for test-time ensembling.
26
+
27
+ Despite its obvious advantages, the use of ensembling for deep networks is not nearly as widespread as it is for other algorithms. One likely reason for this lack of adaptation may be the cost of learning multiple neural networks. Training deep networks can last for weeks, even on high performance hardware with GPU acceleration. As the training cost for ensembles increases linearly, ensembles can quickly becomes uneconomical for most researchers without access to industrial scale computational resources.
28
+
29
+ In this paper we focus on the seemingly-contradictory goal of learning an ensemble of multiple neural networks without incurring any additional training costs. We achieve this goal with a training method that is simple and straight-forward to implement. Our approach leverages the non-convex nature of neural networks and the ability of SGD to converge to and escape from local minima on demand. Instead of training $M$ neural networks independently from scratch, we let SGD converge $M$ times to local minima along its optimization path. Each time the model converges, we save the weights and add the corresponding network to our ensemble. We then restart the optimization with a large learning rate to escape the current local minimum. More specifically, we adopt the cycling procedure suggested by Loshchilov & Hutter (2016), in which the learning rate is abruptly raised and then quickly lowered to follow a cosine function. Because our final ensemble consists of snapshots of the optimization path, we refer to our approach as Snapshot Ensembling. Figure 1 presents a high-level overview of this method.
30
+
31
+ In contrast to traditional ensembles, the training time for the entire ensemble is identical to the time required to train a single traditional model. During testing time, one can evaluate and average the last (and therefore most accurate) $m$ out of $M$ models. Our approach is naturally compatible with other methods to improve the accuracy, such as data augmentation, stochastic depth (Huang et al., 2016b), or batch normalization (Ioffe & Szegedy, 2015). In fact, Snapshot Ensembles can even be ensembled, if for example parallel resources are available during training. In this case, an ensemble of $K$ Snapshot Ensembles yields $K \times M$ models at $K$ times the training cost.
32
+
33
+ We evaluate the efficacy of Snapshot Ensembles on three state-of-the-art deep learning architectures for object recognition: ResNet (He et al., 2016b), Wide-ResNet (Zagoruyko & Komodakis, 2016), and DenseNet (Huang et al., 2016a). We show across four different data sets that Snapshot Ensembles almost always reduce error without increasing training costs. For example, on CIFAR-10 and CIFAR-100, Snapshot Ensembles obtains error rates of $3 . 4 \bar { 4 } \%$ and $1 7 . 4 1 \%$ respectively.
34
+
35
+ # 2 RELATED WORK
36
+
37
+ Neural network ensembles have been widely studied and applied in machine learning (Hansen & Salamon, 1990; Krogh et al., 1995). However, most of these prior studies focus on improving the generalization performance, while few of them address the cost of training ensembles.
38
+
39
+ As an alternative to traditional ensembles, so-called “implicit” ensembles have high efficiency during both training and testing (Srivastava et al., 2014; Wan et al., 2013; Huang et al., 2016b; Singh et al., 2016; Krueger et al., 2016). The Dropout (Srivastava et al., 2014) technique creates an ensemble out of a single model by “dropping” — or zeroing — random sets of hidden nodes during each mini-batch. At test time, no nodes are dropped, and each node is scaled by the probability of surviving during training. Srivastava et al. claim that Dropout reduces overfitting by preventing the co-adaptation of nodes. An alternative explanation is that this mechanism creates an exponential number of networks with shared weights during training, which are then implicitly ensembled at test time. DropConnect (Wan et al., 2013) uses a similar trick to create ensembles at test time by dropping connections (weights) during training instead of nodes. The recently proposed Stochastic Depth technique (Huang et al., 2016b) randomly drops layers during training to create an implicit ensemble of networks with varying depth at test time. Finally, Swapout (Singh et al., 2016) is a stochastic training method that generalizes Dropout and Stochastic Depth. From the perspective of model ensembling, Swapout creates diversified network structures for model averaging. Our proposed method similarly trains only a single model; however, the resulting ensemble is “explicit” in that the models do not share weights. Furthermore, our method can be used in conjunction with any of these implicit ensembling techniques.
40
+
41
+ Several recent publications focus on reducing the test time cost of ensembles, by transferring the “knowledge” of cumbersome ensembles into a single model (Bucilu et al., 2006; Hinton et al., 2015). Hinton et al. (2015) propose to use an ensemble of multiple networks as the target of a single (smaller) network. Our proposed method is complementary to these works as we aim to reduce the training cost of ensembles rather than the test-time cost.
42
+
43
+ Perhaps most similar to our work is that of Swann & Allinson (1998) and Xie et al. (2013), who explore creating ensembles from slices of the learning trajectory. Xie et al. introduce the horizontal and vertical ensembling method, which combines the output of networks within a range of training epochs. More recently, Jean et al. (2014) and Sennrich et al. (2016) show improvement by ensembling the intermediate stages of model training. Laine & Aila (2016) propose a temporal ensembling method for semi-supervised learning, which achieves consensus among models trained with different regularization and augmentation conditions for better generalization performance. Finally, Moghimi et al. (2016) show that boosting can be applied to convolutional neural networks to create strong ensembles. Our work differs from these prior works in that we force the model to visit multiple local minima, and we take snapshots only when the model reaches a minimum. We believe this key insight allows us to leverage more power from our ensembles.
44
+
45
+ Our work is inspired by the recent findings of Loshchilov & Hutter (2016) and Smith (2016), who show that cyclic learning rates can be effective for training convolutional neural networks. The authors show that each cycle produces models which are (almost) competitive to those learned with traditional learning rate schedules while requiring a fraction of training iterations. Although model performance temporarily suffers when the learning rate cycle is restarted, the performance eventually surpasses the previous cycle after annealing the learning rate. The authors suggest that cycling perturbs the parameters of a converged model, which allows the model to find a better local minimum. We build upon these recent findings by (1) showing that there is significant diversity in the local minima visited during each cycle and (2) exploiting this diversity using ensembles. We are not concerned with speeding up or improving the training of a single model; rather, our goal is to extract an ensemble of classifiers while following the optimization path of the final model.
46
+
47
+ # 3 SNAPSHOT ENSEMBLING
48
+
49
+ Snapshot Ensembling produces an ensemble of accurate and diverse models from a single training process. At the heart of Snapshot Ensembling is an optimization process which visits several local minima before converging to a final solution. We take model snapshots at these various minima, and average their predictions at test time.
50
+
51
+ Ensembles work best if the individual models (1) have low test error and (2) do not overlap in the set of examples they misclassify. Along most of the optimization path, the weight assignments of a neural network tend not to correspond to low test error. In fact, it is commonly observed that the validation error drops significantly only after the learning rate has been reduced, which is typically done after several hundred epochs. Our approach is inspired by the observation that training neural networks for fewer epochs and dropping the learning rate earlier has minor impact on the final test error (Loshchilov & Hutter, 2016). This seems to suggest that local minima along the optimization path become promising (in terms of generalization error) after only a few epochs.
52
+
53
+ Cyclic Cosine Annealing. To converge to multiple local minima, we follow a cyclic annealing schedule as proposed by Loshchilov & Hutter (2016). We lower the learning rate at a very fast pace, encouraging the model to converge towards its first local minimum after as few as 50 epochs. The optimization is then continued at a larger learning rate, which perturbs the model and dislodges it from the minimum. We repeat this process several times to obtain multiple convergences. Formally, the learning rate $\alpha$ has the form:
54
+
55
+ $$
56
+ \alpha ( t ) = f \left( \mathrm { m o d } \left( t - 1 , \lceil T / M \rceil \right) \right) ,
57
+ $$
58
+
59
+ where $t$ is the iteration number, $T$ is the total number of training iterations, and $f$ is a monotonically decreasing function. In other words, we split the training process into $M$ cycles, each of which starts with a large learning rate, which is annealed to a smaller learning rate. The large learning rate $\alpha = f ( 0 )$ provides the model enough energy to escape from a critical point, while the small learning rate
60
+
61
+ $\alpha = f ( \lceil \bar { T } / M \rceil )$ drives the model to a well behaved local minimum. In our experiments, we set $f$ to be the shifted cosine function proposed by Loshchilov & Hutter (2016):
62
+
63
+ ![](images/c2ae06eb52ffabda7683124a8942902ce5bda2833139f06bb9d05f36fe2f9af3.jpg)
64
+ Figure 2: Training loss of 100-layer DenseNet on CIFAR10 using standard learning rate (blue) and $M = 6$ cosine annealing cycles (red). The intermediate models, denoted by the dotted lines, form an ensemble at the end of training.
65
+
66
+ $$
67
+ \alpha ( t ) = \frac { \alpha _ { 0 } } { 2 } \left( \cos \left( \frac { \pi \mathrm { m o d } ( t - 1 , \left\lceil T / M \right\rceil ) } { \left\lceil T / M \right\rceil } \right) + 1 \right) ,
68
+ $$
69
+
70
+ where $\alpha _ { 0 }$ is the initial learning rate. Intuitively, this function anneals the learning rate from its initial value $\alpha _ { 0 }$ to $f ( \lceil T / M \rceil ) \approx 0$ over the course of a cycle. Following (Loshchilov & Hutter, 2016), we update the learning rate at each iteration rather than at every epoch. This improves the convergence of short cycles, even when a large initial learning rate is used.
71
+
72
+ Snapshot Ensembling. Figure 2 depicts the training process using cyclic and traditional learning rate schedules. At the end of each training cycle, it is apparent that the model reaches a local minimum with respect to the training loss. Thus, before raising the learning rate, we take a “snapshot” of the model weights (indicated as vertical dashed black lines). After training $M$ cycles, we have $M$ model snapshots, $f _ { 1 } \ldots f _ { M }$ , each of which will be used in the final ensemble. It is important to highlight that the total training time of the $M$ snapshots is the same as training a model with a standard schedule (indicated in blue). In some cases, the standard learning rate schedule achieves lower training loss than the cyclic schedule; however, as we will show in the next section, the benefits of ensembling outweigh this difference.
73
+
74
+ Ensembling at Test Time. The ensemble prediction at test time is the average of the last $m$ $( m ~ \leq ~ M )$ model’s softmax outputs. Let $\mathbf { x }$ be a test sample and let $h _ { i } \left( \mathbf { x } \right)$ be the softmax $i$ t of the ensemble is a simple. We always ensemble the last verage of the last models, as these $m$ models:dels tend $h _ { \mathrm { E n s e m b l e } } = \textstyle { \frac { 1 } { m } } \sum _ { 0 } ^ { m - 1 } h _ { M - i } ( { \bf x } )$ $m$
75
+
76
+ Table 1: Error rates $( \% )$ on CIFAR-10 and CIFAR-100 datasets. All methods in the same group are trained for the same number of iterations. Results of our method are colored in blue, and the best result for each network/dataset pair are bolded. ∗ indicates numbers which we take directly from Huang et al. (2016a).
77
+
78
+ <table><tr><td colspan="2">Method</td><td>C10</td><td>C100</td><td>SVHN</td><td>Tiny ImageNet</td></tr><tr><td rowspan="5">ResNet-110</td><td>Single model</td><td>5.52</td><td>28.02</td><td>1.96</td><td>46.50</td></tr><tr><td>NoCycle Snapshot Ensemble</td><td>5.49</td><td>26.97</td><td>1.78</td><td>43.69</td></tr><tr><td>SingleCycle Ensembles</td><td>6.66</td><td>24.54</td><td>1.74</td><td>42.60</td></tr><tr><td>Snapshot Ensemble (αo = 0.1)</td><td>5.73</td><td>25.55</td><td>1.63</td><td>40.54</td></tr><tr><td>Snapshot Ensemble (αo = 0.2)</td><td>5.32</td><td>24.19</td><td>1.66</td><td>39.40</td></tr><tr><td rowspan="6">Wide-ResNet-32</td><td>Single model</td><td>5.43</td><td>23.55</td><td>1.90</td><td>39.63</td></tr><tr><td>Dropout</td><td>4.68</td><td>22.82</td><td>1.81</td><td>36.58</td></tr><tr><td>NoCycle Snapshot Ensemble</td><td>5.18</td><td>22.81</td><td>1.81</td><td>38.64</td></tr><tr><td>SingleCycle Ensembles</td><td>5.95</td><td>21.38</td><td>1.65</td><td>35.53</td></tr><tr><td>Snapshot Ensemble (αo = 0.1)</td><td>4.41</td><td>21.26</td><td>1.64</td><td>35.45</td></tr><tr><td>Snapshot Ensemble(αo = 0.2)</td><td>4.73</td><td>21.56</td><td>1.51</td><td>32.90</td></tr><tr><td rowspan="6">DenseNet-40</td><td>Single model</td><td>5.24*</td><td>24.42*</td><td>1.77</td><td>39.09</td></tr><tr><td>Dropout</td><td>6.08</td><td>25.79</td><td>1.79*</td><td>39.68</td></tr><tr><td>NoCycle Snapshot Ensemble</td><td>5.20</td><td>24.63</td><td>1.80</td><td>38.51</td></tr><tr><td>SingleCycle Ensembles</td><td>5.43</td><td>22.51</td><td>1.87</td><td>38.00</td></tr><tr><td>Snapshot Ensemble (αo = 0.1)</td><td>4.99</td><td>23.34</td><td>1.64</td><td>37.25</td></tr><tr><td>Snapshot Ensemble (αo = 0.2)</td><td>4.84</td><td>21.93</td><td>1.73</td><td>36.61</td></tr><tr><td rowspan="6">DenseNet-100</td><td>Single model</td><td>3.74*</td><td>19.25*</td><td></td><td>=</td></tr><tr><td>Dropout</td><td>3.65</td><td>18.77</td><td>=</td><td></td></tr><tr><td>NoCycle Snapshot Ensemble</td><td>3.80</td><td>19.30</td><td></td><td></td></tr><tr><td>SingleCycle Ensembles</td><td>4.52</td><td>18.38</td><td></td><td></td></tr><tr><td>Snapshot Ensemble (αo = 0.1)</td><td>3.57</td><td>18.12</td><td></td><td></td></tr><tr><td>Snapshot Ensemble (αo = 0.2)</td><td>3.44</td><td>17.41</td><td></td><td></td></tr></table>
79
+
80
+ # 4 EXPERIMENTS
81
+
82
+ We demonstrate the effectiveness of Snapshot Ensembles on several benchmark datasets, comparing with competitive baselines. We run all experiments with Torch 7 (Collobert et al., 2011)2.
83
+
84
+ # 4.1 DATASETS
85
+
86
+ CIFAR. The two CIFAR datasets (Krizhevsky & Hinton, 2009) consist of colored natural images sized at $3 2 \times 3 2$ pixels. CIFAR-10 (C10) and CIFAR-100 (C100) images are drawn from 10 and 100 classes, respectively. For each dataset, there are 50,000 training images and 10,000 images reserved for testing. We use a standard data augmentation scheme (Lin et al., 2013; Romero et al., 2014; Lee et al., 2015; Springenberg et al., 2014; Srivastava et al., 2015; Huang et al., 2016b; Larsson et al., 2016), in which the images are zero-padded with 4 pixels on each side, randomly cropped to produce $3 2 \times 3 2$ images, and horizontally mirrored with probability 0.5.
87
+
88
+ SVHN. The Street View House Numbers (SVHN) dataset (Netzer et al., 2011) contains $3 2 \times 3 2$ colored digit images from Google Street View, with one class for each digit. There are 73,257 images in the training set and 26,032 images in the test set. Following common practice (Sermanet et al., 2012; Goodfellow et al., 2013; Huang et al., 2016a), we withhold 6,000 training images for validation, and train on the remaining images without data augmentation.
89
+
90
+ Tiny ImageNet. The Tiny ImageNet dataset3 consists of a subset of ImageNet images (Deng et al., 2009). There are 200 classes, each of which has 500 training images and 50 validation images. Each image is resized to $6 4 \times 6 4$ and augmented with random crops, horizontal mirroring, and RGB intensity scaling (Krizhevsky et al., 2012).
91
+
92
+ ImageNet. The ILSVRC 2012 classification dataset (Deng et al., 2009) consists of 1000 images classes, with a total of 1.2 million training images and 50,000 validation images. We adopt the same data augmentation scheme as in (He et al., 2016a; Huang et al., 2016a) and apply a $2 2 4 \times 2 2 4$ center crop to images at test time.
93
+
94
+ ![](images/df25f7f707d94414d7c1652a477801315c74e97f6c0090e95d159019b040c02a.jpg)
95
+ Figure 3: DenseNet-100 Snapshot Ensemble performance on CIFAR-10 and CIFAR-100 with restart learning rate $\alpha _ { 0 } = 0 . 1$ (left two) and $\alpha _ { 0 } = 0 . 2$ (right two). Each ensemble is trained with $M = 6$ annealing cycles (50 epochs per each).
96
+
97
+ # 4.2 TRAINING SETTING
98
+
99
+ Architectures. We test several state-of-the-art architectures, including residual networks (ResNet) (He et al., 2016a), Wide ResNet (Zagoruyko & Komodakis, 2016) and DenseNet (Huang et al., 2016a). For ResNet, we use the original 110-layer network introduced by He et al. (2016a). Wide-ResNet is a 32-layer ResNet with 4 times as many convolutional features per layer as a standard ResNet. For DenseNet, our large model follows the same setup as (Huang et al., 2016a), with depth $L = 1 0 0$ and growth rate $k = 2 4$ . In addition, we also evaluate our method on a small DenseNet, with depth $L = 4 0$ and $k = 1 2$ . To adapt all these networks to Tiny ImageNet, we add a stride of 2 to the first layer of the models, which downsamples the images to $3 2 \times 3 2$ . For ImageNet, we test the 50-layer ResNet proposed in (He et al., 2016a). We use a mini batch size of 64.4
100
+
101
+ Baselines. Snapshot Ensembles incur the training cost of a single model; therefore, we compare with baselines that require the same amount of training. First, we compare against a Single Model trained with a standard learning rate schedule, dropping the learning rate from 0.1 to 0.01 halfway through training, and then to 0.001 when training is at $7 5 \%$ . Additionally, to compare against implicit ensembling methods, we test against a single model trained with Dropout. This baseline uses the same learning rate as above, and drops nodes during training with a probability of 0.2.
102
+
103
+ We then test the Snapshot Ensemble algorithm trained with the cyclic cosine learning rate as described in (2). We test models with the max learning rate $\alpha _ { 0 }$ set to 0.1 and 0.2. In both cases, we divide the training process into learning rate cycles. Model snapshots are taken after each learning rate cycle. Additionally, we train a Snapshot Ensemble with a non-cyclic learning rate schedule. This NoCycle Snapshot Ensemble, which uses the same schedule as the Single Model and Dropout baselines, is meant to highlight the impact of cyclic learning rates for our method. To accurately compare with the cyclic Snapshot Ensembles, we take the same number of snapshots equally spaced throughout the training process. Finally, we compare against SingleCycle Ensembles, a Snapshot Ensemble variant in which the network is re-initialized at the beginning of every cosine learning rate cycle, rather than using the parameters from the previous optimization cycle. This baseline essentially creates a traditional ensemble, yet each network only has $1 / M$ of the typical training time. This variant is meant to highlight the tradeoff between model diversity and model convergence. Though SingleCycle Ensembles should in theory explore more of the parameter space, the models do not benefit from the optimization of previous cycles.
104
+
105
+ Training Budget. On CIFAR datasets, the training budget is $B = 3 0 0$ epochs for DenseNet-40 and DenseNet-100, and $B = 2 0 0$ for ResNet and Wide ResNet models. Snapshot variants are trained with $M = 6$ cycles of $B / M = 5 0$ epochs for DenseNets, and $M = 5$ cycles of $B / M = 4 0$ epochs for ResNets/Wide ResNets. SVHN models are trained with a budget of $B = 4 0$ epochs (5 cycles of 8 epochs). For Tiny ImageNet, we use a training budget of $B = 1 5 0$ (6 cycles of 25 epochs). Finally, ImageNet is trained with a budget of $B = 9 0$ epochs, and we trained 2 Snapshot variants: one with $M = 2$ cycles and one with $M = 3$ .
106
+
107
+ # 4.3 SNAPSHOT ENSEMBLE RESULTS
108
+
109
+ Accuracy. The main results are summarized in Table 1. In most cases, Snapshot ensembles achieve lower error than any of the baseline methods. Most notably, Snapshot Ensembles yield an error rate of $1 7 . 4 1 \%$ on CIFAR100 using large DenseNets, far outperforming the record of ${ \mathrm { \bar { 1 9 . 2 5 \% } } }$ under the same training cost and architecture (Huang et al., 2016a). Our method has the most success on CIFAR-100 and Tiny ImageNet, which is likely due to the
110
+
111
+ Table 2: Top-1 error rates $( \% )$ on ImageNet validation set using ResNet-50 with varying number of cycles.
112
+
113
+ <table><tr><td>Method</td><td>Val. Error (%)</td></tr><tr><td>Single model</td><td>24.01</td></tr><tr><td>Snapshot Ensemble (M= 2)</td><td>23.33</td></tr><tr><td>Snapshot Ensemble (M= 3)</td><td>23.96</td></tr></table>
114
+
115
+ complexity of these datasets. The softmax outputs for these datasets are high dimensional due to the large number of classes, making it unlikely that any two models make the same predictions. Snapshot Ensembling is also capable of improving the competitive baselines for CIFAR-10 and SVHN as well, reducing error by $1 \%$ and $0 . 4 \%$ respectively with the Wide ResNet architecture.
116
+
117
+ The NoCycle Snapshot Ensemble generally has little effect on performance, and in some instances even increases the test error. This highlights the need for a cyclic learning rate for useful ensembling. The SingleCycle Ensemble has similarly mixed performance. In some cases, e.g., DenseNet-40 on CIFAR-100, the SingleCycle Ensemble is competitive with Snapshot Ensembles. However, as the model size increases to 100 layers, it does not perform as well. This is because it is difficult to train a large model from scratch in only a few epochs. These results demonstrate that Snapshot Ensembles tend to work best when utilizing information from previous cycles. Effectively, Snapshot Ensembles strike a balance between model diversity and optimization.
118
+
119
+ Table 2 shows Snapshot Ensemble results on ImageNet. The Snapshot Ensemble with $M \ = \ 2$ achieves $2 3 . 3 3 \%$ validation error, outperforming the single model baseline with $2 4 . 0 1 \%$ validation error. It appears that 2 cycles is the optimal choice for the ImageNet dataset. Provided with the limited total training budget $B = 9 0$ epochs, we hypothesize that allocating fewer than $B / 2 = 4 5$ epochs per training cycle is insufficient for the model to converge on such a large dataset.
120
+
121
+ Ensemble Size. In some applications, it may be beneficial to vary the size of the ensemble dynamically at test time depending on available resources. Figure 3 displays the performance of DenseNet-40 on the CIFAR-100 dataset as the effective ensemble size, $m$ , is varied. Each ensemble consists of snapshots from later cycles, as these snapshots have received the most training and therefore have likely converged to better minima. Although ensembling more models generally gives better performance, we observe significant drops in error when the second and third models are added to the ensemble. In most cases, an ensemble of two models outperforms the baseline model.
122
+
123
+ Restart Learning Rate. The effect of the restart learning rate can be observed in Figure 3. The left two plots show performance when using a restart learning rate of $\alpha _ { 0 } = 0 . 1$ at the beginning of each cycle, and the right two plots show $\alpha _ { 0 } = 0 . 2$ . In most cases, ensembles with the larger restart learning rate perform better, presumably because the strong perturbation in between cycles increases the diversity of local minima.
124
+
125
+ Table 3: Error rates of a DenseNet-40 Snapshot Ensemble on CIFAR-100, varying $M .$ —the number of models (cycles) used in the ensemble.
126
+
127
+ <table><tr><td colspan="2">M Test Error (%)</td></tr><tr><td>2</td><td>22.92</td></tr><tr><td>4</td><td>22.07</td></tr><tr><td>6</td><td>21.93</td></tr><tr><td>8</td><td>21.89</td></tr><tr><td>10</td><td>22.16</td></tr></table>
128
+
129
+ Varying Number of Cycles. Given a fixed training budget, there is a trade-off between the number of learning rate cycles and their length. Therefore, we investigate how the number of cycles $M$ affects the ensemble performance, given a fixed training budget. We train a 40-layer DenseNet on the CIFAR-100 dataset with an initial learning rate of $\alpha _ { 0 } = 0 . 2$ . We fix the total training budget $B = 3 0 0$ epochs, and vary the value of $M \in \{ 2 , 4 , 6 , 8 , 1 0 \}$ . As shown in Table 3, our method is relatively robust with respect to different values of $M$ . At the extremes, $M = 2$ and $M = 1 0$ , we find a slight degradation in performance, as the cycles are either too few or too short. In practice, we find that setting $M$ to be $4 \sim 8$ works reasonably well.
130
+
131
+ Varying Training Budget. The left and middle panels of Figure 4 show the performance of Snapshot Ensembles and SingleCycle Ensembles as a function of training budget (where the number of cycles is fixed at $M = 6$ ). We train a 40-layer DenseNet on CIFAR-10 and CIFAR-100, with an initial learning rate of $\alpha _ { 0 } = 0 . 1$ , varying the total number of training epochs from 60 to 300. We observe that both Snapshot Ensembles and SingleCycle Ensembles become more accurate as training budget increases. However, we note that as training budget decreases, Snapshot Ensembles still yield competitive results, while the performance of the SingleCycle Ensembles degrades rapidly. These results highlight the improvements that Snapshot Ensembles obtain when the budget is low. If the budget is high, then the SingleCycle baseline approaches true ensembles and outperforms Snapshot ensembles eventually.
132
+
133
+ ![](images/566524b31fa444cf21bb9f53d27b9fd61198bb62d5bc3d79c805e92b61609d81.jpg)
134
+
135
+ ![](images/ace68d823d81d79b87bf51385623b321d3d01faffbff255db025ad20176534f5.jpg)
136
+ Figure 4: Snapshot Ensembles under different training budgets on (Left) CIFAR-10 and (Middle) CIFAR-100. Right: Comparison of Snapshot Ensembles with true ensembles.
137
+ Figure 5: Interpolations in parameter space between the final model (sixth snapshot) and all intermediate snapshots. $\lambda = 0$ represents an intermediate snapshot model, while $\lambda = 1$ represents the final model. Left: A Snapshot Ensemble, with cosine annealing cycles $\mathrm { \Delta } \alpha _ { 0 } = 0 . 2$ every $B / M \stackrel { \textstyle - } { = } 5 0$ epochs). Right: A NoCycle Snapshot Ensemble, (two learning rate drops, snapshots every 50 epochs).
138
+
139
+ Comparison with True Ensembles. We compare Snapshot Ensembles with the traditional ensembling method. The right panel of Figure 4 shows the test error rates of DenseNet-40 on CIFAR-100. The true ensemble method averages models that are trained with 300 full epochs, each with different weight initializations. Given the same number of models at test time, the error rate of the true ensemble can be seen as a lower bound of our method. Our method achieves performance that is comparable with ensembling of 2 independent models, but with the training cost of one model.
140
+
141
+ # 4.4 DIVERSITY OF MODEL ENSEMBLES
142
+
143
+ Parameter Space. We hypothesize that the cyclic learning rate schedule creates snapshots which are not only accurate but also diverse with respect to model predictions. We qualitatively measure this diversity by visualizing the local minima that models converge to. To do so, we linearly interpolate snapshot models, as described by Goodfellow et al. (2014). Let $J \left( \theta \right)$ be the test error of a model using parameters $\theta$ . Given $\theta _ { 1 }$ and $\theta _ { 2 }$ — the parameters from models 1 and 2 respectively — we can compute the loss for a convex combination of model parameters: $J \left( \lambda \left( \theta _ { 1 } \right) + \left( 1 - \lambda \right) ( \bar { \theta _ { 2 } } ) \right)$ , where $\lambda$ is a mixing coefficient. Setting $\lambda$ to 1 results in a parameters that are entirely $\theta _ { 1 }$ while setting $\lambda$ to 0 gives the parameters $\theta _ { 2 }$ . By sweeping the values of $\lambda$ , we can examine a linear slice of the parameter space. Two models that converge to a similar minimum will have smooth parameter interpolations, whereas models that converge to different minima will likely have a non-convex interpolation, with a spike in error when $\lambda$ is between 0 and 1.
144
+
145
+ Figure 5 displays interpolations between the final model of DenseNet-40 (sixth snapshot) and all intermediate snapshots. The left two plots show Snapshot Ensemble models trained with a cyclic learning rate, while the right two plots show NoCycle Snapshot models. $\lambda = 0$ represents a model which is entirely snapshot parameters, while $\lambda = 1$ represents a model which is entirely the parameters of the final model. From this figure, it is clear that there are differences between cyclic and non-cyclic learning rate schedules. Firstly, all of the cyclic snapshots achieve roughly the same error as the final cyclical model, as the error is similar for $\lambda = 0$ and $\lambda = 1$ . Additionally, it appears that most snapshots do not lie in the same minimum as the final model. Thus the snapshots are likely to misclassify different samples. Conversely, the first three snapshots achieve much higher error than the final model. This can be observed by the sharp minima around $\lambda = 1$ , which suggests that mixing in any amount of the snapshot parameters will worsen performance. While the final two snapshots achieve low error, the figures suggests that they lie in the same minimum as the final model, and therefore likely add limited diversity to the ensemble.
146
+
147
+ Activation space. To further explore the diversity of models, we compute the pairwise correlation of softmax outputs for every pair of snapshots. Figure 6 displays the average correlation for both cyclic snapshots and non-cyclical snapshots. Firstly, there are large correlations between the last 3 snapshots of the non-cyclic training schedule (right). These snapshots are taken after dropping the learning rate, suggesting that each snapshot has converged to the same minimum. Though there is more diversity amongst the earlier snapshots, these snapshots have much higher error rates and are therefore not ideal for ensembling. Conversely, there is less correlation between all cyclic snapshots (left). Because all snapshots have similar accuracy (as can be seen in Figure 5), these differences in predictions can be exploited to create effective ensembles.
148
+
149
+ ![](images/5bfed1b78edee266e038c3f02397a3d98cee181dce976e7a3786fe57b7db3a99.jpg)
150
+ Figure 6: Pairwise correlation of softmax outputs between any two snapshots for DenseNet-100. Left: A Snapshot Ensemble, with cosine annealing cycles (restart with $\alpha _ { 0 } ~ = ~ 0 . 2$ every 50 epochs). Right: A NoCycle Snapshot Ensemble, (two learning rate drops, snapshots every 50 epochs).
151
+
152
+ # 5 DISCUSSION
153
+
154
+ We introduce Snapshot Ensembling, a simple method to obtain ensembles of neural networks without any additional training cost. Our method exploits the ability of SGD to converge to and escape from local minima as the learning rate is lowered, which allows the model to visit several weight assignments that lead to increasingly accurate predictions over the course of training. We harness this power with the cyclical learning rate schedule proposed by Loshchilov & Hutter (2016), saving model snapshots at each point of convergence. We show in several experiments that all snapshots are accurate, yet produce different predictions from one another, and therefore are well suited for test-time ensembles. Ensembles of these snapshots significantly improve the state-of-the-art on CIFAR-10, CIFAR-100 and SVHN. Future work will explore combining Snapshot Ensembles with traditional ensembles. In particular, we will investigate how to balance growing an ensemble with new models (with random initializations) and refining existing models with further training cycles under a fixed training budget.
155
+
156
+ # ACKNOWLEDGEMENTS
157
+
158
+ We thank Ilya Loshchilov and Frank Hutter for their insightful comments on the cyclic cosineshaped learning rate. The authors are supported in part by the, III-1618134, III-1526012, IIS1149882 grants from the National Science Foundation, US Army Research Office W911NF-14- 1-0477, and the Bill and Melinda Gates Foundation.
159
+
160
+ # REFERENCES
161
+
162
+ Leon Bottou. Large-scale machine learning with stochastic gradient descent. In ´ COMPSTAT. 2010.
163
+
164
+ Cristian Bucilu, Rich Caruana, and Alexandru Niculescu-Mizil. Model compression. In KDD, 2006.
165
+
166
+ Rich Caruana, Alexandru Niculescu-Mizil, Geoff Crew, and Alex Ksikes. Ensemble selection from libraries of models. In ICML, 2004.
167
+
168
+ Ronan Collobert, Koray Kavukcuoglu, and Clement Farabet. Torch7: A matlab-like environment ´ for machine learning. In BigLearn, NIPS Workshop, 2011.
169
+
170
+ Yann N Dauphin, Razvan Pascanu, Caglar Gulcehre, Kyunghyun Cho, Surya Ganguli, and Yoshua Bengio. Identifying and attacking the saddle point problem in high-dimensional non-convex optimization. In NIPS, 2014.
171
+
172
+ Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009.
173
+
174
+ John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of Machine Learning Research, 12(Jul):2121–2159, 2011.
175
+
176
+ Ian J Goodfellow, David Warde-Farley, Mehdi Mirza, Aaron Courville, and Yoshua Bengio. Maxout networks. In ICML, 2013.
177
+
178
+ Ian J Goodfellow, Oriol Vinyals, and Andrew M Saxe. Qualitatively characterizing neural network optimization problems. arXiv preprint arXiv:1412.6544, 2014.
179
+
180
+ Lars Kai Hansen and Peter Salamon. Neural network ensembles. IEEE transactions on pattern analysis and machine intelligence, 12:993–1001, 1990.
181
+
182
+ Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016a.
183
+
184
+ Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In ECCV, 2016b.
185
+
186
+ Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015.
187
+
188
+ Gao Huang, Zhuang Liu, and Kilian Q Weinberger. Densely connected convolutional networks. arXiv preprint arXiv:1608.06993, 2016a.
189
+
190
+ Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Weinberger. Deep networks with stochastic depth. In ECCV, 2016b.
191
+
192
+ Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICCV, 2015.
193
+
194
+ Sbastien Jean, Kyunghyun Cho, Roland Memisevic, and Yoshua Bengio. On using very large target vocabulary for neural machine translation. arXiv preprint arXiv:1412.2007, 2014.
195
+
196
+ Kenji Kawaguchi. Deep learning without poor local minima. arXiv preprint arXiv:1605.07110, 2016.
197
+
198
+ Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. arXiv preprint arXiv:1609.04836, 2016.
199
+
200
+ Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.
201
+
202
+ Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. 2009.
203
+
204
+ Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NIPS, 2012.
205
+
206
+ Anders Krogh, Jesper Vedelsby, et al. Neural network ensembles, cross validation, and active learning. In NIPS, volume 7, 1995.
207
+
208
+ David Krueger, Tegan Maharaj, Janos Kram ´ ar, Mohammad Pezeshki, Nicolas Ballas, Nan Rose- ´ mary Ke, Anirudh Goyal, Yoshua Bengio, Hugo Larochelle, Aaron Courville, et al. Zoneout: Regularizing rnns by randomly preserving hidden activations. arXiv preprint arXiv:1606.01305, 2016.
209
+
210
+ Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016.
211
+
212
+ Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Fractalnet: Ultra-deep neural networks without residuals. arXiv preprint arXiv:1605.07648, 2016.
213
+
214
+ Chen-Yu Lee, Saining Xie, Patrick Gallagher, Zhengyou Zhang, and Zhuowen Tu. Deeplysupervised nets. In AISTATS, 2015.
215
+
216
+ Min Lin, Qiang Chen, and Shuicheng Yan. Network in network. arXiv preprint arXiv:1312.4400, 2013.
217
+
218
+ Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with restarts. arXiv preprint arXiv:1608.03983, 2016.
219
+
220
+ Mohammad Moghimi, Mohammad Saberian, Jian Yang, Li-Jia Li, Nuno Vasconcelos, and Serge Belongie. Boosted convolutional neural networks. 2016.
221
+
222
+ Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning, 2011. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning, 2011.
223
+
224
+ Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014.
225
+
226
+ Rico Sennrich, Barry Haddow, and Alexandra Birch. Edinburgh neural machine translation systems for wmt 16. arXiv preprint arXiv:1606.02891, 2016.
227
+
228
+ Pierre Sermanet, Soumith Chintala, and Yann LeCun. Convolutional neural networks applied to house numbers digit classification. In ICPR, 2012.
229
+
230
+ Saurabh Singh, Derek Hoiem, and David Forsyth. Swapout: Learning an ensemble of deep architectures. arXiv preprint arXiv:1605.06465, 2016.
231
+
232
+ Leslie N. Smith. No more pesky learning rate guessing games. CoRR, abs/1506.01186, 2016. URL http://arxiv.org/abs/1506.01186.
233
+
234
+ Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806, 2014.
235
+
236
+ Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(1):1929–1958, 2014.
237
+
238
+ Rupesh Kumar Srivastava, Klaus Greff, and Jurgen Schmidhuber. Highway networks. ¨ arXiv preprint arXiv:1505.00387, 2015.
239
+
240
+ A Swann and N Allinson. Fast committee learning: Preliminary results. Electronics Letters, 34(14): 1408–1410, 1998.
241
+
242
+ Li Wan, Matthew Zeiler, Sixin Zhang, Yann L Cun, and Rob Fergus. Regularization of neural networks using dropconnect. In ICML, 2013.
243
+
244
+ Jingjing Xie, Bing Xu, and Zhang Chuang. Horizontal and vertical ensemble with deep representation for classification. arXiv preprint arXiv:1306.2759, 2013.
245
+
246
+ Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016.
247
+
248
+ # SUPPLEMENTARY
249
+
250
+ # A. Single model and Snapshot Ensemble performance over time
251
+
252
+ In Figures 7-9, we compare the test error of Snapshot Ensembles with the error of individual model snapshots. The blue curve shows the test error of a single model snapshot using a cyclic cosine learning rate. The green curve shows the test error when ensembling model snapshots over time. (Note that, unlike Figure 3, we construct these ensembles beginning with the earliest snapshots.) As a reference, the red dashed line in each panel represents the test error of single model trained for 300 epochs using a standard learning rate schedule. Without Snapshot Ensembles, in about half of the cases, the test error of final model using a cyclic learning rate—the right most point in the blue curve—is no better than using a standard learning rate schedule.
253
+
254
+ One can observe that under almost all settings, complete Snapshot Ensembles—the right most points of the green curves—outperform the single model baselines. In many cases, ensembles of just 2 or 3 model snapshots are able to match the performance of the single model trained with a standard learning rate. Not surprisingly, the ensembles of model snapshots consistently outperform any of its members, yielding a smooth curve of test error over time.
255
+
256
+ ![](images/492226a2f04cafb128e77b6bb42f1568ab5ee2737c3d10eb5641d1478d3d2cf7.jpg)
257
+ Figure 7: Single model and Snapshot Ensemble performance over time (part 1).
258
+
259
+ ![](images/b2db0a7558184db89e197051f6ece7697c6a617925b9be55972e0134e1ce7226.jpg)
260
+ Figure 8: Single model and Snapshot Ensemble performance over time (part 2).
261
+
262
+ ![](images/1119911817b4c7e1d7782e2e076880f8d39c5011c289884028bfb86ba527e0a3.jpg)
263
+ Figure 9: Single model and Snapshot Ensemble performance over time (part 3).
parse/train/BJYwwY9ll/BJYwwY9ll_content_list.json ADDED
@@ -0,0 +1,1446 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "type": "text",
4
+ "text": "SNAPSHOT ENSEMBLES: TRAIN 1, GET M FOR FREE ",
5
+ "text_level": 1,
6
+ "bbox": [
7
+ 174,
8
+ 98,
9
+ 803,
10
+ 122
11
+ ],
12
+ "page_idx": 0
13
+ },
14
+ {
15
+ "type": "text",
16
+ "text": "Gao Huang∗, Yixuan Li∗, Geoff Pleiss \nCornell University \n{gh349, yl2363}@cornell.edu, geoff@cs.cornell.edu \nZhuang Liu \nTsinghua University \nliuzhuangthu@gmail.com ",
17
+ "bbox": [
18
+ 183,
19
+ 147,
20
+ 661,
21
+ 190
22
+ ],
23
+ "page_idx": 0
24
+ },
25
+ {
26
+ "type": "text",
27
+ "text": "",
28
+ "bbox": [
29
+ 184,
30
+ 212,
31
+ 401,
32
+ 253
33
+ ],
34
+ "page_idx": 0
35
+ },
36
+ {
37
+ "type": "text",
38
+ "text": "John E. Hopcroft, Kilian Q. Weinberger Cornell University jeh@cs.cornell.edu, kqw4@cornell.edu ",
39
+ "bbox": [
40
+ 437,
41
+ 210,
42
+ 794,
43
+ 253
44
+ ],
45
+ "page_idx": 0
46
+ },
47
+ {
48
+ "type": "text",
49
+ "text": "ABSTRACT ",
50
+ "text_level": 1,
51
+ "bbox": [
52
+ 454,
53
+ 290,
54
+ 544,
55
+ 305
56
+ ],
57
+ "page_idx": 0
58
+ },
59
+ {
60
+ "type": "text",
61
+ "text": "Ensembles of neural networks are known to be much more robust and accurate than individual networks. However, training multiple deep networks for model averaging is computationally expensive. In this paper, we propose a method to obtain the seemingly contradictory goal of ensembling multiple neural networks at no additional training cost. We achieve this goal by training a single neural network, converging to several local minima along its optimization path and saving the model parameters. To obtain repeated rapid convergence, we leverage recent work on cyclic learning rate schedules. The resulting technique, which we refer to as Snapshot Ensembling, is simple, yet surprisingly effective. We show in a series of experiments that our approach is compatible with diverse network architectures and learning tasks. It consistently yields lower error rates than state-of-the-art single models at no additional training cost, and compares favorably with traditional network ensembles. On CIFAR-10 and CIFAR-100 our DenseNet Snapshot Ensembles obtain error rates of $3 . 4 \\%$ and $1 7 . 4 \\%$ respectively. ",
62
+ "bbox": [
63
+ 233,
64
+ 323,
65
+ 764,
66
+ 516
67
+ ],
68
+ "page_idx": 0
69
+ },
70
+ {
71
+ "type": "text",
72
+ "text": "1 INTRODUCTION ",
73
+ "text_level": 1,
74
+ "bbox": [
75
+ 176,
76
+ 546,
77
+ 336,
78
+ 561
79
+ ],
80
+ "page_idx": 0
81
+ },
82
+ {
83
+ "type": "text",
84
+ "text": "Stochastic Gradient Descent (SGD) (Bottou, 2010) and its accelerated variants (Kingma & Ba, 2014; Duchi et al., 2011) have become the de-facto approaches for optimizing deep neural networks. The popularity of SGD can be attributed to its ability to avoid and even escape spurious saddle-points and local minima (Dauphin et al., 2014). Although avoiding these spurious solutions is generally considered positive, in this paper we argue that these local minima contain useful information that may in fact improve model performance. ",
85
+ "bbox": [
86
+ 174,
87
+ 578,
88
+ 823,
89
+ 661
90
+ ],
91
+ "page_idx": 0
92
+ },
93
+ {
94
+ "type": "text",
95
+ "text": "Although deep networks typically never converge to a global minimum, there is a notion of “good” and “bad” local minima with respect to generalization. Keskar et al. (2016) argue that local minima with flat basins tend to generalize better. SGD tends to avoid sharper local minima because gradients are computed from small mini-batches and are therefore inexact (Keskar et al., 2016). If the learningrate is sufficiently large, the intrinsic random motion across gradient steps prevents the optimizer from reaching any of the sharp basins along its optimization path. However, if the learning rate is small, the model tends to converge into the closest local minimum. These two very different behaviors of SGD are typically exploited in different phases of optimization (He et al., 2016a). Initially the learning rate is kept high to move into the general vicinity of a flat local minimum. Once this search has reached a stage in which no further progress is made, the learning rate is dropped (once or twice), triggering a descent, and ultimately convergence, to the final local minimum. ",
96
+ "bbox": [
97
+ 174,
98
+ 669,
99
+ 825,
100
+ 821
101
+ ],
102
+ "page_idx": 0
103
+ },
104
+ {
105
+ "type": "text",
106
+ "text": "It is well established (Kawaguchi, 2016) that the number of possible local minima grows exponentially with the number of parameters—of which modern neural networks can have millions. It is therefore not surprising that two identical architectures optimized with different initializations or minibatch orderings will converge to different solutions. Although different local minima often have very similar error rates, the corresponding neural networks tend to make different mistakes. This diversity can be exploited through ensembling, in which multiple neural networks are trained from different initializations and then combined with majority voting or averaging (Caruana et al., 2004). Ensembling often leads to drastic reductions in error rates. In fact, most high profile competitions, e.g. Imagenet (Deng et al., 2009) or Kaggle1, are won by ensembles of deep learning architectures. ",
107
+ "bbox": [
108
+ 174,
109
+ 829,
110
+ 823,
111
+ 898
112
+ ],
113
+ "page_idx": 0
114
+ },
115
+ {
116
+ "type": "image",
117
+ "img_path": "images/1f88b86ec2fd97a2263949838ab06c8e2bf496e67dc1f5e93984e1f4e23ef598.jpg",
118
+ "image_caption": [
119
+ "Figure 1: Left: Illustration of SGD optimization with a typical learning rate schedule. The model converges to a minimum at the end of training. Right: Illustration of Snapshot Ensembling. The model undergoes several learning rate annealing cycles, converging to and escaping from multiple local minima. We take a snapshot at each minimum for test-time ensembling. "
120
+ ],
121
+ "image_footnote": [],
122
+ "bbox": [
123
+ 202,
124
+ 88,
125
+ 795,
126
+ 272
127
+ ],
128
+ "page_idx": 1
129
+ },
130
+ {
131
+ "type": "text",
132
+ "text": "",
133
+ "bbox": [
134
+ 174,
135
+ 375,
136
+ 823,
137
+ 431
138
+ ],
139
+ "page_idx": 1
140
+ },
141
+ {
142
+ "type": "text",
143
+ "text": "Despite its obvious advantages, the use of ensembling for deep networks is not nearly as widespread as it is for other algorithms. One likely reason for this lack of adaptation may be the cost of learning multiple neural networks. Training deep networks can last for weeks, even on high performance hardware with GPU acceleration. As the training cost for ensembles increases linearly, ensembles can quickly becomes uneconomical for most researchers without access to industrial scale computational resources. ",
144
+ "bbox": [
145
+ 174,
146
+ 438,
147
+ 825,
148
+ 521
149
+ ],
150
+ "page_idx": 1
151
+ },
152
+ {
153
+ "type": "text",
154
+ "text": "In this paper we focus on the seemingly-contradictory goal of learning an ensemble of multiple neural networks without incurring any additional training costs. We achieve this goal with a training method that is simple and straight-forward to implement. Our approach leverages the non-convex nature of neural networks and the ability of SGD to converge to and escape from local minima on demand. Instead of training $M$ neural networks independently from scratch, we let SGD converge $M$ times to local minima along its optimization path. Each time the model converges, we save the weights and add the corresponding network to our ensemble. We then restart the optimization with a large learning rate to escape the current local minimum. More specifically, we adopt the cycling procedure suggested by Loshchilov & Hutter (2016), in which the learning rate is abruptly raised and then quickly lowered to follow a cosine function. Because our final ensemble consists of snapshots of the optimization path, we refer to our approach as Snapshot Ensembling. Figure 1 presents a high-level overview of this method. ",
155
+ "bbox": [
156
+ 173,
157
+ 529,
158
+ 825,
159
+ 694
160
+ ],
161
+ "page_idx": 1
162
+ },
163
+ {
164
+ "type": "text",
165
+ "text": "In contrast to traditional ensembles, the training time for the entire ensemble is identical to the time required to train a single traditional model. During testing time, one can evaluate and average the last (and therefore most accurate) $m$ out of $M$ models. Our approach is naturally compatible with other methods to improve the accuracy, such as data augmentation, stochastic depth (Huang et al., 2016b), or batch normalization (Ioffe & Szegedy, 2015). In fact, Snapshot Ensembles can even be ensembled, if for example parallel resources are available during training. In this case, an ensemble of $K$ Snapshot Ensembles yields $K \\times M$ models at $K$ times the training cost. ",
166
+ "bbox": [
167
+ 174,
168
+ 702,
169
+ 825,
170
+ 799
171
+ ],
172
+ "page_idx": 1
173
+ },
174
+ {
175
+ "type": "text",
176
+ "text": "We evaluate the efficacy of Snapshot Ensembles on three state-of-the-art deep learning architectures for object recognition: ResNet (He et al., 2016b), Wide-ResNet (Zagoruyko & Komodakis, 2016), and DenseNet (Huang et al., 2016a). We show across four different data sets that Snapshot Ensembles almost always reduce error without increasing training costs. For example, on CIFAR-10 and CIFAR-100, Snapshot Ensembles obtains error rates of $3 . 4 \\bar { 4 } \\%$ and $1 7 . 4 1 \\%$ respectively. ",
177
+ "bbox": [
178
+ 174,
179
+ 806,
180
+ 825,
181
+ 876
182
+ ],
183
+ "page_idx": 1
184
+ },
185
+ {
186
+ "type": "text",
187
+ "text": "2 RELATED WORK ",
188
+ "text_level": 1,
189
+ "bbox": [
190
+ 176,
191
+ 102,
192
+ 344,
193
+ 117
194
+ ],
195
+ "page_idx": 2
196
+ },
197
+ {
198
+ "type": "text",
199
+ "text": "Neural network ensembles have been widely studied and applied in machine learning (Hansen & Salamon, 1990; Krogh et al., 1995). However, most of these prior studies focus on improving the generalization performance, while few of them address the cost of training ensembles. ",
200
+ "bbox": [
201
+ 176,
202
+ 133,
203
+ 823,
204
+ 175
205
+ ],
206
+ "page_idx": 2
207
+ },
208
+ {
209
+ "type": "text",
210
+ "text": "As an alternative to traditional ensembles, so-called “implicit” ensembles have high efficiency during both training and testing (Srivastava et al., 2014; Wan et al., 2013; Huang et al., 2016b; Singh et al., 2016; Krueger et al., 2016). The Dropout (Srivastava et al., 2014) technique creates an ensemble out of a single model by “dropping” — or zeroing — random sets of hidden nodes during each mini-batch. At test time, no nodes are dropped, and each node is scaled by the probability of surviving during training. Srivastava et al. claim that Dropout reduces overfitting by preventing the co-adaptation of nodes. An alternative explanation is that this mechanism creates an exponential number of networks with shared weights during training, which are then implicitly ensembled at test time. DropConnect (Wan et al., 2013) uses a similar trick to create ensembles at test time by dropping connections (weights) during training instead of nodes. The recently proposed Stochastic Depth technique (Huang et al., 2016b) randomly drops layers during training to create an implicit ensemble of networks with varying depth at test time. Finally, Swapout (Singh et al., 2016) is a stochastic training method that generalizes Dropout and Stochastic Depth. From the perspective of model ensembling, Swapout creates diversified network structures for model averaging. Our proposed method similarly trains only a single model; however, the resulting ensemble is “explicit” in that the models do not share weights. Furthermore, our method can be used in conjunction with any of these implicit ensembling techniques. ",
211
+ "bbox": [
212
+ 174,
213
+ 183,
214
+ 825,
215
+ 417
216
+ ],
217
+ "page_idx": 2
218
+ },
219
+ {
220
+ "type": "text",
221
+ "text": "Several recent publications focus on reducing the test time cost of ensembles, by transferring the “knowledge” of cumbersome ensembles into a single model (Bucilu et al., 2006; Hinton et al., 2015). Hinton et al. (2015) propose to use an ensemble of multiple networks as the target of a single (smaller) network. Our proposed method is complementary to these works as we aim to reduce the training cost of ensembles rather than the test-time cost. ",
222
+ "bbox": [
223
+ 174,
224
+ 426,
225
+ 823,
226
+ 494
227
+ ],
228
+ "page_idx": 2
229
+ },
230
+ {
231
+ "type": "text",
232
+ "text": "Perhaps most similar to our work is that of Swann & Allinson (1998) and Xie et al. (2013), who explore creating ensembles from slices of the learning trajectory. Xie et al. introduce the horizontal and vertical ensembling method, which combines the output of networks within a range of training epochs. More recently, Jean et al. (2014) and Sennrich et al. (2016) show improvement by ensembling the intermediate stages of model training. Laine & Aila (2016) propose a temporal ensembling method for semi-supervised learning, which achieves consensus among models trained with different regularization and augmentation conditions for better generalization performance. Finally, Moghimi et al. (2016) show that boosting can be applied to convolutional neural networks to create strong ensembles. Our work differs from these prior works in that we force the model to visit multiple local minima, and we take snapshots only when the model reaches a minimum. We believe this key insight allows us to leverage more power from our ensembles. ",
233
+ "bbox": [
234
+ 174,
235
+ 502,
236
+ 825,
237
+ 655
238
+ ],
239
+ "page_idx": 2
240
+ },
241
+ {
242
+ "type": "text",
243
+ "text": "Our work is inspired by the recent findings of Loshchilov & Hutter (2016) and Smith (2016), who show that cyclic learning rates can be effective for training convolutional neural networks. The authors show that each cycle produces models which are (almost) competitive to those learned with traditional learning rate schedules while requiring a fraction of training iterations. Although model performance temporarily suffers when the learning rate cycle is restarted, the performance eventually surpasses the previous cycle after annealing the learning rate. The authors suggest that cycling perturbs the parameters of a converged model, which allows the model to find a better local minimum. We build upon these recent findings by (1) showing that there is significant diversity in the local minima visited during each cycle and (2) exploiting this diversity using ensembles. We are not concerned with speeding up or improving the training of a single model; rather, our goal is to extract an ensemble of classifiers while following the optimization path of the final model. ",
244
+ "bbox": [
245
+ 174,
246
+ 661,
247
+ 825,
248
+ 815
249
+ ],
250
+ "page_idx": 2
251
+ },
252
+ {
253
+ "type": "text",
254
+ "text": "3 SNAPSHOT ENSEMBLING ",
255
+ "text_level": 1,
256
+ "bbox": [
257
+ 176,
258
+ 835,
259
+ 413,
260
+ 852
261
+ ],
262
+ "page_idx": 2
263
+ },
264
+ {
265
+ "type": "text",
266
+ "text": "Snapshot Ensembling produces an ensemble of accurate and diverse models from a single training process. At the heart of Snapshot Ensembling is an optimization process which visits several local minima before converging to a final solution. We take model snapshots at these various minima, and average their predictions at test time. ",
267
+ "bbox": [
268
+ 174,
269
+ 867,
270
+ 823,
271
+ 922
272
+ ],
273
+ "page_idx": 2
274
+ },
275
+ {
276
+ "type": "text",
277
+ "text": "Ensembles work best if the individual models (1) have low test error and (2) do not overlap in the set of examples they misclassify. Along most of the optimization path, the weight assignments of a neural network tend not to correspond to low test error. In fact, it is commonly observed that the validation error drops significantly only after the learning rate has been reduced, which is typically done after several hundred epochs. Our approach is inspired by the observation that training neural networks for fewer epochs and dropping the learning rate earlier has minor impact on the final test error (Loshchilov & Hutter, 2016). This seems to suggest that local minima along the optimization path become promising (in terms of generalization error) after only a few epochs. ",
278
+ "bbox": [
279
+ 173,
280
+ 103,
281
+ 825,
282
+ 215
283
+ ],
284
+ "page_idx": 3
285
+ },
286
+ {
287
+ "type": "text",
288
+ "text": "Cyclic Cosine Annealing. To converge to multiple local minima, we follow a cyclic annealing schedule as proposed by Loshchilov & Hutter (2016). We lower the learning rate at a very fast pace, encouraging the model to converge towards its first local minimum after as few as 50 epochs. The optimization is then continued at a larger learning rate, which perturbs the model and dislodges it from the minimum. We repeat this process several times to obtain multiple convergences. Formally, the learning rate $\\alpha$ has the form: ",
289
+ "bbox": [
290
+ 174,
291
+ 219,
292
+ 485,
293
+ 386
294
+ ],
295
+ "page_idx": 3
296
+ },
297
+ {
298
+ "type": "equation",
299
+ "img_path": "images/81c9fab27ee87a992a7cae779f668793fa7975ad241ed401ed5a2968e38a8f76.jpg",
300
+ "text": "$$\n\\alpha ( t ) = f \\left( \\mathrm { m o d } \\left( t - 1 , \\lceil T / M \\rceil \\right) \\right) ,\n$$",
301
+ "text_format": "latex",
302
+ "bbox": [
303
+ 217,
304
+ 404,
305
+ 439,
306
+ 422
307
+ ],
308
+ "page_idx": 3
309
+ },
310
+ {
311
+ "type": "text",
312
+ "text": "where $t$ is the iteration number, $T$ is the total number of training iterations, and $f$ is a monotonically decreasing function. In other words, we split the training process into $M$ cycles, each of which starts with a large learning rate, which is annealed to a smaller learning rate. The large learning rate $\\alpha = f ( 0 )$ provides the model enough energy to escape from a critical point, while the small learning rate ",
313
+ "bbox": [
314
+ 174,
315
+ 439,
316
+ 483,
317
+ 564
318
+ ],
319
+ "page_idx": 3
320
+ },
321
+ {
322
+ "type": "text",
323
+ "text": "$\\alpha = f ( \\lceil \\bar { T } / M \\rceil )$ drives the model to a well behaved local minimum. In our experiments, we set $f$ to be the shifted cosine function proposed by Loshchilov & Hutter (2016): ",
324
+ "bbox": [
325
+ 174,
326
+ 564,
327
+ 823,
328
+ 593
329
+ ],
330
+ "page_idx": 3
331
+ },
332
+ {
333
+ "type": "image",
334
+ "img_path": "images/c2ae06eb52ffabda7683124a8942902ce5bda2833139f06bb9d05f36fe2f9af3.jpg",
335
+ "image_caption": [
336
+ "Figure 2: Training loss of 100-layer DenseNet on CIFAR10 using standard learning rate (blue) and $M = 6$ cosine annealing cycles (red). The intermediate models, denoted by the dotted lines, form an ensemble at the end of training. "
337
+ ],
338
+ "image_footnote": [],
339
+ "bbox": [
340
+ 496,
341
+ 234,
342
+ 818,
343
+ 458
344
+ ],
345
+ "page_idx": 3
346
+ },
347
+ {
348
+ "type": "equation",
349
+ "img_path": "images/a7e219004a04fed6d1eb14024e8c3cd8ee01c3441c5d53732dc269b8c456237d.jpg",
350
+ "text": "$$\n\\alpha ( t ) = \\frac { \\alpha _ { 0 } } { 2 } \\left( \\cos \\left( \\frac { \\pi \\mathrm { m o d } ( t - 1 , \\left\\lceil T / M \\right\\rceil ) } { \\left\\lceil T / M \\right\\rceil } \\right) + 1 \\right) ,\n$$",
351
+ "text_format": "latex",
352
+ "bbox": [
353
+ 331,
354
+ 609,
355
+ 665,
356
+ 646
357
+ ],
358
+ "page_idx": 3
359
+ },
360
+ {
361
+ "type": "text",
362
+ "text": "where $\\alpha _ { 0 }$ is the initial learning rate. Intuitively, this function anneals the learning rate from its initial value $\\alpha _ { 0 }$ to $f ( \\lceil T / M \\rceil ) \\approx 0$ over the course of a cycle. Following (Loshchilov & Hutter, 2016), we update the learning rate at each iteration rather than at every epoch. This improves the convergence of short cycles, even when a large initial learning rate is used. ",
363
+ "bbox": [
364
+ 174,
365
+ 661,
366
+ 825,
367
+ 718
368
+ ],
369
+ "page_idx": 3
370
+ },
371
+ {
372
+ "type": "text",
373
+ "text": "Snapshot Ensembling. Figure 2 depicts the training process using cyclic and traditional learning rate schedules. At the end of each training cycle, it is apparent that the model reaches a local minimum with respect to the training loss. Thus, before raising the learning rate, we take a “snapshot” of the model weights (indicated as vertical dashed black lines). After training $M$ cycles, we have $M$ model snapshots, $f _ { 1 } \\ldots f _ { M }$ , each of which will be used in the final ensemble. It is important to highlight that the total training time of the $M$ snapshots is the same as training a model with a standard schedule (indicated in blue). In some cases, the standard learning rate schedule achieves lower training loss than the cyclic schedule; however, as we will show in the next section, the benefits of ensembling outweigh this difference. ",
374
+ "bbox": [
375
+ 173,
376
+ 722,
377
+ 825,
378
+ 848
379
+ ],
380
+ "page_idx": 3
381
+ },
382
+ {
383
+ "type": "text",
384
+ "text": "Ensembling at Test Time. The ensemble prediction at test time is the average of the last $m$ $( m ~ \\leq ~ M )$ model’s softmax outputs. Let $\\mathbf { x }$ be a test sample and let $h _ { i } \\left( \\mathbf { x } \\right)$ be the softmax $i$ t of the ensemble is a simple. We always ensemble the last verage of the last models, as these $m$ models:dels tend $h _ { \\mathrm { E n s e m b l e } } = \\textstyle { \\frac { 1 } { m } } \\sum _ { 0 } ^ { m - 1 } h _ { M - i } ( { \\bf x } )$ $m$ ",
385
+ "bbox": [
386
+ 174,
387
+ 852,
388
+ 823,
389
+ 924
390
+ ],
391
+ "page_idx": 3
392
+ },
393
+ {
394
+ "type": "table",
395
+ "img_path": "images/e7b464b855b23017174fc0c16e804533ccf6279849a6b6b2937b959a7941a549.jpg",
396
+ "table_caption": [
397
+ "Table 1: Error rates $( \\% )$ on CIFAR-10 and CIFAR-100 datasets. All methods in the same group are trained for the same number of iterations. Results of our method are colored in blue, and the best result for each network/dataset pair are bolded. ∗ indicates numbers which we take directly from Huang et al. (2016a). "
398
+ ],
399
+ "table_footnote": [],
400
+ "table_body": "<table><tr><td colspan=\"2\">Method</td><td>C10</td><td>C100</td><td>SVHN</td><td>Tiny ImageNet</td></tr><tr><td rowspan=\"5\">ResNet-110</td><td>Single model</td><td>5.52</td><td>28.02</td><td>1.96</td><td>46.50</td></tr><tr><td>NoCycle Snapshot Ensemble</td><td>5.49</td><td>26.97</td><td>1.78</td><td>43.69</td></tr><tr><td>SingleCycle Ensembles</td><td>6.66</td><td>24.54</td><td>1.74</td><td>42.60</td></tr><tr><td>Snapshot Ensemble (αo = 0.1)</td><td>5.73</td><td>25.55</td><td>1.63</td><td>40.54</td></tr><tr><td>Snapshot Ensemble (αo = 0.2)</td><td>5.32</td><td>24.19</td><td>1.66</td><td>39.40</td></tr><tr><td rowspan=\"6\">Wide-ResNet-32</td><td>Single model</td><td>5.43</td><td>23.55</td><td>1.90</td><td>39.63</td></tr><tr><td>Dropout</td><td>4.68</td><td>22.82</td><td>1.81</td><td>36.58</td></tr><tr><td>NoCycle Snapshot Ensemble</td><td>5.18</td><td>22.81</td><td>1.81</td><td>38.64</td></tr><tr><td>SingleCycle Ensembles</td><td>5.95</td><td>21.38</td><td>1.65</td><td>35.53</td></tr><tr><td>Snapshot Ensemble (αo = 0.1)</td><td>4.41</td><td>21.26</td><td>1.64</td><td>35.45</td></tr><tr><td>Snapshot Ensemble(αo = 0.2)</td><td>4.73</td><td>21.56</td><td>1.51</td><td>32.90</td></tr><tr><td rowspan=\"6\">DenseNet-40</td><td>Single model</td><td>5.24*</td><td>24.42*</td><td>1.77</td><td>39.09</td></tr><tr><td>Dropout</td><td>6.08</td><td>25.79</td><td>1.79*</td><td>39.68</td></tr><tr><td>NoCycle Snapshot Ensemble</td><td>5.20</td><td>24.63</td><td>1.80</td><td>38.51</td></tr><tr><td>SingleCycle Ensembles</td><td>5.43</td><td>22.51</td><td>1.87</td><td>38.00</td></tr><tr><td>Snapshot Ensemble (αo = 0.1)</td><td>4.99</td><td>23.34</td><td>1.64</td><td>37.25</td></tr><tr><td>Snapshot Ensemble (αo = 0.2)</td><td>4.84</td><td>21.93</td><td>1.73</td><td>36.61</td></tr><tr><td rowspan=\"6\">DenseNet-100</td><td>Single model</td><td>3.74*</td><td>19.25*</td><td></td><td>=</td></tr><tr><td>Dropout</td><td>3.65</td><td>18.77</td><td>=</td><td></td></tr><tr><td>NoCycle Snapshot Ensemble</td><td>3.80</td><td>19.30</td><td></td><td></td></tr><tr><td>SingleCycle Ensembles</td><td>4.52</td><td>18.38</td><td></td><td></td></tr><tr><td>Snapshot Ensemble (αo = 0.1)</td><td>3.57</td><td>18.12</td><td></td><td></td></tr><tr><td>Snapshot Ensemble (αo = 0.2)</td><td>3.44</td><td>17.41</td><td></td><td></td></tr></table>",
401
+ "bbox": [
402
+ 181,
403
+ 101,
404
+ 816,
405
+ 438
406
+ ],
407
+ "page_idx": 4
408
+ },
409
+ {
410
+ "type": "text",
411
+ "text": "4 EXPERIMENTS ",
412
+ "text_level": 1,
413
+ "bbox": [
414
+ 176,
415
+ 515,
416
+ 326,
417
+ 530
418
+ ],
419
+ "page_idx": 4
420
+ },
421
+ {
422
+ "type": "text",
423
+ "text": "We demonstrate the effectiveness of Snapshot Ensembles on several benchmark datasets, comparing with competitive baselines. We run all experiments with Torch 7 (Collobert et al., 2011)2. ",
424
+ "bbox": [
425
+ 176,
426
+ 547,
427
+ 825,
428
+ 575
429
+ ],
430
+ "page_idx": 4
431
+ },
432
+ {
433
+ "type": "text",
434
+ "text": "4.1 DATASETS ",
435
+ "text_level": 1,
436
+ "bbox": [
437
+ 174,
438
+ 594,
439
+ 287,
440
+ 609
441
+ ],
442
+ "page_idx": 4
443
+ },
444
+ {
445
+ "type": "text",
446
+ "text": "CIFAR. The two CIFAR datasets (Krizhevsky & Hinton, 2009) consist of colored natural images sized at $3 2 \\times 3 2$ pixels. CIFAR-10 (C10) and CIFAR-100 (C100) images are drawn from 10 and 100 classes, respectively. For each dataset, there are 50,000 training images and 10,000 images reserved for testing. We use a standard data augmentation scheme (Lin et al., 2013; Romero et al., 2014; Lee et al., 2015; Springenberg et al., 2014; Srivastava et al., 2015; Huang et al., 2016b; Larsson et al., 2016), in which the images are zero-padded with 4 pixels on each side, randomly cropped to produce $3 2 \\times 3 2$ images, and horizontally mirrored with probability 0.5. ",
447
+ "bbox": [
448
+ 173,
449
+ 618,
450
+ 825,
451
+ 717
452
+ ],
453
+ "page_idx": 4
454
+ },
455
+ {
456
+ "type": "text",
457
+ "text": "SVHN. The Street View House Numbers (SVHN) dataset (Netzer et al., 2011) contains $3 2 \\times 3 2$ colored digit images from Google Street View, with one class for each digit. There are 73,257 images in the training set and 26,032 images in the test set. Following common practice (Sermanet et al., 2012; Goodfellow et al., 2013; Huang et al., 2016a), we withhold 6,000 training images for validation, and train on the remaining images without data augmentation. ",
458
+ "bbox": [
459
+ 174,
460
+ 719,
461
+ 825,
462
+ 790
463
+ ],
464
+ "page_idx": 4
465
+ },
466
+ {
467
+ "type": "text",
468
+ "text": "Tiny ImageNet. The Tiny ImageNet dataset3 consists of a subset of ImageNet images (Deng et al., 2009). There are 200 classes, each of which has 500 training images and 50 validation images. Each image is resized to $6 4 \\times 6 4$ and augmented with random crops, horizontal mirroring, and RGB intensity scaling (Krizhevsky et al., 2012). ",
469
+ "bbox": [
470
+ 174,
471
+ 794,
472
+ 825,
473
+ 851
474
+ ],
475
+ "page_idx": 4
476
+ },
477
+ {
478
+ "type": "text",
479
+ "text": "ImageNet. The ILSVRC 2012 classification dataset (Deng et al., 2009) consists of 1000 images classes, with a total of 1.2 million training images and 50,000 validation images. We adopt the same data augmentation scheme as in (He et al., 2016a; Huang et al., 2016a) and apply a $2 2 4 \\times 2 2 4$ center crop to images at test time. ",
480
+ "bbox": [
481
+ 174,
482
+ 854,
483
+ 823,
484
+ 882
485
+ ],
486
+ "page_idx": 4
487
+ },
488
+ {
489
+ "type": "image",
490
+ "img_path": "images/df25f7f707d94414d7c1652a477801315c74e97f6c0090e95d159019b040c02a.jpg",
491
+ "image_caption": [
492
+ "Figure 3: DenseNet-100 Snapshot Ensemble performance on CIFAR-10 and CIFAR-100 with restart learning rate $\\alpha _ { 0 } = 0 . 1$ (left two) and $\\alpha _ { 0 } = 0 . 2$ (right two). Each ensemble is trained with $M = 6$ annealing cycles (50 epochs per each). "
493
+ ],
494
+ "image_footnote": [],
495
+ "bbox": [
496
+ 178,
497
+ 98,
498
+ 823,
499
+ 217
500
+ ],
501
+ "page_idx": 5
502
+ },
503
+ {
504
+ "type": "text",
505
+ "text": "",
506
+ "bbox": [
507
+ 176,
508
+ 285,
509
+ 823,
510
+ 313
511
+ ],
512
+ "page_idx": 5
513
+ },
514
+ {
515
+ "type": "text",
516
+ "text": "4.2 TRAINING SETTING ",
517
+ "text_level": 1,
518
+ "bbox": [
519
+ 176,
520
+ 335,
521
+ 351,
522
+ 349
523
+ ],
524
+ "page_idx": 5
525
+ },
526
+ {
527
+ "type": "text",
528
+ "text": "Architectures. We test several state-of-the-art architectures, including residual networks (ResNet) (He et al., 2016a), Wide ResNet (Zagoruyko & Komodakis, 2016) and DenseNet (Huang et al., 2016a). For ResNet, we use the original 110-layer network introduced by He et al. (2016a). Wide-ResNet is a 32-layer ResNet with 4 times as many convolutional features per layer as a standard ResNet. For DenseNet, our large model follows the same setup as (Huang et al., 2016a), with depth $L = 1 0 0$ and growth rate $k = 2 4$ . In addition, we also evaluate our method on a small DenseNet, with depth $L = 4 0$ and $k = 1 2$ . To adapt all these networks to Tiny ImageNet, we add a stride of 2 to the first layer of the models, which downsamples the images to $3 2 \\times 3 2$ . For ImageNet, we test the 50-layer ResNet proposed in (He et al., 2016a). We use a mini batch size of 64.4 ",
529
+ "bbox": [
530
+ 174,
531
+ 361,
532
+ 825,
533
+ 487
534
+ ],
535
+ "page_idx": 5
536
+ },
537
+ {
538
+ "type": "text",
539
+ "text": "Baselines. Snapshot Ensembles incur the training cost of a single model; therefore, we compare with baselines that require the same amount of training. First, we compare against a Single Model trained with a standard learning rate schedule, dropping the learning rate from 0.1 to 0.01 halfway through training, and then to 0.001 when training is at $7 5 \\%$ . Additionally, to compare against implicit ensembling methods, we test against a single model trained with Dropout. This baseline uses the same learning rate as above, and drops nodes during training with a probability of 0.2. ",
540
+ "bbox": [
541
+ 174,
542
+ 491,
543
+ 825,
544
+ 574
545
+ ],
546
+ "page_idx": 5
547
+ },
548
+ {
549
+ "type": "text",
550
+ "text": "We then test the Snapshot Ensemble algorithm trained with the cyclic cosine learning rate as described in (2). We test models with the max learning rate $\\alpha _ { 0 }$ set to 0.1 and 0.2. In both cases, we divide the training process into learning rate cycles. Model snapshots are taken after each learning rate cycle. Additionally, we train a Snapshot Ensemble with a non-cyclic learning rate schedule. This NoCycle Snapshot Ensemble, which uses the same schedule as the Single Model and Dropout baselines, is meant to highlight the impact of cyclic learning rates for our method. To accurately compare with the cyclic Snapshot Ensembles, we take the same number of snapshots equally spaced throughout the training process. Finally, we compare against SingleCycle Ensembles, a Snapshot Ensemble variant in which the network is re-initialized at the beginning of every cosine learning rate cycle, rather than using the parameters from the previous optimization cycle. This baseline essentially creates a traditional ensemble, yet each network only has $1 / M$ of the typical training time. This variant is meant to highlight the tradeoff between model diversity and model convergence. Though SingleCycle Ensembles should in theory explore more of the parameter space, the models do not benefit from the optimization of previous cycles. ",
551
+ "bbox": [
552
+ 174,
553
+ 582,
554
+ 825,
555
+ 776
556
+ ],
557
+ "page_idx": 5
558
+ },
559
+ {
560
+ "type": "text",
561
+ "text": "Training Budget. On CIFAR datasets, the training budget is $B = 3 0 0$ epochs for DenseNet-40 and DenseNet-100, and $B = 2 0 0$ for ResNet and Wide ResNet models. Snapshot variants are trained with $M = 6$ cycles of $B / M = 5 0$ epochs for DenseNets, and $M = 5$ cycles of $B / M = 4 0$ epochs for ResNets/Wide ResNets. SVHN models are trained with a budget of $B = 4 0$ epochs (5 cycles of 8 epochs). For Tiny ImageNet, we use a training budget of $B = 1 5 0$ (6 cycles of 25 epochs). Finally, ImageNet is trained with a budget of $B = 9 0$ epochs, and we trained 2 Snapshot variants: one with $M = 2$ cycles and one with $M = 3$ . ",
562
+ "bbox": [
563
+ 173,
564
+ 780,
565
+ 825,
566
+ 877
567
+ ],
568
+ "page_idx": 5
569
+ },
570
+ {
571
+ "type": "text",
572
+ "text": "4.3 SNAPSHOT ENSEMBLE RESULTS ",
573
+ "text_level": 1,
574
+ "bbox": [
575
+ 176,
576
+ 103,
577
+ 436,
578
+ 117
579
+ ],
580
+ "page_idx": 6
581
+ },
582
+ {
583
+ "type": "text",
584
+ "text": "Accuracy. The main results are summarized in Table 1. In most cases, Snapshot ensembles achieve lower error than any of the baseline methods. Most notably, Snapshot Ensembles yield an error rate of $1 7 . 4 1 \\%$ on CIFAR100 using large DenseNets, far outperforming the record of ${ \\mathrm { \\bar { 1 9 . 2 5 \\% } } }$ under the same training cost and architecture (Huang et al., 2016a). Our method has the most success on CIFAR-100 and Tiny ImageNet, which is likely due to the ",
585
+ "bbox": [
586
+ 174,
587
+ 128,
588
+ 483,
589
+ 267
590
+ ],
591
+ "page_idx": 6
592
+ },
593
+ {
594
+ "type": "table",
595
+ "img_path": "images/f4c5d33a1c4c89f9486f91836b6fa120405af452eff344a081e38f68846e8dc0.jpg",
596
+ "table_caption": [
597
+ "Table 2: Top-1 error rates $( \\% )$ on ImageNet validation set using ResNet-50 with varying number of cycles. "
598
+ ],
599
+ "table_footnote": [],
600
+ "table_body": "<table><tr><td>Method</td><td>Val. Error (%)</td></tr><tr><td>Single model</td><td>24.01</td></tr><tr><td>Snapshot Ensemble (M= 2)</td><td>23.33</td></tr><tr><td>Snapshot Ensemble (M= 3)</td><td>23.96</td></tr></table>",
601
+ "bbox": [
602
+ 504,
603
+ 145,
604
+ 813,
605
+ 212
606
+ ],
607
+ "page_idx": 6
608
+ },
609
+ {
610
+ "type": "text",
611
+ "text": "complexity of these datasets. The softmax outputs for these datasets are high dimensional due to the large number of classes, making it unlikely that any two models make the same predictions. Snapshot Ensembling is also capable of improving the competitive baselines for CIFAR-10 and SVHN as well, reducing error by $1 \\%$ and $0 . 4 \\%$ respectively with the Wide ResNet architecture. ",
612
+ "bbox": [
613
+ 174,
614
+ 267,
615
+ 825,
616
+ 323
617
+ ],
618
+ "page_idx": 6
619
+ },
620
+ {
621
+ "type": "text",
622
+ "text": "The NoCycle Snapshot Ensemble generally has little effect on performance, and in some instances even increases the test error. This highlights the need for a cyclic learning rate for useful ensembling. The SingleCycle Ensemble has similarly mixed performance. In some cases, e.g., DenseNet-40 on CIFAR-100, the SingleCycle Ensemble is competitive with Snapshot Ensembles. However, as the model size increases to 100 layers, it does not perform as well. This is because it is difficult to train a large model from scratch in only a few epochs. These results demonstrate that Snapshot Ensembles tend to work best when utilizing information from previous cycles. Effectively, Snapshot Ensembles strike a balance between model diversity and optimization. ",
623
+ "bbox": [
624
+ 173,
625
+ 329,
626
+ 825,
627
+ 440
628
+ ],
629
+ "page_idx": 6
630
+ },
631
+ {
632
+ "type": "text",
633
+ "text": "Table 2 shows Snapshot Ensemble results on ImageNet. The Snapshot Ensemble with $M \\ = \\ 2$ achieves $2 3 . 3 3 \\%$ validation error, outperforming the single model baseline with $2 4 . 0 1 \\%$ validation error. It appears that 2 cycles is the optimal choice for the ImageNet dataset. Provided with the limited total training budget $B = 9 0$ epochs, we hypothesize that allocating fewer than $B / 2 = 4 5$ epochs per training cycle is insufficient for the model to converge on such a large dataset. ",
634
+ "bbox": [
635
+ 174,
636
+ 448,
637
+ 825,
638
+ 517
639
+ ],
640
+ "page_idx": 6
641
+ },
642
+ {
643
+ "type": "text",
644
+ "text": "Ensemble Size. In some applications, it may be beneficial to vary the size of the ensemble dynamically at test time depending on available resources. Figure 3 displays the performance of DenseNet-40 on the CIFAR-100 dataset as the effective ensemble size, $m$ , is varied. Each ensemble consists of snapshots from later cycles, as these snapshots have received the most training and therefore have likely converged to better minima. Although ensembling more models generally gives better performance, we observe significant drops in error when the second and third models are added to the ensemble. In most cases, an ensemble of two models outperforms the baseline model. ",
645
+ "bbox": [
646
+ 174,
647
+ 521,
648
+ 645,
649
+ 660
650
+ ],
651
+ "page_idx": 6
652
+ },
653
+ {
654
+ "type": "text",
655
+ "text": "Restart Learning Rate. The effect of the restart learning rate can be observed in Figure 3. The left two plots show performance when using a restart learning rate of $\\alpha _ { 0 } = 0 . 1$ at the beginning of each cycle, and the right two plots show $\\alpha _ { 0 } = 0 . 2$ . In most cases, ensembles with the larger restart learning rate perform better, presumably because the strong perturbation in between cycles increases the diversity of local minima. ",
656
+ "bbox": [
657
+ 174,
658
+ 665,
659
+ 645,
660
+ 747
661
+ ],
662
+ "page_idx": 6
663
+ },
664
+ {
665
+ "type": "table",
666
+ "img_path": "images/48aa4335d51dd953254e7bbf9d387c21c47f40da53d5f6997b4942016c9959a2.jpg",
667
+ "table_caption": [
668
+ "Table 3: Error rates of a DenseNet-40 Snapshot Ensemble on CIFAR-100, varying $M .$ —the number of models (cycles) used in the ensemble. "
669
+ ],
670
+ "table_footnote": [],
671
+ "table_body": "<table><tr><td colspan=\"2\">M Test Error (%)</td></tr><tr><td>2</td><td>22.92</td></tr><tr><td>4</td><td>22.07</td></tr><tr><td>6</td><td>21.93</td></tr><tr><td>8</td><td>21.89</td></tr><tr><td>10</td><td>22.16</td></tr></table>",
672
+ "bbox": [
673
+ 665,
674
+ 537,
675
+ 816,
676
+ 630
677
+ ],
678
+ "page_idx": 6
679
+ },
680
+ {
681
+ "type": "text",
682
+ "text": "Varying Number of Cycles. Given a fixed training budget, there is a trade-off between the number of learning rate cycles and their length. Therefore, we investigate how the number of cycles $M$ affects the ensemble performance, given a fixed training budget. We train a 40-layer DenseNet on the CIFAR-100 dataset with an initial learning rate of $\\alpha _ { 0 } = 0 . 2$ . We fix the total training budget $B = 3 0 0$ epochs, and vary the value of $M \\in \\{ 2 , 4 , 6 , 8 , 1 0 \\}$ . As shown in Table 3, our method is relatively robust with respect to different values of $M$ . At the extremes, $M = 2$ and $M = 1 0$ , we find a slight degradation in performance, as the cycles are either too few or too short. In practice, we find that setting $M$ to be $4 \\sim 8$ works reasonably well. ",
683
+ "bbox": [
684
+ 173,
685
+ 752,
686
+ 825,
687
+ 863
688
+ ],
689
+ "page_idx": 6
690
+ },
691
+ {
692
+ "type": "text",
693
+ "text": "Varying Training Budget. The left and middle panels of Figure 4 show the performance of Snapshot Ensembles and SingleCycle Ensembles as a function of training budget (where the number of cycles is fixed at $M = 6$ ). We train a 40-layer DenseNet on CIFAR-10 and CIFAR-100, with an initial learning rate of $\\alpha _ { 0 } = 0 . 1$ , varying the total number of training epochs from 60 to 300. We observe that both Snapshot Ensembles and SingleCycle Ensembles become more accurate as training budget increases. However, we note that as training budget decreases, Snapshot Ensembles still yield competitive results, while the performance of the SingleCycle Ensembles degrades rapidly. These results highlight the improvements that Snapshot Ensembles obtain when the budget is low. If the budget is high, then the SingleCycle baseline approaches true ensembles and outperforms Snapshot ensembles eventually. ",
694
+ "bbox": [
695
+ 174,
696
+ 867,
697
+ 823,
698
+ 924
699
+ ],
700
+ "page_idx": 6
701
+ },
702
+ {
703
+ "type": "image",
704
+ "img_path": "images/566524b31fa444cf21bb9f53d27b9fd61198bb62d5bc3d79c805e92b61609d81.jpg",
705
+ "image_caption": [],
706
+ "image_footnote": [],
707
+ "bbox": [
708
+ 171,
709
+ 98,
710
+ 826,
711
+ 243
712
+ ],
713
+ "page_idx": 7
714
+ },
715
+ {
716
+ "type": "image",
717
+ "img_path": "images/ace68d823d81d79b87bf51385623b321d3d01faffbff255db025ad20176534f5.jpg",
718
+ "image_caption": [
719
+ "Figure 4: Snapshot Ensembles under different training budgets on (Left) CIFAR-10 and (Middle) CIFAR-100. Right: Comparison of Snapshot Ensembles with true ensembles. ",
720
+ "Figure 5: Interpolations in parameter space between the final model (sixth snapshot) and all intermediate snapshots. $\\lambda = 0$ represents an intermediate snapshot model, while $\\lambda = 1$ represents the final model. Left: A Snapshot Ensemble, with cosine annealing cycles $\\mathrm { \\Delta } \\alpha _ { 0 } = 0 . 2$ every $B / M \\stackrel { \\textstyle - } { = } 5 0$ epochs). Right: A NoCycle Snapshot Ensemble, (two learning rate drops, snapshots every 50 epochs). "
721
+ ],
722
+ "image_footnote": [],
723
+ "bbox": [
724
+ 169,
725
+ 281,
726
+ 825,
727
+ 402
728
+ ],
729
+ "page_idx": 7
730
+ },
731
+ {
732
+ "type": "text",
733
+ "text": "",
734
+ "bbox": [
735
+ 174,
736
+ 479,
737
+ 825,
738
+ 564
739
+ ],
740
+ "page_idx": 7
741
+ },
742
+ {
743
+ "type": "text",
744
+ "text": "Comparison with True Ensembles. We compare Snapshot Ensembles with the traditional ensembling method. The right panel of Figure 4 shows the test error rates of DenseNet-40 on CIFAR-100. The true ensemble method averages models that are trained with 300 full epochs, each with different weight initializations. Given the same number of models at test time, the error rate of the true ensemble can be seen as a lower bound of our method. Our method achieves performance that is comparable with ensembling of 2 independent models, but with the training cost of one model. ",
745
+ "bbox": [
746
+ 174,
747
+ 568,
748
+ 825,
749
+ 651
750
+ ],
751
+ "page_idx": 7
752
+ },
753
+ {
754
+ "type": "text",
755
+ "text": "4.4 DIVERSITY OF MODEL ENSEMBLES ",
756
+ "text_level": 1,
757
+ "bbox": [
758
+ 176,
759
+ 671,
760
+ 457,
761
+ 684
762
+ ],
763
+ "page_idx": 7
764
+ },
765
+ {
766
+ "type": "text",
767
+ "text": "Parameter Space. We hypothesize that the cyclic learning rate schedule creates snapshots which are not only accurate but also diverse with respect to model predictions. We qualitatively measure this diversity by visualizing the local minima that models converge to. To do so, we linearly interpolate snapshot models, as described by Goodfellow et al. (2014). Let $J \\left( \\theta \\right)$ be the test error of a model using parameters $\\theta$ . Given $\\theta _ { 1 }$ and $\\theta _ { 2 }$ — the parameters from models 1 and 2 respectively — we can compute the loss for a convex combination of model parameters: $J \\left( \\lambda \\left( \\theta _ { 1 } \\right) + \\left( 1 - \\lambda \\right) ( \\bar { \\theta _ { 2 } } ) \\right)$ , where $\\lambda$ is a mixing coefficient. Setting $\\lambda$ to 1 results in a parameters that are entirely $\\theta _ { 1 }$ while setting $\\lambda$ to 0 gives the parameters $\\theta _ { 2 }$ . By sweeping the values of $\\lambda$ , we can examine a linear slice of the parameter space. Two models that converge to a similar minimum will have smooth parameter interpolations, whereas models that converge to different minima will likely have a non-convex interpolation, with a spike in error when $\\lambda$ is between 0 and 1. ",
768
+ "bbox": [
769
+ 173,
770
+ 694,
771
+ 825,
772
+ 847
773
+ ],
774
+ "page_idx": 7
775
+ },
776
+ {
777
+ "type": "text",
778
+ "text": "Figure 5 displays interpolations between the final model of DenseNet-40 (sixth snapshot) and all intermediate snapshots. The left two plots show Snapshot Ensemble models trained with a cyclic learning rate, while the right two plots show NoCycle Snapshot models. $\\lambda = 0$ represents a model which is entirely snapshot parameters, while $\\lambda = 1$ represents a model which is entirely the parameters of the final model. From this figure, it is clear that there are differences between cyclic and non-cyclic learning rate schedules. Firstly, all of the cyclic snapshots achieve roughly the same error as the final cyclical model, as the error is similar for $\\lambda = 0$ and $\\lambda = 1$ . Additionally, it appears that most snapshots do not lie in the same minimum as the final model. Thus the snapshots are likely to misclassify different samples. Conversely, the first three snapshots achieve much higher error than the final model. This can be observed by the sharp minima around $\\lambda = 1$ , which suggests that mixing in any amount of the snapshot parameters will worsen performance. While the final two snapshots achieve low error, the figures suggests that they lie in the same minimum as the final model, and therefore likely add limited diversity to the ensemble. ",
779
+ "bbox": [
780
+ 174,
781
+ 854,
782
+ 823,
783
+ 924
784
+ ],
785
+ "page_idx": 7
786
+ },
787
+ {
788
+ "type": "text",
789
+ "text": "",
790
+ "bbox": [
791
+ 173,
792
+ 103,
793
+ 825,
794
+ 214
795
+ ],
796
+ "page_idx": 8
797
+ },
798
+ {
799
+ "type": "text",
800
+ "text": "Activation space. To further explore the diversity of models, we compute the pairwise correlation of softmax outputs for every pair of snapshots. Figure 6 displays the average correlation for both cyclic snapshots and non-cyclical snapshots. Firstly, there are large correlations between the last 3 snapshots of the non-cyclic training schedule (right). These snapshots are taken after dropping the learning rate, suggesting that each snapshot has converged to the same minimum. Though there is more diversity amongst the earlier snapshots, these snapshots have much higher error rates and are therefore not ideal for ensembling. Conversely, there is less correlation between all cyclic snapshots (left). Because all snapshots have similar accuracy (as can be seen in Figure 5), these differences in predictions can be exploited to create effective ensembles. ",
801
+ "bbox": [
802
+ 174,
803
+ 219,
804
+ 483,
805
+ 482
806
+ ],
807
+ "page_idx": 8
808
+ },
809
+ {
810
+ "type": "image",
811
+ "img_path": "images/5bfed1b78edee266e038c3f02397a3d98cee181dce976e7a3786fe57b7db3a99.jpg",
812
+ "image_caption": [
813
+ "Figure 6: Pairwise correlation of softmax outputs between any two snapshots for DenseNet-100. Left: A Snapshot Ensemble, with cosine annealing cycles (restart with $\\alpha _ { 0 } ~ = ~ 0 . 2$ every 50 epochs). Right: A NoCycle Snapshot Ensemble, (two learning rate drops, snapshots every 50 epochs). "
814
+ ],
815
+ "image_footnote": [],
816
+ "bbox": [
817
+ 496,
818
+ 234,
819
+ 823,
820
+ 362
821
+ ],
822
+ "page_idx": 8
823
+ },
824
+ {
825
+ "type": "text",
826
+ "text": "5 DISCUSSION ",
827
+ "text_level": 1,
828
+ "bbox": [
829
+ 174,
830
+ 502,
831
+ 310,
832
+ 518
833
+ ],
834
+ "page_idx": 8
835
+ },
836
+ {
837
+ "type": "text",
838
+ "text": "We introduce Snapshot Ensembling, a simple method to obtain ensembles of neural networks without any additional training cost. Our method exploits the ability of SGD to converge to and escape from local minima as the learning rate is lowered, which allows the model to visit several weight assignments that lead to increasingly accurate predictions over the course of training. We harness this power with the cyclical learning rate schedule proposed by Loshchilov & Hutter (2016), saving model snapshots at each point of convergence. We show in several experiments that all snapshots are accurate, yet produce different predictions from one another, and therefore are well suited for test-time ensembles. Ensembles of these snapshots significantly improve the state-of-the-art on CIFAR-10, CIFAR-100 and SVHN. Future work will explore combining Snapshot Ensembles with traditional ensembles. In particular, we will investigate how to balance growing an ensemble with new models (with random initializations) and refining existing models with further training cycles under a fixed training budget. ",
839
+ "bbox": [
840
+ 174,
841
+ 534,
842
+ 825,
843
+ 700
844
+ ],
845
+ "page_idx": 8
846
+ },
847
+ {
848
+ "type": "text",
849
+ "text": "ACKNOWLEDGEMENTS ",
850
+ "text_level": 1,
851
+ "bbox": [
852
+ 176,
853
+ 720,
854
+ 367,
855
+ 736
856
+ ],
857
+ "page_idx": 8
858
+ },
859
+ {
860
+ "type": "text",
861
+ "text": "We thank Ilya Loshchilov and Frank Hutter for their insightful comments on the cyclic cosineshaped learning rate. The authors are supported in part by the, III-1618134, III-1526012, IIS1149882 grants from the National Science Foundation, US Army Research Office W911NF-14- 1-0477, and the Bill and Melinda Gates Foundation. ",
862
+ "bbox": [
863
+ 174,
864
+ 751,
865
+ 823,
866
+ 806
867
+ ],
868
+ "page_idx": 8
869
+ },
870
+ {
871
+ "type": "text",
872
+ "text": "REFERENCES ",
873
+ "text_level": 1,
874
+ "bbox": [
875
+ 174,
876
+ 827,
877
+ 285,
878
+ 842
879
+ ],
880
+ "page_idx": 8
881
+ },
882
+ {
883
+ "type": "text",
884
+ "text": "Leon Bottou. Large-scale machine learning with stochastic gradient descent. In ´ COMPSTAT. 2010. ",
885
+ "bbox": [
886
+ 173,
887
+ 849,
888
+ 821,
889
+ 864
890
+ ],
891
+ "page_idx": 8
892
+ },
893
+ {
894
+ "type": "text",
895
+ "text": "Cristian Bucilu, Rich Caruana, and Alexandru Niculescu-Mizil. Model compression. In KDD, 2006. ",
896
+ "bbox": [
897
+ 173,
898
+ 872,
899
+ 820,
900
+ 887
901
+ ],
902
+ "page_idx": 8
903
+ },
904
+ {
905
+ "type": "text",
906
+ "text": "Rich Caruana, Alexandru Niculescu-Mizil, Geoff Crew, and Alex Ksikes. Ensemble selection from libraries of models. In ICML, 2004. ",
907
+ "bbox": [
908
+ 174,
909
+ 895,
910
+ 823,
911
+ 922
912
+ ],
913
+ "page_idx": 8
914
+ },
915
+ {
916
+ "type": "text",
917
+ "text": "Ronan Collobert, Koray Kavukcuoglu, and Clement Farabet. Torch7: A matlab-like environment ´ for machine learning. In BigLearn, NIPS Workshop, 2011. ",
918
+ "bbox": [
919
+ 171,
920
+ 103,
921
+ 823,
922
+ 132
923
+ ],
924
+ "page_idx": 9
925
+ },
926
+ {
927
+ "type": "text",
928
+ "text": "Yann N Dauphin, Razvan Pascanu, Caglar Gulcehre, Kyunghyun Cho, Surya Ganguli, and Yoshua Bengio. Identifying and attacking the saddle point problem in high-dimensional non-convex optimization. In NIPS, 2014. ",
929
+ "bbox": [
930
+ 176,
931
+ 140,
932
+ 821,
933
+ 183
934
+ ],
935
+ "page_idx": 9
936
+ },
937
+ {
938
+ "type": "text",
939
+ "text": "Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. ",
940
+ "bbox": [
941
+ 171,
942
+ 191,
943
+ 823,
944
+ 220
945
+ ],
946
+ "page_idx": 9
947
+ },
948
+ {
949
+ "type": "text",
950
+ "text": "John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of Machine Learning Research, 12(Jul):2121–2159, 2011. ",
951
+ "bbox": [
952
+ 171,
953
+ 229,
954
+ 823,
955
+ 258
956
+ ],
957
+ "page_idx": 9
958
+ },
959
+ {
960
+ "type": "text",
961
+ "text": "Ian J Goodfellow, David Warde-Farley, Mehdi Mirza, Aaron Courville, and Yoshua Bengio. Maxout networks. In ICML, 2013. ",
962
+ "bbox": [
963
+ 171,
964
+ 267,
965
+ 823,
966
+ 296
967
+ ],
968
+ "page_idx": 9
969
+ },
970
+ {
971
+ "type": "text",
972
+ "text": "Ian J Goodfellow, Oriol Vinyals, and Andrew M Saxe. Qualitatively characterizing neural network optimization problems. arXiv preprint arXiv:1412.6544, 2014. ",
973
+ "bbox": [
974
+ 173,
975
+ 304,
976
+ 823,
977
+ 334
978
+ ],
979
+ "page_idx": 9
980
+ },
981
+ {
982
+ "type": "text",
983
+ "text": "Lars Kai Hansen and Peter Salamon. Neural network ensembles. IEEE transactions on pattern analysis and machine intelligence, 12:993–1001, 1990. ",
984
+ "bbox": [
985
+ 173,
986
+ 342,
987
+ 823,
988
+ 371
989
+ ],
990
+ "page_idx": 9
991
+ },
992
+ {
993
+ "type": "text",
994
+ "text": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016a. ",
995
+ "bbox": [
996
+ 174,
997
+ 378,
998
+ 823,
999
+ 409
1000
+ ],
1001
+ "page_idx": 9
1002
+ },
1003
+ {
1004
+ "type": "text",
1005
+ "text": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In ECCV, 2016b. ",
1006
+ "bbox": [
1007
+ 171,
1008
+ 416,
1009
+ 823,
1010
+ 446
1011
+ ],
1012
+ "page_idx": 9
1013
+ },
1014
+ {
1015
+ "type": "text",
1016
+ "text": "Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. ",
1017
+ "bbox": [
1018
+ 171,
1019
+ 454,
1020
+ 825,
1021
+ 484
1022
+ ],
1023
+ "page_idx": 9
1024
+ },
1025
+ {
1026
+ "type": "text",
1027
+ "text": "Gao Huang, Zhuang Liu, and Kilian Q Weinberger. Densely connected convolutional networks. arXiv preprint arXiv:1608.06993, 2016a. ",
1028
+ "bbox": [
1029
+ 171,
1030
+ 492,
1031
+ 823,
1032
+ 521
1033
+ ],
1034
+ "page_idx": 9
1035
+ },
1036
+ {
1037
+ "type": "text",
1038
+ "text": "Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Weinberger. Deep networks with stochastic depth. In ECCV, 2016b. ",
1039
+ "bbox": [
1040
+ 173,
1041
+ 529,
1042
+ 823,
1043
+ 559
1044
+ ],
1045
+ "page_idx": 9
1046
+ },
1047
+ {
1048
+ "type": "text",
1049
+ "text": "Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICCV, 2015. ",
1050
+ "bbox": [
1051
+ 171,
1052
+ 568,
1053
+ 823,
1054
+ 597
1055
+ ],
1056
+ "page_idx": 9
1057
+ },
1058
+ {
1059
+ "type": "text",
1060
+ "text": "Sbastien Jean, Kyunghyun Cho, Roland Memisevic, and Yoshua Bengio. On using very large target vocabulary for neural machine translation. arXiv preprint arXiv:1412.2007, 2014. ",
1061
+ "bbox": [
1062
+ 171,
1063
+ 604,
1064
+ 821,
1065
+ 633
1066
+ ],
1067
+ "page_idx": 9
1068
+ },
1069
+ {
1070
+ "type": "text",
1071
+ "text": "Kenji Kawaguchi. Deep learning without poor local minima. arXiv preprint arXiv:1605.07110, 2016. ",
1072
+ "bbox": [
1073
+ 171,
1074
+ 642,
1075
+ 823,
1076
+ 671
1077
+ ],
1078
+ "page_idx": 9
1079
+ },
1080
+ {
1081
+ "type": "text",
1082
+ "text": "Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. arXiv preprint arXiv:1609.04836, 2016. ",
1083
+ "bbox": [
1084
+ 176,
1085
+ 679,
1086
+ 823,
1087
+ 723
1088
+ ],
1089
+ "page_idx": 9
1090
+ },
1091
+ {
1092
+ "type": "text",
1093
+ "text": "Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. ",
1094
+ "bbox": [
1095
+ 169,
1096
+ 731,
1097
+ 823,
1098
+ 761
1099
+ ],
1100
+ "page_idx": 9
1101
+ },
1102
+ {
1103
+ "type": "text",
1104
+ "text": "Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. 2009. ",
1105
+ "bbox": [
1106
+ 171,
1107
+ 768,
1108
+ 821,
1109
+ 785
1110
+ ],
1111
+ "page_idx": 9
1112
+ },
1113
+ {
1114
+ "type": "text",
1115
+ "text": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NIPS, 2012. ",
1116
+ "bbox": [
1117
+ 169,
1118
+ 791,
1119
+ 823,
1120
+ 821
1121
+ ],
1122
+ "page_idx": 9
1123
+ },
1124
+ {
1125
+ "type": "text",
1126
+ "text": "Anders Krogh, Jesper Vedelsby, et al. Neural network ensembles, cross validation, and active learning. In NIPS, volume 7, 1995. ",
1127
+ "bbox": [
1128
+ 171,
1129
+ 830,
1130
+ 823,
1131
+ 859
1132
+ ],
1133
+ "page_idx": 9
1134
+ },
1135
+ {
1136
+ "type": "text",
1137
+ "text": "David Krueger, Tegan Maharaj, Janos Kram ´ ar, Mohammad Pezeshki, Nicolas Ballas, Nan Rose- ´ mary Ke, Anirudh Goyal, Yoshua Bengio, Hugo Larochelle, Aaron Courville, et al. Zoneout: Regularizing rnns by randomly preserving hidden activations. arXiv preprint arXiv:1606.01305, 2016. ",
1138
+ "bbox": [
1139
+ 174,
1140
+ 867,
1141
+ 825,
1142
+ 922
1143
+ ],
1144
+ "page_idx": 9
1145
+ },
1146
+ {
1147
+ "type": "text",
1148
+ "text": "Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. ",
1149
+ "bbox": [
1150
+ 173,
1151
+ 103,
1152
+ 823,
1153
+ 132
1154
+ ],
1155
+ "page_idx": 10
1156
+ },
1157
+ {
1158
+ "type": "text",
1159
+ "text": "Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Fractalnet: Ultra-deep neural networks without residuals. arXiv preprint arXiv:1605.07648, 2016. ",
1160
+ "bbox": [
1161
+ 174,
1162
+ 140,
1163
+ 823,
1164
+ 171
1165
+ ],
1166
+ "page_idx": 10
1167
+ },
1168
+ {
1169
+ "type": "text",
1170
+ "text": "Chen-Yu Lee, Saining Xie, Patrick Gallagher, Zhengyou Zhang, and Zhuowen Tu. Deeplysupervised nets. In AISTATS, 2015. ",
1171
+ "bbox": [
1172
+ 174,
1173
+ 178,
1174
+ 821,
1175
+ 208
1176
+ ],
1177
+ "page_idx": 10
1178
+ },
1179
+ {
1180
+ "type": "text",
1181
+ "text": "Min Lin, Qiang Chen, and Shuicheng Yan. Network in network. arXiv preprint arXiv:1312.4400, 2013. ",
1182
+ "bbox": [
1183
+ 173,
1184
+ 215,
1185
+ 825,
1186
+ 246
1187
+ ],
1188
+ "page_idx": 10
1189
+ },
1190
+ {
1191
+ "type": "text",
1192
+ "text": "Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with restarts. arXiv preprint arXiv:1608.03983, 2016. ",
1193
+ "bbox": [
1194
+ 173,
1195
+ 253,
1196
+ 823,
1197
+ 284
1198
+ ],
1199
+ "page_idx": 10
1200
+ },
1201
+ {
1202
+ "type": "text",
1203
+ "text": "Mohammad Moghimi, Mohammad Saberian, Jian Yang, Li-Jia Li, Nuno Vasconcelos, and Serge Belongie. Boosted convolutional neural networks. 2016. ",
1204
+ "bbox": [
1205
+ 173,
1206
+ 291,
1207
+ 825,
1208
+ 321
1209
+ ],
1210
+ "page_idx": 10
1211
+ },
1212
+ {
1213
+ "type": "text",
1214
+ "text": "Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning, 2011. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning, 2011. ",
1215
+ "bbox": [
1216
+ 178,
1217
+ 329,
1218
+ 823,
1219
+ 373
1220
+ ],
1221
+ "page_idx": 10
1222
+ },
1223
+ {
1224
+ "type": "text",
1225
+ "text": "Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014. ",
1226
+ "bbox": [
1227
+ 174,
1228
+ 381,
1229
+ 821,
1230
+ 411
1231
+ ],
1232
+ "page_idx": 10
1233
+ },
1234
+ {
1235
+ "type": "text",
1236
+ "text": "Rico Sennrich, Barry Haddow, and Alexandra Birch. Edinburgh neural machine translation systems for wmt 16. arXiv preprint arXiv:1606.02891, 2016. ",
1237
+ "bbox": [
1238
+ 174,
1239
+ 419,
1240
+ 821,
1241
+ 449
1242
+ ],
1243
+ "page_idx": 10
1244
+ },
1245
+ {
1246
+ "type": "text",
1247
+ "text": "Pierre Sermanet, Soumith Chintala, and Yann LeCun. Convolutional neural networks applied to house numbers digit classification. In ICPR, 2012. ",
1248
+ "bbox": [
1249
+ 173,
1250
+ 457,
1251
+ 821,
1252
+ 486
1253
+ ],
1254
+ "page_idx": 10
1255
+ },
1256
+ {
1257
+ "type": "text",
1258
+ "text": "Saurabh Singh, Derek Hoiem, and David Forsyth. Swapout: Learning an ensemble of deep architectures. arXiv preprint arXiv:1605.06465, 2016. ",
1259
+ "bbox": [
1260
+ 173,
1261
+ 494,
1262
+ 820,
1263
+ 523
1264
+ ],
1265
+ "page_idx": 10
1266
+ },
1267
+ {
1268
+ "type": "text",
1269
+ "text": "Leslie N. Smith. No more pesky learning rate guessing games. CoRR, abs/1506.01186, 2016. URL http://arxiv.org/abs/1506.01186. ",
1270
+ "bbox": [
1271
+ 174,
1272
+ 532,
1273
+ 821,
1274
+ 561
1275
+ ],
1276
+ "page_idx": 10
1277
+ },
1278
+ {
1279
+ "type": "text",
1280
+ "text": "Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806, 2014. ",
1281
+ "bbox": [
1282
+ 173,
1283
+ 569,
1284
+ 820,
1285
+ 599
1286
+ ],
1287
+ "page_idx": 10
1288
+ },
1289
+ {
1290
+ "type": "text",
1291
+ "text": "Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(1):1929–1958, 2014. ",
1292
+ "bbox": [
1293
+ 176,
1294
+ 607,
1295
+ 823,
1296
+ 651
1297
+ ],
1298
+ "page_idx": 10
1299
+ },
1300
+ {
1301
+ "type": "text",
1302
+ "text": "Rupesh Kumar Srivastava, Klaus Greff, and Jurgen Schmidhuber. Highway networks. ¨ arXiv preprint arXiv:1505.00387, 2015. ",
1303
+ "bbox": [
1304
+ 173,
1305
+ 659,
1306
+ 823,
1307
+ 689
1308
+ ],
1309
+ "page_idx": 10
1310
+ },
1311
+ {
1312
+ "type": "text",
1313
+ "text": "A Swann and N Allinson. Fast committee learning: Preliminary results. Electronics Letters, 34(14): 1408–1410, 1998. ",
1314
+ "bbox": [
1315
+ 173,
1316
+ 696,
1317
+ 821,
1318
+ 727
1319
+ ],
1320
+ "page_idx": 10
1321
+ },
1322
+ {
1323
+ "type": "text",
1324
+ "text": "Li Wan, Matthew Zeiler, Sixin Zhang, Yann L Cun, and Rob Fergus. Regularization of neural networks using dropconnect. In ICML, 2013. ",
1325
+ "bbox": [
1326
+ 176,
1327
+ 734,
1328
+ 823,
1329
+ 765
1330
+ ],
1331
+ "page_idx": 10
1332
+ },
1333
+ {
1334
+ "type": "text",
1335
+ "text": "Jingjing Xie, Bing Xu, and Zhang Chuang. Horizontal and vertical ensemble with deep representation for classification. arXiv preprint arXiv:1306.2759, 2013. ",
1336
+ "bbox": [
1337
+ 173,
1338
+ 772,
1339
+ 823,
1340
+ 801
1341
+ ],
1342
+ "page_idx": 10
1343
+ },
1344
+ {
1345
+ "type": "text",
1346
+ "text": "Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. ",
1347
+ "bbox": [
1348
+ 174,
1349
+ 809,
1350
+ 825,
1351
+ 839
1352
+ ],
1353
+ "page_idx": 10
1354
+ },
1355
+ {
1356
+ "type": "text",
1357
+ "text": "SUPPLEMENTARY ",
1358
+ "text_level": 1,
1359
+ "bbox": [
1360
+ 176,
1361
+ 103,
1362
+ 320,
1363
+ 117
1364
+ ],
1365
+ "page_idx": 11
1366
+ },
1367
+ {
1368
+ "type": "text",
1369
+ "text": "A. Single model and Snapshot Ensemble performance over time ",
1370
+ "text_level": 1,
1371
+ "bbox": [
1372
+ 174,
1373
+ 133,
1374
+ 619,
1375
+ 148
1376
+ ],
1377
+ "page_idx": 11
1378
+ },
1379
+ {
1380
+ "type": "text",
1381
+ "text": "In Figures 7-9, we compare the test error of Snapshot Ensembles with the error of individual model snapshots. The blue curve shows the test error of a single model snapshot using a cyclic cosine learning rate. The green curve shows the test error when ensembling model snapshots over time. (Note that, unlike Figure 3, we construct these ensembles beginning with the earliest snapshots.) As a reference, the red dashed line in each panel represents the test error of single model trained for 300 epochs using a standard learning rate schedule. Without Snapshot Ensembles, in about half of the cases, the test error of final model using a cyclic learning rate—the right most point in the blue curve—is no better than using a standard learning rate schedule. ",
1382
+ "bbox": [
1383
+ 173,
1384
+ 155,
1385
+ 825,
1386
+ 266
1387
+ ],
1388
+ "page_idx": 11
1389
+ },
1390
+ {
1391
+ "type": "text",
1392
+ "text": "One can observe that under almost all settings, complete Snapshot Ensembles—the right most points of the green curves—outperform the single model baselines. In many cases, ensembles of just 2 or 3 model snapshots are able to match the performance of the single model trained with a standard learning rate. Not surprisingly, the ensembles of model snapshots consistently outperform any of its members, yielding a smooth curve of test error over time. ",
1393
+ "bbox": [
1394
+ 174,
1395
+ 273,
1396
+ 825,
1397
+ 343
1398
+ ],
1399
+ "page_idx": 11
1400
+ },
1401
+ {
1402
+ "type": "image",
1403
+ "img_path": "images/492226a2f04cafb128e77b6bb42f1568ab5ee2737c3d10eb5641d1478d3d2cf7.jpg",
1404
+ "image_caption": [
1405
+ "Figure 7: Single model and Snapshot Ensemble performance over time (part 1). "
1406
+ ],
1407
+ "image_footnote": [],
1408
+ "bbox": [
1409
+ 228,
1410
+ 388,
1411
+ 767,
1412
+ 872
1413
+ ],
1414
+ "page_idx": 11
1415
+ },
1416
+ {
1417
+ "type": "image",
1418
+ "img_path": "images/b2db0a7558184db89e197051f6ece7697c6a617925b9be55972e0134e1ce7226.jpg",
1419
+ "image_caption": [
1420
+ "Figure 8: Single model and Snapshot Ensemble performance over time (part 2). "
1421
+ ],
1422
+ "image_footnote": [],
1423
+ "bbox": [
1424
+ 230,
1425
+ 270,
1426
+ 767,
1427
+ 752
1428
+ ],
1429
+ "page_idx": 12
1430
+ },
1431
+ {
1432
+ "type": "image",
1433
+ "img_path": "images/1119911817b4c7e1d7782e2e076880f8d39c5011c289884028bfb86ba527e0a3.jpg",
1434
+ "image_caption": [
1435
+ "Figure 9: Single model and Snapshot Ensemble performance over time (part 3). "
1436
+ ],
1437
+ "image_footnote": [],
1438
+ "bbox": [
1439
+ 230,
1440
+ 271,
1441
+ 769,
1442
+ 661
1443
+ ],
1444
+ "page_idx": 13
1445
+ }
1446
+ ]
parse/train/BJYwwY9ll/BJYwwY9ll_middle.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/BJYwwY9ll/BJYwwY9ll_model.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/BJgr4kSFDS/BJgr4kSFDS.md ADDED
@@ -0,0 +1,385 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # QUERY2BOX: REASONING OVER KNOWLEDGEGRAPHS IN VECTOR SPACE USING BOX EMBEDDINGS
2
+
3
+ Hongyu Ren∗, Weihua $\mathbf { H u } ^ { * }$ , Jure Leskovec Department of Computer Science, Stanford University {hyren,weihuahu,jure}@cs.stanford.edu
4
+
5
+ # ABSTRACT
6
+
7
+ Answering complex logical queries on large-scale incomplete knowledge graphs (KGs) is a fundamental yet challenging task. Recently, a promising approach to this problem has been to embed KG entities as well as the query into a vector space such that entities that answer the query are embedded close to the query. However, prior work models queries as single points in the vector space, which is problematic because a complex query represents a potentially large set of its answer entities, but it is unclear how such a set can be represented as a single point. Furthermore, prior work can only handle queries that use conjunctions $( \wedge )$ and existential quantifiers (∃). Handling queries with logical disjunctions (∨) remains an open problem. Here we propose QUERY2BOX, an embedding-based framework for reasoning over arbitrary queries with $\wedge , \vee$ , and ∃ operators in massive and incomplete KGs. Our main insight is that queries can be embedded as boxes (i.e., hyper-rectangles), where a set of points inside the box corresponds to a set of answer entities of the query. We show that conjunctions can be naturally represented as intersections of boxes and also prove a negative result that handling disjunctions would require embedding with dimension proportional to the number of KG entities. However, we show that by transforming queries into a Disjunctive Normal Form, QUERY2BOX is capable of handling arbitrary logical queries with ∧, ∨, ∃ in a scalable manner. We demonstrate the effectiveness of QUERY2BOX on three large KGs and show that QUERY2BOX achieves up to $2 5 \%$ relative improvement over the state of the art.
8
+
9
+ # 1 INTRODUCTION
10
+
11
+ Knowledge graphs (KGs) capture different types of relationships between entities, e.g., Canada citizen−−−−→ Hinton. Answering arbitrary logical queries, such as “where did Canadian citizens with Turing Award graduate?”, over such KGs is a fundamental task in question answering, knowledge base reasoning, as well as AI more broadly.
12
+
13
+ First-order logical queries can be represented as Directed Acyclic Graphs (DAGs) (Fig. 1(A)) and be reasoned according to the DAGs to obtain a set of answers (Fig. 1(C)). While simple and intuitive, such approach has many drawbacks: (1) Computational complexity of subgraph matching is exponential in the query size, and thus cannot scale to modern KGs; (2) Subgraph matching is very sensitive as it cannot correctly answer queries with missing relations. To remedy (2) one could impute missing relations (Koller et al., 2007; Džeroski, 2009; De Raedt, 2008; Nickel et al., 2016) but that would only make the KG denser, which would further exacerbate issue (1) (Dalvi & Suciu, 2007; Krompaß et al., 2014).
14
+
15
+ Recently, a promising alternative approach has emerged, where logical queries as well as KG entities are embedded into a low-dimensional vector space such that entities that answer the query are embedded close to the query (Guu et al., 2015; Hamilton et al., 2018; Das et al., 2017). Such approach robustly handles missing relations (Hamilton et al., 2018) and is also orders of magnitude faster, as answering an arbitrary logical query is reduced to simply identifying entities nearest to the embedding of the query in the vector space.
16
+
17
+ ![](images/7bf7f77dd269885d1b5032b0dec09237cf01a6d5d84bde0dd0c77dea4edd5f59.jpg)
18
+ Figure 1: Query2Box reasoning framework. (A) A given conjunctive query “Where did Canadian citizens with Turing Award graduate?” can be represented with a dependency graph. (B) Computation graph specifies the reasoning procedure to obtain a set of answers for the query in (A). (C) Example knowledge graph, where green nodes/entities denote answers to the query. Bold arrows indicate subgraphs that match the query graph in (A). (D) In QUERY2BOX, nodes of the KG are embedded as points in the vector space. We then obtain query embedding according to the computation graph (B) as a sequence of box operations: start with two nodes TuringAward and Canada and apply Win and Citizen projection operators, followed by an intersection operator (denoted as a shaded intersection of yellow and orange boxes) and another projection operator. The final embedding of the query is a green box and query’s answers are the entities inside the box.
19
+
20
+ However, prior work embeds a query into a single point in the vector space. This is problematic because answering a logical query requires modeling a set of active entities while traversing the KG (Fig. 1(C)), and how to effectively model a set with a single point is unclear. Furthermore, it is also unnatural to define logical operators (e.g., set intersection) of two points in the vector space. Another fundamental limitation of prior work is that it can only handle conjunctive queries, a subset of first-order logic that only involves conjunction $( \wedge )$ and existential quantifier (∃), but not disjunction $( \vee )$ . It remains an open question how to handle disjunction effectively in the vector space.
21
+
22
+ Here we present QUERY2BOX, an embedding-based framework for reasoning over KGs that is capable of handling arbitrary Existential Positive First-order (EPFO) logical queries (i.e., queries that include any set of $\wedge , \vee$ , and ∃) in a scalable manner. First, to accurately model a set of entities, our key idea is to use a closed region rather than a single point in the vector space. Specifically, we use a box (axis-aligned hyper-rectangle) to represent a query (Fig. 1(D)). This provides three important benefits: (1) Boxes naturally model sets of entities they enclose; (2) Logical operators (e.g., set intersection) can naturally be defined over boxes similarly as in Venn diagrams (Venn, 1880); (3) Executing logical operators over boxes results in new boxes, which means that the operations are closed; thus, logical reasoning can be efficiently performed in QUERY2BOX by iteratively updating boxes according to the query computation graph (Fig. 1(B)(D)).
23
+
24
+ We show that QUERY2BOX can naturally handle conjunctive queries. We first prove a negative result that embedding EPFO queries to only single points or boxes is intractable as it would require embedding dimension proportional to the number of KG entities. However, we provide an elegant solution, where we transform a given EPFO logical query into a Disjunctive Normal Form (DNF) (Davey & Priestley, 2002), i.e., disjunction of conjunctive queries. Given any EPFO query, QUERY2BOX represents it as a set of individual boxes, where each box is obtained for each conjunctive query in the DNF. We then return nearest neighbor entities to any of the boxes as the answers to the query. This means that to answer any EPFO query we first answer individual conjunctive queries and then take the union of the answer entities.
25
+
26
+ We evaluate QUERY2BOX on three standard KG benchmarks and show: (1) QUERY2BOX provides strong generalization as it can answer complex queries; (2) QUERY2BOX can generalize to new logical query structures that it has never seen during training; (3) QUERY2BOX is able to implicitly impute missing relations as it can answer any EPFO query with high accuracy even when relations involving answering the query are missing in the KG; (4) QUERY2BOX provides up to $2 5 \%$ relative improvement in accuracy of answering EPFO queries over state-of-the-art baselines.
27
+
28
+ # 2 FURTHER RELATED WORK
29
+
30
+ Most related to our work are embedding approaches for multi-hop reasoning over KGs (Bordes et al., 2013; Das et al., 2017; Guu et al., 2015; Hamilton et al., 2018). Crucial difference is that we provide a way to tractably handle a larger subset of the first-order logic (EPFO queries vs. conjunctive queries) and that we embed queries as boxes, which provides better accuracy and generalization.
31
+
32
+ Second line of related work is on structured embeddings, which associate images, words, sentences, or knowledge base concepts with geometric objects such as regions (Erk, 2009; Vilnis et al., 2018; Li et al., 2019), densities (Vilnis & McCallum, 2014; He et al., 2015; Athiwaratkun & Wilson, 2018), and orderings (Vendrov et al., 2016; Lai & Hockenmaier, 2017; Li et al., 2017). While the above work uses geometric objects to model individual entities and their pairwise relations, we use the geometric objects to model sets of entities and reason over those sets. In this sense our work is also related to classical Venn Diagrams (Venn, 1880), where boxes are essentially the Venn Diagrams in vector space, but our boxes and entity embeddings are jointly learned, which allows us to reason over incomplete KGs.
33
+
34
+ Box embeddings have also been used to model hierarchical nature of concepts in an ontology with uncertainty (Vilnis et al., 2018; Li et al., 2019). While our work is also based on box embeddings we employ them for logical reasoning in massive heterogeneous knowledge graphs.
35
+
36
+ # 3 QUERY2BOX: LOGICAL REASONING OVER KGS IN VECTOR SPACE
37
+
38
+ Here we present the QUERY2BOX, where we will define an objective function that allows us to learn embeddings of entities in the KG, and at the same time also learn parameterized geometric logical operators over boxes. Then given an arbitrary EPFO query $q$ (Fig. 1(A)), we will identify its computation graph (Fig. 1(B)), and embed the query by executing a set of geometric operators over boxes (Fig. 1(D)). Entities that are enclosed in the final box embedding are returned as answers to the query (Fig. 1(D)).
39
+
40
+ In order to train our system, we generate a set of queries together with their answers at training time and then learn entity embeddings and geometric operators such that queries can be accurately answered. We show in the following sections that our approach is able to generalize to queries and logical structures never seen during training. Furthermore, as we show in experiments, our approach is able to implicitly impute missing relations and answer queries that would be impossible to answer with traditional graph traversal methods.
41
+
42
+ In the following we first only consider conjunctive queries (conjunction and existential operator) and then we extend our method to also include disjunction.
43
+
44
+ # 3.1 KNOWLEDGE GRAPHS AND CONJUNCTIVE QUERIES
45
+
46
+ We denote a KG as $\mathcal { G } = ( \nu , \mathcal { R } )$ , where $v \in \mathcal V$ represents an entity, and $r \in \mathcal { R }$ is a binary function $r : \mathcal { V } \times \mathcal { V } \{ \mathrm { T r u e , F a l s e } \}$ , indicating whether the relation $r$ holds between a pair of entities or not. In the KG, such binary output indicates the existence of the directed edge between a pair of entities, i.e., $v \stackrel { r } { } v ^ { \prime }$ iff $r ( v , v ^ { \prime } ) = \mathrm { T r u e }$ .
47
+
48
+ Conjunctive queries are a subclass of the first-order logical queries that use existential $\textcircled{1}$ and conjunction $( \wedge )$ operations. They are formally defined as follows.
49
+
50
+ $$
51
+ \begin{array} { r l } & { q [ V _ { ? } ] = V _ { ? } \cdot \exists V _ { 1 } , \dotsc , V _ { k } : e _ { 1 } \wedge e _ { 2 } \wedge \dotsc \wedge e _ { n } , } \\ & { \mathrm { w h e r e ~ } e _ { i } = r ( v _ { a } , V ) , V \in \{ V _ { ? } , V _ { 1 } , \dotsc , V _ { k } \} , v _ { a } \in \mathcal { V } , r \in \mathcal { R } , } \\ & { \qquad \mathrm { o r ~ } e _ { i } = r ( V , V ^ { \prime } ) , V , V ^ { \prime } \in \{ V _ { ? } , V _ { 1 } , \dotsc , V _ { k } \} , V \neq V ^ { \prime } , r \in \mathcal { R } , } \end{array}
52
+ $$
53
+
54
+ where $v _ { a }$ represents non-variable anchor entity, $V _ { 1 } , \ldots , V _ { k }$ are existentially quantified bound variables, $V _ { ? }$ is the target variable. The goal of answering the logical query $q$ is to find a set of entities $[ [ q ] ] \subseteq \nu$ such that $v \in [ [ q ] ]$ iff $q [ v ] = { \mathrm { T r u e } }$ . We call $[ [ q ] ]$ the denotation set (i.e., answer set) of query $q$ .
55
+
56
+ As shown in Fig. 1(A), the dependency graph is a graphical representation of conjunctive query $q$ where nodes correspond to variable or non-variable entities in $q$ and edges correspond to relations in $q$ . In order for the query to be valid, the corresponding dependency graph needs to be a Directed Acyclic Graph (DAG), with the anchor entities as the source nodes of the DAG and the query target $V _ { ? }$ as the unique sink node (Hamilton et al., 2018).
57
+
58
+ From the dependency graph of query $q$ , one can also derive the computation graph, which consists of two types of directed edges that represent operators over sets of entities:
59
+
60
+ • Projection: Given a set of entities $S \subseteq \nu$ , and relation $r \in \mathcal { R }$ , this operator obtains $\cup _ { v \in S } A _ { r } ( v )$ , where $A _ { r } ( v ) \equiv \{ v ^ { \prime } \in \mathcal { V } : \ r ( v , v ^ { \prime } ) = \mathrm { T r u e } \}$ . • Intersection: Given a set of entity sets $\{ S _ { 1 } , S _ { 2 } , \ldots , S _ { n } \}$ , this operator obtains $\cap _ { i = 1 } ^ { n } S _ { i }$ .
61
+
62
+ For a given query $q$ , the computation graph specifies the procedure of reasoning to obtain a set of answer entities, i.e., starting from a set of anchor nodes, the above two operators are applied iteratively until the unique sink target node is reached. The entire procedure is analogous to traversing KGs following the computation graph (Guu et al., 2015).
63
+
64
+ # 3.2 REASONING OVER SETS OF ENTITIES USING BOX EMBEDDINGS
65
+
66
+ So far we have defined conjunctive queries as computation graphs that can be executed directly over the nodes and edges in the KG. Now, we define logical reasoning in the vector space. Our intuition follows Fig. 1: Given a complex query, we shall decompose it into a sequence of logical operations, and then execute these operations in the vector space. This way we will obtain the embedding of the query, and answers to the query will be entities that are enclosed in the final query embedding box.
67
+
68
+ In the following, we detail our two methodological advances: (1) the use of box embeddings to efficiently model and reason over sets of entities in the vector space, and (2) how to tractably handle disjunction operator (∨), expanding the class of first-order logic that can be modeled in the vector space (Section 3.3).
69
+
70
+ Box embeddings. To efficiently model a set of entities in the vector space, we use boxes (i.e., axis-aligned hyper-rectangles). The benefit is that unlike a single point, the box has the interior; thus, if an entity is in a set, it is natural to model the entity embedding to be a point inside the box. Formally, we operate on $\mathbb { R } ^ { d }$ , and define a box in $\mathbb { R } ^ { d }$ by $\mathbf { p } \overset { \cdot } { = } ( \mathbf { C e n } ( \mathbf { p } ) , \mathbf { \bar { O } f f } ( \mathbf { p } ) ) \overset { \cdot } { \in } \mathbb { R } ^ { 2 d }$ as:
71
+
72
+ $$
73
+ \operatorname { B o x } _ { \mathbf { p } } \equiv \{ \mathbf { v } \in \mathbb { R } ^ { d } : \operatorname { C e n } ( \mathbf { p } ) - \operatorname { O f f } ( \mathbf { p } ) \preceq \mathbf { v } \preceq \operatorname { C e n } ( \mathbf { p } ) + \operatorname { O f f } ( \mathbf { p } ) \} ,
74
+ $$
75
+
76
+ where $\preceq$ is element-wise inequality, $\mathbf { C e n } ( \mathbf { p } ) \in \mathbb { R } ^ { d }$ is the center of the box, and $\mathrm { O f f } ( \mathbf { p } ) \in \mathbb { R } _ { \geq 0 } ^ { d }$ is the positive offset of the box, modeling the size of the box. Each entity $v \in \mathcal V$ in KG is assigned a single vector $\mathbf { v } \in \mathbb { R } ^ { d }$ (i.e., a zero-size box), and the box embedding $\mathbf { p }$ models $\{ v \in \mathcal { V } : \mathbf { v } \in \mathrm { B o x } _ { \mathbf { p } } \}$ , i.e., a set of entities whose vectors are inside the box. For the rest of the paper, we use the bold face to denote the embedding, e.g., embedding of $v$ is denoted by $\mathbf { v }$ .
77
+
78
+ Our framework reasons over KGs in the vector space following the computation graph of the query, as shown in Fig. 1(D): we start from the initial box embeddings of the source nodes (anchor entities) and sequentially update the embeddings according to the logical operators. Below, we describe how we set initial box embeddings for the source nodes, as well as how we model projection and intersection operators (defined in Sec. 3.1) as geometric operators that operate over boxes. After that, we describe our entity-to-box distance function and the overall objective that learns embeddings as well as the geometric operators.
79
+
80
+ Initial boxes for source nodes. Each source node represents an anchor entity $v \in \mathcal V$ , which we can regard as a set that only contains the single entity. Such a single-element set can be naturally modeled by a box of size/offset zero centered at $\mathbf { v }$ . Formally, we set the initial box embedding as $( \mathbf { v } , \mathbf { 0 } )$ , where $\mathbf { v } \in \mathbb { R } ^ { d }$ is the anchor entity vector and 0 is a $d$ -dimensional all-zero vector.
81
+
82
+ Geometric projection operator. We associate each relation $r \in \mathcal { R }$ with relation embedding $\mathbf { r } =$ $( \mathbf { C e n } ( \mathbf { r } ) , \mathbf { O f f } ( \mathbf { r } ) ) \in \mathbb { R } ^ { 2 d }$ with $\mathrm { O f f } ( \mathbf { r } ) \succeq \mathbf { 0 }$ . Given an input box embedding p, we model the projection by $\mathbf { p } + \mathbf { r }$ , where we sum the centers and sum the offsets. This gives us a new box with the translated center and larger offset because $\mathrm { O f f } ( \mathbf { r } ) \succeq \mathbf { 0 }$ , as illustrated in Fig. 2(A). The adaptive box size effectively models a different number of entities/vectors in the set.
83
+
84
+ Geometric intersection operator. We model the intersection of a set of box embeddings $\{ \mathbf { p _ { 1 } } , \dotsc , \mathbf { p _ { n } } \}$ as $\mathbf { p } _ { \mathrm { i n t e r } } = ( \mathbf { C e n } ( \mathbf { p } _ { \mathrm { i n t e r } } ) , \mathbf { O f f } ( \mathbf { p } _ { \mathrm { i n t e r } } ) )$ , which is calculated by performing attention over the box centers (Bahdanau et al., 2015) and shrinking the box offset using the sigmoid function:
85
+
86
+ ![](images/d966764f055bb0ba60e5acdb07162da730ed5bf426089eff4bbc7892e6413e55.jpg)
87
+ Figure 2: The geometric intuition of the two operations and distance function in QUERY2BOX. (A) Projection generates a larger box with a translated center. $\mathbf { ( B ) }$ Intersection generates a smaller box lying inside the given set of boxes. (C) Distance $\mathrm { d i s t _ { b o x } }$ is the weighted sum of $\mathrm { d i s t _ { o u t s i d e } }$ and $\mathrm { d i s t } _ { \mathrm { i n s i d e } }$ , where the latter is weighted less.
88
+
89
+ $$
90
+ \begin{array} { r l } & { \displaystyle \mathrm { C e n } ( { \bf p } _ { \mathrm { i n t e r } } ) = \sum _ { i } { \bf a } _ { i } \odot \mathrm { C e n } ( { \bf p } _ { \mathrm { i } } ) , ~ { \bf a } _ { i } = \frac { \exp ( \mathrm { M L P } ( { \bf p } _ { \mathrm { i } } ) ) } { \sum _ { j } \exp ( \mathrm { M L P } ( { \bf p } _ { \mathrm { j } } ) ) } , } \\ & { \displaystyle \mathrm { O f f } ( { \bf p } _ { \mathrm { i n t e r } } ) = \mathrm { M i n } ( \{ \mathrm { O f f } ( { \bf p } _ { 1 } ) , \dots , \mathrm { O f f } ( { \bf p } _ { \mathrm { n } } ) \} ) \odot \sigma ( \mathrm { D e e p S e t s } ( \{ { \bf p } _ { 1 } , \dots , { \bf p } _ { \mathrm { n } } \} ) ) , } \end{array}
91
+ $$
92
+
93
+ where $\odot$ is the dimension-wise product, $\mathrm { M L P } ( \cdot ) : \mathbb { R } ^ { 2 d } \to \mathbb { R } ^ { d }$ is the Multi-Layer Perceptron, $\sigma ( \cdot )$ is the sigmoid function, $\mathrm { D e e p S e t s } ( \cdot )$ is the permutation-invariant deep architecture (Zaheer et al., 2017), and both $\mathrm { M i n } ( \cdot )$ and $\exp ( \cdot )$ are applied in a dimension-wise manner. Following Hamilton et al. (2018), we model all the deep sets by $\begin{array} { r } { \mathrm { D e e p S e t s } ( \{ \mathbf { x _ { 1 } } , . . . , \mathbf { x _ { N } } \} ) = \mathrm { M L P } ( ( 1 / N ) \cdot \sum _ { i = 1 } ^ { N } \mathrm { M L P } ( \mathbf { x _ { i } } ) ) , } \end{array}$ where all the hidden dimensionalities of the two MLPs are the same as the input dimensionality. The intuition behind our geometric intersection is to generate a smaller box that lies inside a set of boxes, as illustrated in Fig. 2(B).1 Different from the generic deep sets to model the intersection (Hamilton et al., 2018), our geometric intersection operator effectively constrains the center position and models the shrinking set size.
94
+
95
+ Entity-to-box distance. Given a query box $ { \mathbf { q } } \in \mathbb { R } ^ { 2 d }$ and an entity vector $\mathbf { v } \in \mathbb { R } ^ { d }$ , we define their distance as
96
+
97
+ $$
98
+ \mathrm { d i s t _ { b o x } ( { \bf v } ; { \bf q } ) } = \mathrm { d i s t _ { o u t s i d e } ( { \bf v } ; { \bf q } ) } + \alpha \cdot \mathrm { d i s t _ { i n s i d e } ( { \bf v } ; { \bf q } ) } ,
99
+ $$
100
+
101
+ where $\mathbf { q } _ { \mathrm { m a x } } = \mathbf { C e n } ( \mathbf { q } ) + \mathbf { O f f } ( \mathbf { q } ) \in \mathbb { R } ^ { d }$ , $\mathbf { q } _ { \mathrm { m i n } } = \mathbf { C e n } ( \mathbf { q } ) - \mathbf { O f f } ( \mathbf { q } ) \in \mathbb { R } ^ { d }$ and $0 < \alpha < 1$ is a fixed scalar, and
102
+
103
+ $$
104
+ \begin{array} { r l } & { \mathrm { d i s t } _ { \mathrm { o u t s i d e } } ( \mathbf { v } ; \mathbf { q } ) = \| \mathrm { M a x } ( \mathbf { v } - \mathbf { q } _ { \mathrm { m a x } } , \mathbf { 0 } ) + \mathrm { M a x } ( \mathbf { q } _ { \mathrm { m i n } } - \mathbf { v } , \mathbf { 0 } ) \| _ { 1 } , } \\ & { \mathrm { d i s t } _ { \mathrm { i n s i d e } } ( \mathbf { v } ; \mathbf { q } ) = \| \mathrm { C e n } ( \mathbf { q } ) - \mathrm { M i n } ( \mathbf { q } _ { \mathrm { m a x } } , \mathrm { M a x } ( \mathbf { q } _ { \mathrm { m i n } } , \mathbf { v } ) ) \| _ { 1 } . } \end{array}
105
+ $$
106
+
107
+ As illustrated in Fig. 2(C), $\mathrm { d i s t _ { o u t s i d e } }$ corresponds to the distance between the entity and closest corner/side of the box. Analogously, $\mathrm { d i s t } _ { \mathrm { i n s i d e } }$ corresponds to the distance between the center of the box and its side/corner (or the entity itself if the entity is inside the box).
108
+
109
+ The key here is to downweight the distance inside the box by using $0 < \alpha < 1$ . This means that as long as entity vectors are inside the box, we regard them as “close enough” to the query center $( i . e . , \mathrm { d i s t _ { o u t s i d e } }$ is 0, and $\mathrm { d i s t } _ { \mathrm { i n s i d e } }$ is scaled by $\alpha$ ). When $\alpha = 1$ , $\mathrm { d i s t } _ { \mathrm { b o x } }$ reduces to the ordinary $L _ { 1 }$ distance, i.e., $\| \mathbf { C e n } ( \mathbf { q } ) - \mathbf { v } \| _ { 1 }$ , which is used by the conventional TransE (Bordes et al., 2013) as well as prior query embedding methods (Guu et al., 2015; Hamilton et al., 2018).
110
+
111
+ Training objective. Our next goal is to learn entity embeddings as well as geometric projection and intersection operators.
112
+
113
+ Given a training set of queries and their answers, we optimize a negative sampling loss (Mikolov et al., 2013) to effectively optimize our distance-based model (Sun et al., 2019):
114
+
115
+ $$
116
+ L = - \log \sigma \left( \gamma - \mathrm { d i s t } _ { \mathrm { b o x } } ( \mathbf { v } ; \mathbf { q } ) \right) - \sum _ { i = 1 } ^ { k } \frac { 1 } { k } \log \sigma \left( \mathrm { d i s t } _ { \mathrm { b o x } } ( \mathbf { v } _ { \mathbf { i } } ^ { \prime } ; \mathbf { q } ) - \gamma \right) ,
117
+ $$
118
+
119
+ where $\gamma$ represents a fixed scalar margin, $v \in \mathbb { I } q \mathbb { I }$ is a positive entity (i.e., answer to the query $q$ ), and $v _ { i } ^ { \prime } \notin \ [ q ]$ is the $i \cdot$ J K-th negative entity (non-answer to the query $q$ ) and $k$ is the number of negative entities.
120
+
121
+ 3.3 TRACTABLE HANDLING OF DISJUNCTION USING DISJUNCTIVE NORMAL FORM
122
+
123
+ So far we have focused on conjunctive queries, and our aim here is to tractably handle in the vector space a wider class of logical queries, called Existential Positive First-order (EPFO) queries (Dalvi & Suciu, 2012) that involve $\vee$ in addition to $\exists$ and $\wedge$ . We specifically focus on EPFO queries whose computation graphs are a DAG, same as that of conjunctive queries (Section 3.1), except that we now have an additional type of directed edge, called union defined as follows:
124
+
125
+ • Union: Given a set of entity sets $\{ S _ { 1 } , S _ { 2 } , \ldots , S _ { n } \}$ , this operator obtains $\cup _ { i = 1 } ^ { n } S _ { i }$
126
+
127
+ A straightforward approach here would be to define another geometric operator for union and embed the query as we did in the previous sections. An immediate challenge for our box embeddings is that boxes can be located anywhere in the vector space, so their union would no longer be a simple box. In other words, union operation over boxes is not closed.
128
+
129
+ Theoretically, we prove a general negative result that holds for any embedding-based method that embeds query $q$ into $\mathbf { q }$ and uses some distance function to retrieve entities, i.e., $\mathrm { d i s t } ( \mathbf { v } ; \mathbf { q } ) \le \beta$ iff $v \in [ [ q ] ]$ . Here, $\operatorname { d i s t } ( \mathbf { v } ; \mathbf { q } )$ is the distance between entity and query embeddings, e.g., $\mathrm { d i s t } _ { \mathrm { b o x } } ( \mathbf { v } ; \mathbf { q } )$ or $\| \mathbf { v } - \mathbf { q } \| _ { 1 }$ , and $\beta$ is a fixed threshold.
130
+
131
+ Theorem 1. Consider any $M$ conjunctive queries $q _ { 1 } , \ldots , q _ { M }$ whose denotation sets $[ [ q _ { 1 } ] ] , \dots , [ [ q _ { M } ] ]$ are disjoint with each other, $\forall \ : i \neq j$ , $[ [ q _ { i } ] ] \cap [ [ q _ { j } ] ] = \emptyset$ . Let $D$ be the $V C$ J K J K dimension of the function class $\{ \operatorname { s i g n } ( \beta - \mathrm { d i s t } ( \cdot ; \mathbf { q } ) ) : \mathbf { q } \in \Xi \}$ J K, where $\Xi$ Krepresents the query embedding space and $\mathrm { s i g n } ( \cdot )$ is the sign function. Then, we need $D \geq M$ to model any EPFO query, i.e., $\operatorname { d i s t } ( \mathbf { v } ; \mathbf { q } ) \leq \beta \Leftrightarrow v \in [ [ q ] ]$ is satisfied for every EPFO query $q$ .
132
+
133
+ The proof is provided in Appendix A, where the key is that with the introduction of the union operation any subset of denotation sets can be the answer, which forces us to model the powerset $\bigl \{ \mathsf { U } _ { q _ { i } \in S } \bigl [ \| q _ { i } \| : \mathsf { S } \subseteq \{ q _ { 1 } , \ldots , q _ { M } \} \bigr \}$ in a vector space.
134
+
135
+ For a real-world KG, there are $M \approx | \nu |$ conjunctive queries with non-overlapping answers. For example, in the commonly-used FB15k dataset (Bordes et al., 2013), derived from the Freebase (Bollacker et al., 2008), we find $M = 1 3 { , } 3 6 5$ , while $| \nu |$ is 14,951 (see Appendix B for the details).
136
+
137
+ Theorem 1 shows that in order to accurately model any EPFO query with the existing framework, the complexity of the distance function measured by the VC dimension needs to be as large as the number of KG entities. This implies that if we use common distance functions based on hyper-plane, Euclidean sphere, or axis-aligned rectangle,2 their parameter dimensionality needs to be $\Theta ( M )$ , which is $\Theta ( \bar { | \nu | } )$ for real KGs we are interested in. In other words, the dimensionality of the logical query embeddings needs to be $\Theta ( | \nu | )$ , which is not low-dimensional; thus not scalable to large KGs and not generalizable in the presence of unobserved KG edges.
138
+
139
+ To rectify this issue, our key idea is to transform a given EPFO query into a Disjunctive Normal Form (DNF) (Davey & Priestley, 2002), i.e., disjunction of conjunctive queries, so that union operation only appears in the last step. Each of the conjunctive queries can then be reasoned in the low-dimensional space, after which we can aggregate the results by a simple and intuitive procedure. In the following, we describe the transformation to DNF and the aggregation procedure.
140
+
141
+ Transformation to DNF. Any first-order logic can be transformed into the equivalent DNF (Davey & Priestley, 2002). We perform such transformation directly in the space of computation graph, i.e., moving all the edges of type “union” to the last step of the computation graph. Let $G _ { q } = ( V _ { q } , E _ { q } )$ be the computation graph for a given EPFO query $q$ , and let $V _ { \mathrm { u n i o n } } \subset V _ { q }$ be a set of nodes whose in-coming edges are of type “union”. For each $v \in V _ { \mathrm { u n i o n } }$ , define $P _ { v } \subset V _ { q }$ as a set of its parent nodes. We first generate N = Qv∈Vunion different computation graphs $\bar { G } _ { q ^ { ( 1 ) } } , \dots , G _ { q ^ { ( N ) } }$ as follows, each with different choices of $v _ { \mathrm { p a r e n t } }$ in the first step.
142
+
143
+ 1. For every $v \in V _ { \mathrm { u n i o n } }$ , select one parent node $v _ { \mathrm { p a r e n t } } \in P _ { v }$ .
144
+
145
+ ![](images/84688316f065b743a9eb9d02d333e7ed60bc9285055f06f42902dc536bf42148.jpg)
146
+ Figure 3: Illustration of converting a computation graph of an EPFO query into an equivalent computation graph of the Disjunctive Normal Form.
147
+
148
+ 2. Remove all the edges of type ‘union.’
149
+
150
+ 3. Merge $v$ and $v _ { \mathrm { p a r e n t } }$ , while retaining all other edge connections.
151
+
152
+ We then combine the obtained computation graphs $G _ { q ^ { ( 1 ) } } , \dots , G _ { q ^ { ( N ) } }$ as follows to give the final equivalent computation graph.
153
+
154
+ 1. Convert the target sink nodes of all the obtained computation graphs into the existentially quantified bound variables nodes.
155
+ 2. Create a new target sink node $V _ { ? }$ , and draw directed edges of type “union” from all the above variable nodes to the new target node.
156
+
157
+ An example of the entire transformation procedure is illustrated in Fig. 3. By the definition of the union operation, our procedure gives the equivalent computation graph as the original one. Furthermore, as all the union operators are removed from $G _ { q ^ { ( 1 ) } } , \dots , G _ { q ^ { ( N ) } }$ , all of these computation graphs represent conjunctive queries, which we denote as $\boldsymbol q ^ { ( 1 ) } , \ldots , \boldsymbol q ^ { ( N ) }$ . We can then apply existing framework to obtain a set of embeddings for these conjunctive queries as $\mathbf { q } ^ { ( 1 ) } , \ldots , \mathbf { q } ^ { ( \bar { \mathbf { N } } ) }$ .
158
+
159
+ Aggregation. Next we define the distance function between the given EPFO query $q$ and an entity $v \in \nu$ . Since $q$ is logically equivalent to $q ^ { ( 1 ) } \vee \cdots \vee q ^ { ( N ) }$ , we can naturally define the aggregated distance function using the box distance $\mathrm { d i s t } _ { \mathrm { b o x } }$ :
160
+
161
+ $$
162
+ \mathrm { d i s t _ { a g g } ( { \bf v } ; { \boldsymbol q } ) } = \mathrm { M i n } ( \{ \mathrm { d i s t _ { b o x } ( { \bf v } ; { \bf q } ^ { ( 1 ) } ) , \dots , d i s t _ { b o x } ( { \bf v } ; { \bf q } ^ { ( N ) } ) } \} ) ,
163
+ $$
164
+
165
+ where $\mathrm { d i s t _ { a g g } }$ is parameterized by the EPFO query $q$ . When $q$ is a conjunctive query, i.e., $N = 1$ , $\mathrm { d i s t } _ { \mathrm { a g g } } ( \mathbf { v } ; q ) ^ { \sim } = \mathrm { d i s t } _ { \mathrm { b o x } } ( \mathbf { v } ; \mathbf { q } )$ . For $N > 1$ , $\mathrm { d i s t _ { a g g } }$ takes the minimum distance to the closest box as the distance to an entity. This modeling aligns well with the union operation; an entity is inside the union of sets as long as the entity is in one of the sets. Note that our DNF-query rewriting scheme is general and is able to extend any method that works for conjunctive queries (e.g., (Hamilton et al., 2018)) to handle more general class of EPFO queries.
166
+
167
+ Computational complexity. The computational complexity of answering an EPFO query with our framework is equal to that of answering the $N$ conjunctive queries. In practice, $N$ might not be so large, and all the $N$ computations can be parallelized. Furthermore, answering each conjunctive query is very fast as it requires us to execute a sequence of simple box operations (each of which takes constant time) and then perform a range search (Bentley & Friedman, 1979) in the embedding space, which can also be done in constant time using techniques based on Locality Sensitive Hashing (Indyk & Motwani, 1998).
168
+
169
+ # 4 EXPERIMENTS
170
+
171
+ Our goal in the experiment section is to evaluate the performance of QUERY2BOX on discovering answers to complex logical queries that cannot be obtained by traversing the incomplete KG. This means, we will focus on answering queries where one or more missing edges in the KG have to be successfully predicted in order to obtain the additional answers.
172
+
173
+ # 4.1 KNOWLEDGE GRAPHS AND QUERY GENERATION
174
+
175
+ We perform experiments on three standard KG benchmarks, FB15k (Bordes et al., 2013), FB15k-237 (Toutanova & Chen, 2015), and NELL995 (Xiong et al., 2017) (see Appendix E for NELL995 pre-processing details). Dataset statistics are summarized in Table 5 in Appendix F.
176
+
177
+ ![](images/be57143a5bf103712d18b376e7ab22dad69199041222cca70a3ee43886b15b1d.jpg)
178
+ Figure 4: Query structures considered in the experiments, where anchor entities and relations are to be specified to instantiate logical queries. Naming for each query structure is provided under each subfigure, where ‘p’, ‘i’, and ‘u’ stand for ‘projection’, ‘intersection’, and ‘union’, respectively. Models are trained on the first 5 query structures, and evaluated on all 9 query structures. For example, “3p” is a path query of length three, and “2i” is an intersection of cardinality two.
179
+
180
+ <table><tr><td rowspan=1 colspan=1>Dataset</td><td rowspan=1 colspan=1>1p</td><td rowspan=1 colspan=1>2p</td><td rowspan=1 colspan=1>3p</td><td rowspan=1 colspan=1>2i</td><td rowspan=1 colspan=1>3i</td><td rowspan=1 colspan=1>ip</td><td rowspan=1 colspan=1>pi</td><td rowspan=1 colspan=1>2u</td><td rowspan=1 colspan=1>up</td></tr><tr><td rowspan=1 colspan=1>FB15k</td><td rowspan=1 colspan=1>10.8</td><td rowspan=1 colspan=1>255.6</td><td rowspan=1 colspan=1>250.0</td><td rowspan=1 colspan=1>90.3</td><td rowspan=1 colspan=1>64.1</td><td rowspan=1 colspan=1>593.8</td><td rowspan=1 colspan=1>190.1</td><td rowspan=1 colspan=1>27.8</td><td rowspan=1 colspan=1>227.0</td></tr><tr><td rowspan=1 colspan=1>FB15k-237</td><td rowspan=1 colspan=1>13.3</td><td rowspan=1 colspan=1>131.4</td><td rowspan=1 colspan=1>215.3</td><td rowspan=1 colspan=1>69.0</td><td rowspan=1 colspan=1>48.9</td><td rowspan=1 colspan=1>593.8</td><td rowspan=1 colspan=1>257.7</td><td rowspan=1 colspan=1>35.6</td><td rowspan=1 colspan=1>127.7</td></tr><tr><td rowspan=1 colspan=1>NELL995</td><td rowspan=1 colspan=1>8.5</td><td rowspan=1 colspan=1>56.6</td><td rowspan=1 colspan=1>65.3</td><td rowspan=1 colspan=1>30.3</td><td rowspan=1 colspan=1>15.9</td><td rowspan=1 colspan=1>310.0</td><td rowspan=1 colspan=1>144.9</td><td rowspan=1 colspan=1>14.4</td><td rowspan=1 colspan=1>62.5</td></tr></table>
181
+
182
+ Table 1: Average number of answer entities of test queries with missing edges grouped by different query structures (for a KG with $10 \%$ edges missing).
183
+
184
+ We follow the standard evaluation protocol in KG literture: Given the standard split of edges into training, test, and validation sets, we first augment the KG to also include inverse relations and effectively double the number of edges in the graph. We then create three graphs: $\mathcal { G } _ { \mathrm { t r a i n } }$ , which only contains training edges and we use this graph to train node embeddings as well as box operators. We then also generate two bigger graphs: $\mathcal { G } _ { \mathrm { v a l i d } }$ , which contains $\mathcal { G } _ { \mathrm { t r a i n } }$ plus the validation edges, and $\mathcal { G } _ { \mathrm { t e s t } }$ which includes $\mathcal { G } _ { \mathrm { v a l i d } }$ as well as the test edges.
185
+
186
+ We consider 9 kinds of diverse query structures shown and named in Fig. 4. We use 5 query structures for training and then evaluate on all the 9 query structures. We refer the reader to Appendix D for full details on query generation and Table 6 in Appendix F for statistics of the generated logical queries. Given a query $q$ , let $[ [ q ] ] _ { \mathrm { t r a i n } }$ , $[ [ q ] ] _ { \mathrm { v a l } }$ , and $[ [ q ] ] _ { \mathrm { t e s t } }$ denote a set of answer entities obtained by Jrunning subgraph matching of $q$ Kon $\mathcal { G } _ { \mathrm { t r a i n } }$ , $\mathcal { G } _ { \mathrm { v a l i d } }$ J K, and $\mathcal { G } _ { \mathrm { t e s t } }$ , respectively. At the training time, we use $[ [ q ] ] _ { \mathrm { t r a i n } }$ as positive examples for the query and other random entities as negative examples. However, J Kat the test/validation time we proceed differently. Note that we focus on answering queries where generalization performance is crucial and at least one edge needs to be imputed in order to answer the queries. Thus, rather than evaluating a given query on the full validation (or test) set $[ [ q ] ] _ { \mathrm { v a l } }$ $\left( \mathbb { I } q \mathbb { I } _ { \mathrm { t e s t } } \right)$ J K of answers, we validate the method only on answers that include missing relations. Given J Khow we constructed ${ \mathcal { G } } _ { \mathrm { t r a i n } } \subseteq { \mathcal { G } } _ { \mathrm { v a l i d } } \subseteq { \mathcal { G } } _ { \mathrm { t e s t } } .$ , we have $[ [ q ] ] _ { \mathrm { t r a i n } } \subseteq [ [ q ] ] _ { \mathrm { v a l } } \subseteq [ [ q ] ] _ { \mathrm { t e s t } }$ and thus we evaluate the method on $[ [ q ] _ { \mathrm { { l v a l } } } \backslash [ [ q ] ] _ { \mathrm { { t r a i n } } }$ J K J K J Kto tune hyper-parameters and then report results identifying answer entities in $[ [ q ] _ { \mathrm { t e s t } } \bar { \backslash } \bar { [ [ q ] ] _ { \mathrm { v a l } } }$ J K. This means we always evaluate on queries/entities that were not part of the J K J Ktraining set and the method has not seen them before. Furthermore, for these queries, traditional graph traversal techniques would not be able to find the answers (due to missing relations).
187
+
188
+ Table 1 shows the average number of answer entities for different query structures. We observe that complex logical queries (especially 2p, 3p, ip, pi, up) indeed require modeling a much larger number of answer entities (often more than 10 times) than the simple 1p queries do. Therefore, we expect our box embeddings to work particularly well in handling complex queries with many answer entities.3
189
+
190
+ # 4.2 EVALUATION PROTOCOL
191
+
192
+ Given a test query $q$ , for each of its non-trivial answers $v \in [ [ q ] ] _ { \mathrm { t e s t } } \backslash [ [ q ] ] _ { \mathrm { v a l } }$ , we use $\mathrm { d i s t } _ { \mathrm { b o x } }$ in Eq. 3 to rank $v$ among $\mathcal { V } \backslash \ [ q \ ] _ { \mathrm { t e s t } }$ . Denoting the rank of $v$ by Rank $( v )$ J K J K, we then calculate evaluation metrics for answering query $q$ K, such as Mean Reciprocal Rank (MRR) and Hits at $K$ $( \mathrm { H } @ K )$ :
193
+
194
+ $$
195
+ \mathrm { M e t r i c s } ( q ) = \frac { 1 } { | [ q ] _ { \mathrm { t e s t } } \setminus [ [ q ] _ { \mathrm { v a l } } ] } \sum _ { \substack { v \in [ [ q ] _ { \mathrm { t e s t } } \setminus [ q ] _ { \mathrm { v a l } } } } f _ { \mathrm { m e t r i c s } } ( \mathrm { R a n k } ( v ) ) ,
196
+ $$
197
+
198
+ <table><tr><td>Method</td><td>Avg</td><td>1p</td><td>2p</td><td>3p</td><td>2i</td><td>3i</td><td>ip</td><td>pi</td><td>2u</td><td>up</td></tr><tr><td colspan="9"></td></tr><tr><td>Q2B</td><td>0.484</td><td>0.786</td><td>0.413</td><td>FB15k 0.303</td><td>0.593</td><td>0.712</td><td>0.211</td><td>0.397</td><td>0.608</td><td>0.33</td></tr><tr><td>GQE</td><td>0.386</td><td>0.636</td><td>0.345</td><td>0.248</td><td>0.515</td><td>0.624</td><td>0.151</td><td>0.310</td><td>0.376</td><td>0.273</td></tr><tr><td>GQE-DOUBLE</td><td>0.384</td><td>0.630</td><td>0.346</td><td>0.250</td><td>0.515</td><td>0.611</td><td>0.153</td><td>0.320</td><td>0.362</td><td>0.271</td></tr><tr><td colspan="9"></td></tr><tr><td>Q2B</td><td>0.268</td><td>0.467</td><td>0.24</td><td>FB15k-237 0.186</td><td>0.324</td><td>0.453</td><td>0.108</td><td>0.205</td><td>0.239</td><td>0.193</td></tr><tr><td>GQE</td><td>0.228</td><td>0.402</td><td>0.213</td><td>0.155</td><td>0.292</td><td>0.406</td><td>0.083</td><td>0.17</td><td>0.169</td><td>0.163</td></tr><tr><td>GQE-DOUBLE</td><td>0.23</td><td>0.405</td><td>0.213</td><td>0.153</td><td>0.298</td><td>0.411</td><td>0.085</td><td>0.182</td><td>0.167</td><td>0.16</td></tr><tr><td colspan="9">NELL995</td></tr><tr><td>Q2B</td><td>0.306</td><td>0.555</td><td>0.266</td><td>0.233</td><td>0.343</td><td>0.48</td><td>0.132</td><td>0.212</td><td>0.369</td><td>0.163</td></tr><tr><td>GQE</td><td>0.247</td><td>0.418</td><td>0.228</td><td>0.205</td><td>0.316</td><td>0.447</td><td>0.081</td><td>0.186</td><td>0.199</td><td>0.139</td></tr><tr><td>GQE-DOUBLE</td><td>0.248</td><td>0.417</td><td>0.231</td><td>0.203</td><td>0.318</td><td>0.454</td><td>0.081</td><td>0.188</td><td>0.2</td><td>0.139</td></tr></table>
199
+
200
+ Table 2: $\mathrm { H @ 3 }$ results of QUERY2BOX vs. GQE on FB15k, FB15k-237 and NELL995.
201
+
202
+ where $\begin{array} { r } { f _ { \mathrm { m e t r i c s } } ( x ) = \frac { 1 } { x } } \end{array}$ for MRR, and $f _ { \mathrm { m e t r i c s } } ( x ) = \mathbf { 1 } [ x \leq K ]$ for $\mathrm { H @ } K$
203
+
204
+ We then average Eq. 6 over all the queries within the same query structure,4 and report the results separately for different query structures. The same evaluation protocol is applied to the validation stage except that we evaluate on $[ [ q ] _ { \mathrm { v a l } } \backslash [ [ q ] ] _ { \mathrm { t r a i n } }$ rather than $[ [ q ] _ { \mathrm { l e s t } } \backslash [ [ q ] ] _ { \mathrm { v a l } }$ .
205
+
206
+ # 4.3 BASELINE AND MODEL VARIANTS
207
+
208
+ We compare our framework QUERY2BOX against the state-of-the-art GQE (Hamilton et al., 2018). GQE embeds a query to a single vector, and models projection and intersection operators as translation and deep sets (Zaheer et al., 2017), respectively. The $L _ { 1 }$ distance is used as the distance between query and entity vectors. For a fair comparison, we also compare with GQE-DOUBLE (GQE with doubled embedding dimensionality) so that QUERY2BOX and GQE-DOUBLE have the same amount of parameters. Refer to Appendix G for the model hyper-parameters used in our experiments. Although the original GQE cannot handle EPFO queries, we apply our DNF-query rewriting strategy and in our evaluation extend GQE to handle general EPFO queries as well. Furthermore, we perform extensive ablation study by considering several variants of QUERY2BOX (abbreviated as Q2B). We list our method as well as its variants below.
209
+
210
+ • Q2B (our method): The box embeddings are used to model queries, and the attention mechanism is used for the intersection operator.
211
+ • Q2B-AVG: The attention mechanism for intersection is replaced with averaging.
212
+ • Q2B-DEEPSETS: The attention mechanism for intersection is replaced with the deep sets. Q2B-AVG-1P: The variant of Q2B-AVG that is trained with only 1p queries (see Fig. 4); thus, logical operators are not explicitly trained. Q2B-SHAREDOFFSET; The box offset is shared across all queries (every query is represented by a box with the same trainable size).
213
+
214
+ # 4.4 MAIN RESULTS
215
+
216
+ We start by comparing our Q2B with state-of-the-art query embedding method GQE (Hamilton et al., 2018) on FB15k, FB15k-237, and NELL995. As listed in Tables 2, our method significantly and consistently outperforms the state-of-the-art baseline across all the query structures, including those not seen during training as well as those with union operations. On average, we obtain $9 . 8 \%$ $2 5 \%$ relative), $3 . 8 \%$ 7 $1 5 \%$ relative), and $5 . 9 \%$ $24 \%$ relative) higher $\mathrm { H @ 3 }$ than the best baselines on FB15k, FB15k-237, and NELL995, respectively. Notice that naïvely increasing embedding dimensionality in GQE yields limited performance improvement. Our Q2B is able to effectively model a large set of entities by using the box embedding, and achieves a significant performance gain compared with GQE-DOUBLE (with same number of parameters) that represents queries as point vectors. Also notice that Q2B performs well on new queries with the same structure as the training queries as well as on new query structures never seen during training, which demonstrates that Q2B generalizes well within and beyond query structures.
217
+
218
+ <table><tr><td>Method</td><td>Avg</td><td>1p</td><td>2p</td><td>3p</td><td>2i</td><td>3i</td><td>ip</td><td>pi</td><td>2u</td><td>up</td></tr><tr><td colspan="9">FB15k</td></tr><tr><td>Q2B</td><td>0.484</td><td>0.786</td><td>0.413</td><td>0.303</td><td>0.593</td><td>0.712</td><td>0.211</td><td>0.397</td><td>0.608</td><td>0.330</td></tr><tr><td>Q2B-AVG</td><td>0.468</td><td>0.779</td><td>0.407</td><td>0.300</td><td>0.577</td><td>0.673</td><td>0.199</td><td>0.345</td><td>0.607</td><td>0.326</td></tr><tr><td>Q2B-DEEPSETS</td><td>0.467</td><td>0.755</td><td>0.407</td><td>0.294</td><td>0.588</td><td>0.699</td><td>0.197</td><td>0.378</td><td>0.562</td><td>0.324</td></tr><tr><td>Q2B-AVG-1P</td><td>0.385</td><td>0.812</td><td>0.262</td><td>0.173</td><td>0.463</td><td>0.529</td><td>0.126</td><td>0.263</td><td>0.653</td><td>0.187</td></tr><tr><td>Q2B-SHAREDOFFSET</td><td>0.372</td><td>0.684</td><td>0.335</td><td>0.232</td><td>0.442</td><td>0.559</td><td>0.144</td><td>0.282</td><td>0.417</td><td>0.252</td></tr><tr><td colspan="9">FB15k-237</td></tr><tr><td>Q2B</td><td>0.268</td><td>0.467</td><td>0.24</td><td>0.186</td><td>0.324</td><td>0.453</td><td>0.108</td><td>0.205</td><td>0.239</td><td>0.193</td></tr><tr><td>Q2B-AVG</td><td>0.249</td><td>0.462</td><td>0.242</td><td>0.182</td><td>0.278</td><td>0.391</td><td>0.101</td><td>0.158</td><td>0.236</td><td>0.189</td></tr><tr><td>Q2B-DEEPSETS</td><td>0.259</td><td>0.458</td><td>0.243</td><td>0.186</td><td>0.303</td><td>0.432</td><td>0.104</td><td>0.187</td><td>0.231</td><td>0.190</td></tr><tr><td>Q2B-AVG-1P</td><td>0.219</td><td>0.457</td><td>0.193</td><td>0.132</td><td>0.251</td><td>0.319</td><td>0.083</td><td>0.142</td><td>0.241</td><td>0.152</td></tr><tr><td>Q2B-SHAREDOFFSET</td><td>0.207</td><td>0.391</td><td>0.199</td><td>0.139</td><td>0.251</td><td>0.354</td><td>0.082</td><td>0.154</td><td>0.15</td><td>0.142</td></tr><tr><td colspan="9">NELL995</td></tr><tr><td>Q2B</td><td>0.306</td><td>0.555</td><td>0.266</td><td>0.233</td><td>0.343</td><td>0.480</td><td>0.132</td><td>0.212</td><td>0.369</td><td>0.163</td></tr><tr><td>Q2B-AVG</td><td>0.283</td><td>0.543</td><td>0.250</td><td>0.228</td><td>0.300</td><td>0.403</td><td>0.116</td><td>0.188</td><td>0.36</td><td>0.161</td></tr><tr><td>Q2B-DEEPSETS</td><td>0.293</td><td>0.539</td><td>0.26</td><td>0.231</td><td>0.317</td><td>0.467</td><td>0.11</td><td>0.202</td><td>0.349</td><td>0.16</td></tr><tr><td>Q2B-AVG-1P</td><td>0.274</td><td>0.607</td><td>0.229</td><td>0.182</td><td>0.277</td><td>0.315</td><td>0.097</td><td>0.18</td><td>0.443</td><td>0.133</td></tr><tr><td>Q2B-SHAREDOFFSET</td><td>0.237</td><td>0.436</td><td>0.219</td><td>0.201</td><td>0.278</td><td>0.379</td><td>0.096</td><td>0.174</td><td>0.217</td><td>0.137</td></tr></table>
219
+
220
+ Table 3: $\mathrm { H @ 3 }$ results of QUERY2BOX vs. several variants on FB15k, FB15k-237 and NELL995.
221
+
222
+ We also conduct extensive ablation studies (Tables 3). We summarize the results as follows:
223
+
224
+ Importance of attention mechanism. First, we show that our modeling of intersection using the attention mechanism is important. Given a set of box embeddings $\{ \mathbf { p _ { 1 } } , \dotsc , \mathbf { p _ { n } } \}$ , Q2B-AVG is the most naïve way to calculate the center of the resulting box embedding $\mathbf { p } _ { \mathrm { i n t e r } }$ while Q2B-DEEPSETS is too flexible and neglects the fact that the center should be a weighted average of $\mathbf { C e n } ( \mathbf { p _ { 1 } } ) , \dots , \mathbf { C e n } ( \mathbf { p _ { n } } )$ Compared with the two methods, Q2B achieves better performance in answering queries that involve intersection operation, e.g., 2i, 3i, pi, ip. Specifically, on FB15k-237, Q2B obtains more than $4 \%$ and $2 \%$ absolute gain in $\mathrm { H @ 3 }$ compared to Q2B-AVG and Q2B-DEEPSETS, respectively.
225
+
226
+ Necessity of training on complex queries. Second, we observe that explicitly training on complex logical queries beyond one-hop path queries (1p in Fig. 4) improves the reasoning performance. Although Q2B-AVG-1P is able to achieve strong performance on 1p and $2 \mathrm { u }$ , where answering $2 \mathrm { u }$ is essentially answering two 1p queries with an additional minimum operation (see Eq. 5 in Section 3.3), Q2B-AVG-1P fails miserably in answering other types of queries involving logical operators. On the other hand, other methods (Q2B, Q2B-AVG, and Q2B-DEEPSETS) that are explicitly trained on the logical queries achieve much higher accuracy, with up to $10 \%$ absolute average improvement of $\mathrm { H @ 3 }$ on FB15k.
227
+
228
+ Adaptive box size for different queries. Third, we investigate the importance of learning adaptive offsets (box size) for different queries. Q2B-SHAREDOFFSET is a variant of our Q2B where all the box embeddings share the same learnable offset. Q2B-SHAREDOFFSET does not work well on all types of queries. This is most likely because different queries have different numbers of answer entities, and the adaptive box size enables us to better model it. In fact, we find that box offset varies significantly across different relations, and one-to-many relations tend to have larger offset embeddings (see Appendix H for the details).
229
+
230
+ # 5 CONCLUSION
231
+
232
+ In this paper we proposed a reasoning framework called QUERY2BOX that can effectively model and reason over sets of entities as well as handle EPFO queries in a vector space. Given a logical query, we first transform it into DNF, embed each conjunctive query into a box, and output entities closest to their nearest boxes. Our approach is capable of handling all types of EPFO queries scalably and accurately. Experimental results on standard KGs demonstrate that QUERY2BOX significantly outperforms the existing work in answering diverse logical queries.
233
+
234
+ # ACKNOWLEDGMENTS
235
+
236
+ We thank William Hamilton, Rex Ying, and Jiaxuan You for their helpful discussion. W.H is supported by Funai Overseas Scholarship and Masason Foundation Fellowship. J.L is a Chan Zuckerberg Biohub investigator. We gratefully acknowledge the support of DARPA under Nos. FA865018C7880 (ASED), N660011924033 (MCS); ARO under Nos. W911NF-16-1-0342 (MURI), W911NF-16-1-0171 (DURIP); NSF under Nos. OAC-1835598 (CINES), OAC-1934578 (HDR); Stanford Data Science Initiative, Wu Tsai Neurosciences Institute, Chan Zuckerberg Biohub, JD.com, Amazon, Boeing, Docomo, Huawei, Hitachi, Observe, Siemens, UST Global.
237
+
238
+ The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright notation thereon. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views, policies, or endorsements, either expressed or implied, of DARPA, NIH, ARO, or the U.S. Government.
239
+
240
+ # REFERENCES
241
+
242
+ Carl Allen, Ivana Balazevic, and Timothy M Hospedales. On understanding knowledge graph representation. arXiv preprint arXiv:1909.11611, 2019.
243
+
244
+ Ben Athiwaratkun and Andrew Gordon Wilson. Hierarchical density order embeddings. In International Conference on Learning Representations (ICLR), 2018.
245
+
246
+ Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. In International Conference on Learning Representations (ICLR), 2015.
247
+
248
+ Jon Louis Bentley and Jerome H Friedman. Data structures for range searching. ACM Computing Surveys (CSUR), 11(4):397–409, 1979.
249
+
250
+ Kurt Bollacker, Colin Evans, Praveen Paritosh, Tim Sturge, and Jamie Taylor. Freebase: a collaboratively created graph database for structuring human knowledge. In ACM SIGMOD international conference on Management of data (SIGMOD), pp. 1247–1250. AcM, 2008.
251
+
252
+ Antoine Bordes, Nicolas Usunier, Alberto Garcia-Duran, Jason Weston, and Oksana Yakhnenko. Translating embeddings for modeling multi-relational data. In Advances in Neural Information Processing Systems (NeurIPS), pp. 2787–2795, 2013.
253
+
254
+ Nilesh Dalvi and Dan Suciu. Efficient query evaluation on probabilistic databases. VLDB, 16(4): 523–544, 2007.
255
+
256
+ Nilesh Dalvi and Dan Suciu. The dichotomy of probabilistic inference for unions of conjunctive queries. Journal of the ACM (JACM), 59(6):30, 2012.
257
+
258
+ Rajarshi Das, Arvind Neelakantan, David Belanger, and Andrew McCallum. Chains of reasoning over entities, relations, and text using recurrent neural networks. In European Chapter of the Association for Computational Linguistics (EACL), pp. 132–141, 2017.
259
+
260
+ Brian A Davey and Hilary A Priestley. Introduction to lattices and order. Cambridge university press, 2002.
261
+
262
+ Luc De Raedt. Logical and relational learning. Springer Science & Business Media, 2008.
263
+
264
+ Sašo Džeroski. Relational data mining. In Data Mining and Knowledge Discovery Handbook, pp. 887–911. Springer, 2009.
265
+
266
+ Katrin Erk. Representing words as regions in vector space. In Proceedings of the Thirteenth Conference on Computational Natural Language Learning, pp. 57–65. Annual Meeting of the Association for Computational Linguistics (ACL), 2009.
267
+
268
+ Kelvin Guu, John Miller, and Percy Liang. Traversing knowledge graphs in vector space. In Empirical Methods in Natural Language Processing (EMNLP), pp. 318–327, 2015.
269
+
270
+ Will Hamilton, Payal Bajaj, Marinka Zitnik, Dan Jurafsky, and Jure Leskovec. Embedding logical queries on knowledge graphs. In Advances in Neural Information Processing Systems (NeurIPS), pp. 2027–2038, 2018.
271
+
272
+ Shizhu He, Kang Liu, Guoliang Ji, and Jun Zhao. Learning to represent knowledge graphs with gaussian embedding. In Proceedings of the 24th ACM International on Conference on Information and Knowledge Management, pp. 623–632. ACM, 2015.
273
+
274
+ Piotr Indyk and Rajeev Motwani. Approximate nearest neighbors: towards removing the curse of dimensionality. In Proceedings of the thirtieth annual ACM symposium on Theory of computing, pp. 604–613. ACM, 1998.
275
+
276
+ Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015.
277
+
278
+ Daphne Koller, Nir Friedman, Sašo Džeroski, Charles Sutton, Andrew McCallum, Avi Pfeffer, Pieter Abbeel, Ming-Fai Wong, David Heckerman, Chris Meek, et al. Introduction to statistical relational learning. MIT press, 2007.
279
+
280
+ Denis Krompaß, Maximilian Nickel, and Volker Tresp. Querying factorized probabilistic triple databases. In International Semantic Web Conference, pp. 114–129. Springer, 2014.
281
+
282
+ Alice Lai and Julia Hockenmaier. Learning to predict denotational probabilities for modeling entailment. In Annual Meeting of the Association for Computational Linguistics (ACL), pp. 721–730, 2017.
283
+
284
+ Xiang Li, Luke Vilnis, and Andrew McCallum. Improved representation learning for predicting commonsense ontologies. arXiv preprint arXiv:1708.00549, 2017.
285
+
286
+ Xiang Li, Luke Vilnis, Dongxu Zhang, Michael Boratko, and Andrew McCallum. Smoothing the geometry of probabilistic box embeddings. In International Conference on Learning Representations (ICLR), 2019.
287
+
288
+ Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In International Conference on Learning Representations (ICLR), 2013.
289
+
290
+ Maximilian Nickel, Kevin Murphy, Volker Tresp, and Evgeniy Gabrilovich. A review of relational machine learning for knowledge graphs. Proceedings of the IEEE, 104(1):11–33, 2016.
291
+
292
+ Zhiqing Sun, Zhi-Hong Deng, Jian-Yun Nie, and Jian Tang. Rotate: Knowledge graph embedding by relational rotation in complex space. In International Conference on Learning Representations (ICLR), 2019.
293
+
294
+ Kristina Toutanova and Danqi Chen. Observed versus latent features for knowledge base and text inference. In Proceedings of the 3rd Workshop on Continuous Vector Space Models and their Compositionality, pp. 57–66, 2015.
295
+
296
+ Vladimir Vapnik. The nature of statistical learning theory. Springer science & business media, 2013.
297
+
298
+ Ivan Vendrov, Ryan Kiros, Sanja Fidler, and Raquel Urtasun. Order-embeddings of images and language. In International Conference on Learning Representations (ICLR), 2016.
299
+
300
+ John Venn. I. on the diagrammatic and mechanical representation of propositions and reasonings. The London, Edinburgh, and Dublin philosophical magazine and journal of science, 10(59):1–18, 1880.
301
+
302
+ Luke Vilnis and Andrew McCallum. Word representations via gaussian embedding. In International Conference on Learning Representations (ICLR), 2014.
303
+
304
+ Luke Vilnis, Xiang Li, Shikhar Murty, and Andrew McCallum. Probabilistic embedding of knowledge graphs with box lattice measures. In Annual Meeting of the Association for Computational Linguistics (ACL), 2018.
305
+
306
+ Wenhan Xiong, Thien Hoang, and William Yang Wang. Deeppath: A reinforcement learning method for knowledge graph reasoning. In Empirical Methods in Natural Language Processing (EMNLP), 2017.
307
+
308
+ Manzil Zaheer, Satwik Kottur, Siamak Ravanbakhsh, Barnabas Poczos, Ruslan R Salakhutdinov, and Alexander J Smola. Deep sets. In Advances in Neural Information Processing Systems (NeurIPS), pp. 3391–3401, 2017.
309
+
310
+ # A PROOF OF THEOREM 1
311
+
312
+ Proof. To model any EPFO query, we need to at least model a subset of EPFO queries $\mathcal { Q } = \{ \lor _ { q _ { i } \in S } q _ { i } :$ ${ \cal S } \subseteq \{ q _ { 1 } , \dots , q _ { M } \} \}$ , where the corresponding denotation sets are $\left\{ \cup _ { q _ { i } \in S } \left[ \left[ q _ { i } \right] \right] : S \subseteq \left\{ q _ { 1 } , \ldots , { \overset { \cdot } { q } } _ { M } \right\} \right\}$ . For the sake of modeling $\mathcal { Q }$ J K, without loss of generality, we consider assigning a single entity embedding $\bf v _ { q _ { i } }$ to all $v \in [ [ q _ { i } ] ]$ , so there are $M$ kinds of entity vectors, $\mathbf { v _ { q 1 } } , \dots , \mathbf { v _ { q _ { M } } }$ . To model all queries in $\mathcal { Q }$ J K, it is necessary to satisfy the following.
313
+
314
+ $$
315
+ \exists \mathbf { v _ { q _ { 1 } } } , \dotsc , \exists \mathbf { v _ { q _ { M } } } , \forall S \subseteq \{ q _ { 1 } , \dotsc , q _ { M } \} , \exists \mathbf { q } _ { \mathbf { S } } \in \Xi , \mathrm { s u c h ~ t h a t ~ d i s t } ( \mathbf { v _ { q _ { i } } } ; \mathbf { q } _ { \mathbf { S } } ) \left\{ \underset { > \beta } \leq \beta \ \mathrm { i f } \ q _ { i } \in S , \right.
316
+ $$
317
+
318
+ where qS is the embedding of query $\mathsf { V } _ { q _ { i } \in S } q _ { i }$ . Eq. 7 means that we can learn the $M$ kinds of entity vectors such that for every query in $\mathcal { Q }$ , we can obtain its embedding to model the corresponding set using the distance function. Notice that this is agnostic to the specific algorithm to embed query $\mathsf { V } _ { q \in S } q$ into $\mathbf { q } \mathbf { s }$ ; thus, our result is generally applicable to any method that embeds the query into a single vector.
319
+
320
+ Crucially, satisfying Eq. 7 is equivalent to $\{ \operatorname { s i g n } ( \beta - \mathrm { d i s t } ( \cdot ; \mathbf { q } ) ) : \mathbf { q } \in \Xi \}$ being able to shutter $\{ \mathbf { v _ { q _ { 1 } } } , \dotsc , \mathbf { v _ { q _ { M } } } \}$ , i.e., any binary labeling of the points can be perfectly fit by some classifier in the function class. To sum up, in order to model any EPFO query, we need to at least model any query in $\mathcal { Q }$ , which requires the VC dimension of the distance function to be larger than or equal to $M$
321
+
322
+ # B DETAILS ABOUT COMPUTING $M$ IN THEOREM 1
323
+
324
+ Given the full KG $\mathcal { G } _ { \mathrm { t e s t } }$ for the FB15k dataset, our goal is to find conjunctive queries $q _ { 1 } , \dots , q _ { M }$ such that $[ [ q _ { 1 } ] ] , \dots , [ [ q _ { M } ] ]$ are disjoint with each other. For conjunctive queries, we use two types of queries: J K J K‘1p’ and ‘2i’ whose query structures are shown in Figure 4. On the FB15k, we instantiate 308,006 queries of type $ { \mathrm { \cdot } } 1 { \mathrm { p } } ^ { { \prime } }$ , which we denote by $S _ { \mathrm { 1 p } }$ . Out of all the queries in $S _ { \mathrm { 1 p } }$ , 129,717 queries have more than one answer entities, and we denote such a set of the queries by $S _ { \mathrm { 1 p } } ^ { \prime }$ . We then generate a set of queries of type ‘2i’ by first randomly sampling two queries from $S _ { \mathrm { 1 p } } ^ { \prime }$ and then taking conjunction; we denote the resulting set of queries by $S _ { \mathrm { 2 i } }$ .
325
+
326
+ Now, we use $S _ { \mathrm { 1 p } }$ and $S _ { \mathrm { 2 i } }$ to generate a set of conjunctive queries whose denotation sets are disjoint with each other. First, we prepare two empty sets $\nu _ { \mathrm { s e e n } } = \emptyset$ , and $\mathcal { Q } = \mathcal { Q }$ . Then, for every $q \in S _ { \mathrm { 1 p } }$ , if $\mathcal { V } _ { \mathrm { s e e n } } \cap \ [ q ] = \emptyset$ holds, we let $\mathcal { Q } \mathcal { Q } \cup \{ q \}$ and $\mathcal { V } _ { \mathrm { s e e n } } \mathcal { V } _ { \mathrm { s e e n } } \cup [ [ q ] ]$ . This procedure already gives us $\mathcal { Q }$ J K J K, where we have 10, 812 conjunctive queries whose denotation sets are disjoint with each other. We can further apply the analogous procedure for $S _ { \mathrm { 2 i } }$ , which gives us a further increased $\mathcal { Q }$ , where we have 13, 365 conjunctive queries whose denotation sets are disjoint with each other. Therefore, we get $M = 1 3 , 3 6 5$ .
327
+
328
+ # C EXPERIMENTS ON LINK PREDICTION
329
+
330
+ <table><tr><td rowspan=1 colspan=1></td><td rowspan=1 colspan=2>FB15k</td><td rowspan=1 colspan=2>FB15k-237</td><td rowspan=1 colspan=2>NELL995</td></tr><tr><td rowspan=1 colspan=1>Method</td><td rowspan=1 colspan=1>H@3</td><td rowspan=1 colspan=1>MRR</td><td rowspan=1 colspan=1>H@3</td><td rowspan=1 colspan=1>MRR</td><td rowspan=1 colspan=1>H@3</td><td rowspan=1 colspan=1>MRR</td></tr><tr><td rowspan=1 colspan=1>query2box</td><td rowspan=1 colspan=1>0.613</td><td rowspan=1 colspan=1>0.516</td><td rowspan=1 colspan=1>0.331</td><td rowspan=1 colspan=1>0.295</td><td rowspan=1 colspan=1>0.382</td><td rowspan=1 colspan=1>0.303</td></tr><tr><td rowspan=1 colspan=1>query2box-1p</td><td rowspan=1 colspan=1>0.633</td><td rowspan=1 colspan=1>0.531</td><td rowspan=1 colspan=1>0.323</td><td rowspan=1 colspan=1>0.292</td><td rowspan=1 colspan=1>0.415</td><td rowspan=1 colspan=1>0.320</td></tr><tr><td rowspan=1 colspan=1>TransE</td><td rowspan=1 colspan=1>0.611</td><td rowspan=1 colspan=1>0.522</td><td rowspan=1 colspan=1>0.318</td><td rowspan=1 colspan=1>0.289</td><td rowspan=1 colspan=1>0.413</td><td rowspan=1 colspan=1>0.320</td></tr></table>
331
+
332
+ Table 4: Performance comparison on the simple link prediction task on the three datasets.
333
+
334
+ In Table 4, we report the link prediction performance (no multi-hop logical reasoning required) following the conventional metrics (taking average over the triples of head, relation, and tail). Here query2box is trained on all five query structures as shown in Figure 4, and query2box-1p is only trained on simple 1p queries. We found that our query2box is comparable or slightly better than TransE on simple link prediction. Note that in the case of simple link prediction, we do not expect a huge performance gain by using box embeddings as link prediction does not involve logical reasoning nor handling a large set of answer entities. Also, we see that even if we train query2box over diverse queries, its performance on link prediction is still comparable to TransE and query2box-1p, which are trained solely on the link prediction task.
335
+
336
+ ![](images/379d08b6ac8265a218ddea8b504b54dd9c61d79639ae5a06bf1450fb0c8130c7.jpg)
337
+ Figure 5: Example of the degenerated queries, including (1) $r$ and $r ^ { - 1 }$ appear along one path and (2) same anchor node and relation in intersections.
338
+
339
+ # D DETAILS ON QUERY GENERATION
340
+
341
+ Given $\mathcal { G } _ { \mathrm { t r a i n } }$ , $\mathcal { G } _ { \mathrm { v a l i d } }$ , and $\mathcal { G } _ { \mathrm { t e s t } }$ as defined in Section 4.1, we generate training, validation and test queries of different query structures. During training, we consider the first 5 kinds of query structures. For evaluation, we consider all the 9 query structures in Fig. 4, containing query structures that are both seen and unseen during training time. We instantiate queries in the following way.
342
+
343
+ Given a KG and a query structure (which is a DAG), we use pre-order traversal to assign an entity and a relation to each node and edge in the DAG of query structure to instantiate a query. Namely, we start from the root of the DAG (which is the target node), we sample an entity $e$ uniformly from the KG to be the root, then for every node connected to the root in the DAG, we choose a relation $r$ uniformly from the in-coming relations of $e$ in the KG, and a new entity $e ^ { \prime }$ from the set of entities that reaches $e$ by $r$ in the KG. Then we assign the relation $r$ to the edge and $e ^ { \prime }$ to the node, and move on the process based on the pre-order traversal. This iterative process stops after we assign an entity and relation to every node and edge in DAG. The leaf nodes in the DAG serve as the anchor nodes. Note that during the entity and relation assignment, we specifically filter out all the degenerated queries, as shown in Fig. D. Then we perform a post-order traversal of the DAG on the KG, starting from the anchor nodes, to obtain a set of answer entities to this query.
344
+
345
+ When generating validation/test queries, we explicitly filter out trivial queries that can be fully answered by subgraph matching on $\mathcal { G } _ { \mathrm { t r a i n } } / \mathcal { G } _ { \mathrm { v a l i d } }$ .
346
+
347
+ # E DETAILS OF NELL995 DATASET
348
+
349
+ Here we detail our pre-processing of the NELL995 dataset, which is originally presented by Xiong et al. (2017). Following Allen et al. (2019), we first combine the validation and test sets with the training set to create the whole knowledge graph for NELL995. Then we create new validation and test set splits by randomly selecting 20,000 triples each from the whole knowledge graph. Note that we filter out all the entities that only appear in the validation and test sets but not in the training set.
350
+
351
+ <table><tr><td rowspan=1 colspan=1>Dataset</td><td rowspan=1 colspan=1>Entities</td><td rowspan=1 colspan=1>Relations</td><td rowspan=1 colspan=1>Training Edges</td><td rowspan=1 colspan=1>Validation Edges</td><td rowspan=1 colspan=1>Test Edges</td><td rowspan=1 colspan=1>Total Edges</td></tr><tr><td rowspan=1 colspan=1>FB15k</td><td rowspan=1 colspan=1>14,951</td><td rowspan=1 colspan=1>1,345</td><td rowspan=1 colspan=1>483,142</td><td rowspan=1 colspan=1>50,000</td><td rowspan=1 colspan=1>59,071</td><td rowspan=1 colspan=1>592,213</td></tr><tr><td rowspan=1 colspan=1>FB15k-237</td><td rowspan=1 colspan=1>14,505</td><td rowspan=1 colspan=1>237</td><td rowspan=1 colspan=1>272,115</td><td rowspan=1 colspan=1>17,526</td><td rowspan=1 colspan=1>20,438</td><td rowspan=1 colspan=1>310,079</td></tr><tr><td rowspan=1 colspan=1>NELL995</td><td rowspan=1 colspan=1>63,361</td><td rowspan=1 colspan=1>200</td><td rowspan=1 colspan=1>114,213</td><td rowspan=1 colspan=1>14,324</td><td rowspan=1 colspan=1>14,267</td><td rowspan=1 colspan=1>142,804</td></tr></table>
352
+
353
+ Table 5: Knowledge graph dataset statistics as well as the split into training, validation, and test sets.
354
+
355
+ <table><tr><td rowspan=1 colspan=1>Queries</td><td rowspan=1 colspan=2>Training</td><td rowspan=1 colspan=2>Validation</td><td rowspan=1 colspan=2>Test</td></tr><tr><td rowspan=1 colspan=1>Dataset</td><td rowspan=1 colspan=1>1p</td><td rowspan=1 colspan=1>others</td><td rowspan=1 colspan=1>1p</td><td rowspan=1 colspan=1>others</td><td rowspan=1 colspan=1>1p</td><td rowspan=1 colspan=1>others</td></tr><tr><td rowspan=1 colspan=1>FB15k</td><td rowspan=1 colspan=1>273,710</td><td rowspan=1 colspan=1>273,710</td><td rowspan=1 colspan=1>59,097</td><td rowspan=1 colspan=1>8,000</td><td rowspan=1 colspan=1>67,016</td><td rowspan=1 colspan=1>8,000</td></tr><tr><td rowspan=1 colspan=1>FB15k-237</td><td rowspan=1 colspan=1>149,689</td><td rowspan=1 colspan=1>149,689</td><td rowspan=1 colspan=1>20,101</td><td rowspan=1 colspan=1>5,000</td><td rowspan=1 colspan=1>22.812</td><td rowspan=1 colspan=1>5,000</td></tr><tr><td rowspan=1 colspan=1>NELL995</td><td rowspan=1 colspan=1>107,982</td><td rowspan=1 colspan=1>107,982</td><td rowspan=1 colspan=1>16,927</td><td rowspan=1 colspan=1>4,000</td><td rowspan=1 colspan=1>17,034</td><td rowspan=1 colspan=1>4,000</td></tr></table>
356
+
357
+ Table 6: Number of training, validation, and test queries generated for different query structures.
358
+
359
+ # F DATASET STATISTICS
360
+
361
+ Table 5 summarizes the basic statistics of the three datasets used in our experiments. Table 6 summarizes the basic statistics of the generated logical queries.
362
+
363
+ # G HYPER-PARAMETERS
364
+
365
+ We use embedding dimensionality of $d = 4 0 0$ and set $\gamma = 2 4$ , $\alpha = 0 . 2$ for the loss in Eq. 4. We train all types of training queries jointly. In every iteration, we sample a minibatch size of 512 queries for each query structure (details in Appendix D), and we sample 1 answer entity and 128 negative entities for each query. We optimize the loss in Eq. 4 using Adam Optimizer (Kingma & Ba, 2015) with learning rate $= 0 . 0 0 0 1$ . We train all models for 250 epochs, monitor the performance on the validation set, and report the test performance.
366
+
367
+ # H ANALYSIS OF LEARNED BOX OFFSET SIZE
368
+
369
+ Here we study the correlation between the box size (measured by the L1 norm of the box offset) and the average number of entities that are contained in 1p queries using the corresponding relation. Table 7 shows the top 10 relations with smallest/largest box sizes. We observe a clear trend that the size of the box has a strong correlation with the number of entities the box encloses. Specifically, we see that one-to-many relations tend to have larger offset embeddings, which demonstrates that larger boxes are indeed used to model sets of more points (entities).
370
+
371
+ <table><tr><td rowspan=1 colspan=1>Top10 relationswith largest box size</td><td rowspan=1 colspan=1>#Ent</td><td rowspan=1 colspan=1>Box size</td></tr><tr><td rowspan=1 colspan=1>Tcommon/.../topic</td><td rowspan=1 colspan=1>3616.0</td><td rowspan=1 colspan=1>147.0</td></tr><tr><td rowspan=1 colspan=1>/user/...taxonomy</td><td rowspan=1 colspan=1>1.0</td><td rowspan=1 colspan=1>137.2</td></tr><tr><td rowspan=1 colspan=1>Tcommon/.../category</td><td rowspan=1 colspan=1>1.3</td><td rowspan=1 colspan=1>125.6</td></tr><tr><td rowspan=1 colspan=1>/base/../administrative_area_type</td><td rowspan=1 colspan=1>1.0</td><td rowspan=1 colspan=1>123.6</td></tr><tr><td rowspan=1 colspan=1>/medicine/../legal_status</td><td rowspan=1 colspan=1>1.5</td><td rowspan=1 colspan=1>114.9</td></tr><tr><td rowspan=1 colspan=1>/people/.../spouse</td><td rowspan=1 colspan=1>889.8</td><td rowspan=1 colspan=1>114.3</td></tr><tr><td rowspan=1 colspan=1>/sports/.../team</td><td rowspan=1 colspan=1>397.9</td><td rowspan=1 colspan=1>113.9</td></tr><tr><td rowspan=1 colspan=1>/people/.../location_of_ceremony</td><td rowspan=1 colspan=1>132.0</td><td rowspan=1 colspan=1>108.4</td></tr><tr><td rowspan=1 colspan=1>/sports/../team</td><td rowspan=1 colspan=1>83.1</td><td rowspan=1 colspan=1>104.5</td></tr><tr><td rowspan=1 colspan=1>/user/.../subject</td><td rowspan=1 colspan=1>495.0</td><td rowspan=1 colspan=1>104.2</td></tr></table>
372
+
373
+ Table 7: Top 10 relations with smallest/largest box size in FB15k.
374
+
375
+ <table><tr><td rowspan=1 colspan=1>Top10 relationswithsmallestbox size</td><td rowspan=1 colspan=1>#Ent</td><td rowspan=1 colspan=1>Box size</td></tr><tr><td rowspan=1 colspan=1>/architecture/.../owner</td><td rowspan=1 colspan=1>1.0</td><td rowspan=1 colspan=1>2.3</td></tr><tr><td rowspan=1 colspan=1>/base/.../dog_breeds</td><td rowspan=1 colspan=1>2.0</td><td rowspan=1 colspan=1>4.0</td></tr><tr><td rowspan=1 colspan=1>/education/.../campuses</td><td rowspan=1 colspan=1>1.0</td><td rowspan=1 colspan=1>4.3</td></tr><tr><td rowspan=1 colspan=1>/education/.../educational_institution</td><td rowspan=1 colspan=1>1.0</td><td rowspan=1 colspan=1>4.6</td></tr><tr><td rowspan=1 colspan=1>/base/.../collective</td><td rowspan=1 colspan=1>1.0</td><td rowspan=1 colspan=1>5.1</td></tr><tr><td rowspan=1 colspan=1>/base/.../.member</td><td rowspan=1 colspan=1>1.0</td><td rowspan=1 colspan=1>5.1</td></tr><tr><td rowspan=1 colspan=1>/people/.../appointed_by</td><td rowspan=1 colspan=1>1.0</td><td rowspan=1 colspan=1>5.2</td></tr><tr><td rowspan=1 colspan=1>/base/../fashion_models_with_this_hair_color</td><td rowspan=1 colspan=1>2.0</td><td rowspan=1 colspan=1>5.2</td></tr><tr><td rowspan=1 colspan=1>/fictional_universe/.../parents</td><td rowspan=1 colspan=1>1.0</td><td rowspan=1 colspan=1>5.5</td></tr><tr><td rowspan=1 colspan=1>/american_football/.../team</td><td rowspan=1 colspan=1>2.0</td><td rowspan=1 colspan=1>6.7</td></tr></table>
376
+
377
+ I MRR RESULTS
378
+
379
+ <table><tr><td>Method</td><td>Avg</td><td>1p</td><td>2p 3p</td><td>2i</td><td>3i</td><td></td><td>ip</td><td>pi</td><td>2u up</td></tr><tr><td colspan="8">FB15k</td></tr><tr><td>Q2B</td><td>0.41</td><td>0.654 0.373</td><td>0.274</td><td>0.488</td><td>0.602</td><td>0.194</td><td>0.339</td><td>0.468</td><td>0.301</td></tr><tr><td>GQE</td><td>0.328</td><td>0.505</td><td>0.320 0.218</td><td>0.439</td><td>0.536</td><td>0.139</td><td>0.272</td><td>0.3</td><td>0.244</td></tr><tr><td>GQE-DOUBLE</td><td>0.326</td><td>0.49 0.3</td><td>0.222</td><td>0.438</td><td>0.532</td><td>0.142</td><td>0.28</td><td>0.285</td><td>0.242</td></tr><tr><td colspan="10">FB15k-237</td></tr><tr><td>Q2B</td><td>0.235 0.4</td><td></td><td>0.225</td><td>0.173</td><td>0.275 0.378</td><td>0.105</td><td></td><td>0.18</td><td>0.198</td><td>0.178</td></tr><tr><td>GQE</td><td>0.203</td><td>0.346</td><td>0.193</td><td>0.145</td><td>0.25</td><td>0.355</td><td>0.086</td><td>0.156</td><td>0.145</td><td>0.151</td></tr><tr><td>GQE-DOUBLE</td><td>0.205</td><td>0.346</td><td>0.191</td><td>0.144</td><td>0.258</td><td>0.361</td><td>0.087</td><td>0.164</td><td>0.144</td><td>0.149</td></tr><tr><td colspan="9">NELL995</td></tr><tr><td>Q2B</td><td>0.254</td><td>0.413</td><td>0.227</td><td>0.208 0.288</td><td>0.414</td><td></td><td>0.125</td><td>0.193</td><td>0.266</td><td>0.155</td></tr><tr><td>GQE</td><td>0.21</td><td>0.311</td><td>0.193</td><td>0.175</td><td>0.273</td><td>0.399</td><td>0.078</td><td>0.168</td><td>0.159</td><td>0.13</td></tr><tr><td>GQE-DOUBLE</td><td>0.211</td><td>0.309</td><td>0.192</td><td>0.174</td><td>0.275</td><td>0.408</td><td>0.08</td><td>0.17</td><td>0.156</td><td>0.129</td></tr></table>
380
+
381
+ Table 8: MRR results of QUERY2BOX vs. GQE on FB15k, FB15k-237 and NELL995.
382
+
383
+ <table><tr><td>Method</td><td>Avg</td><td>1p</td><td>2p</td><td>3p</td><td>2i</td><td>3i</td><td>ip</td><td>pi</td><td>2u</td><td>up</td></tr><tr><td colspan="9">FB15k</td></tr><tr><td>Q2B</td><td>0.41</td><td>0.654</td><td>0.373</td><td>0.274</td><td>0.488</td><td>0.602</td><td>0.194</td><td>0.339</td><td>0.468</td><td>0.301</td></tr><tr><td>Q2B-AVG</td><td>0.396</td><td>0.648</td><td>0.368</td><td>0.27</td><td>0.476</td><td>0.564</td><td>0.182</td><td>0.295</td><td>0.465</td><td>0.3</td></tr><tr><td>Q2B-DEEPSETS</td><td>0.402</td><td>0.631</td><td>0.371</td><td>0.269</td><td>0.499</td><td>0.605</td><td>0.181</td><td>0.325</td><td>0.437</td><td>0.298</td></tr><tr><td>Q2B-AVG-1P</td><td>0.324</td><td>0.688</td><td>0.236</td><td>0.159</td><td>0.378</td><td>0.435</td><td>0.122</td><td>0.225</td><td>0.498</td><td>0.178</td></tr><tr><td>Q2B-SHAREDOFFSET</td><td>0.296</td><td>0.511</td><td>0.273</td><td>0.199</td><td>0.351</td><td>0.444</td><td>0.132</td><td>0.233</td><td>0.311</td><td>0.213</td></tr><tr><td colspan="11">FB15k-237</td></tr><tr><td>Q2B</td><td></td><td>0.235 0.4</td><td>0.225</td><td>0.173</td><td>0.275</td><td>0.378</td><td>0.105</td><td>0.18</td><td>0.198</td><td>0.178</td></tr><tr><td>Q2B-AVG</td><td>0.219</td><td>0.398</td><td>0.222</td><td>0.171</td><td>0.236</td><td>0.328</td><td>0.1</td><td>0.145</td><td>0.193</td><td>0.177</td></tr><tr><td>Q2B-DEEPSETS</td><td>0.23</td><td>0.395</td><td>0.224</td><td>0.172</td><td>0.264</td><td>0.372</td><td>0.101</td><td>0.168</td><td>0.194</td><td>0.176</td></tr><tr><td>Q2B-AVG-1P</td><td>0.196</td><td>0.41</td><td>0.18</td><td>0.122</td><td>0.217</td><td>0.274</td><td>0.085</td><td>0.127</td><td>0.209</td><td>0.145</td></tr><tr><td>Q2B-SHAREDOFFSET</td><td>0.18</td><td>0.328</td><td>0.18</td><td>0.131 NELL995</td><td>0.207</td><td>0.289</td><td>0.083</td><td>0.136</td><td>0.135</td><td>0.132</td></tr><tr><td colspan="11"></td></tr><tr><td>Q2B</td><td>0.254</td><td>0.413</td><td>0.227</td><td>0.208</td><td>0.288</td><td>0.414</td><td>0.125</td><td>0.193</td><td>0.266</td><td>0.155</td></tr><tr><td>Q2B-AVG</td><td>0.235</td><td>0.406</td><td>0.219</td><td>0.2</td><td>0.251</td><td>0.342</td><td>0.114</td><td>0.174</td><td>0.259</td><td>0.149</td></tr><tr><td>Q2B-DEEPSETS</td><td>0.246</td><td>0.405</td><td>0.226</td><td>0.207</td><td>0.275</td><td>0.403</td><td>0.107</td><td>0.182</td><td>0.256</td><td>0.153</td></tr><tr><td>Q2B-AVG-1P</td><td>0.227</td><td>0.468</td><td>0.191</td><td>0.16</td><td>0.234</td><td>0.275</td><td>0.094</td><td>0.162</td><td>0.332</td><td>0.125</td></tr><tr><td>Q2B-SHAREDOFFSET</td><td>0.196</td><td>0.318</td><td>0.187</td><td>0.172</td><td>0.228</td><td>0.312</td><td>0.098</td><td>0.156</td><td>0.169</td><td>0.127</td></tr></table>
384
+
385
+ Table 9: MRR results of QUERY2BOX vs. several variants on FB15k, FB15k-237 and NELL995.
parse/train/BJgr4kSFDS/BJgr4kSFDS_content_list.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/BJgr4kSFDS/BJgr4kSFDS_middle.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/BJgr4kSFDS/BJgr4kSFDS_model.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/By5ugjyCb/By5ugjyCb.md ADDED
@@ -0,0 +1,354 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # PACT: PARAMETERIZED CLIPPING ACTIVATION FOR QUANTIZED NEURAL NETWORKS
2
+
3
+ Anonymous authors Paper under double-blind review
4
+
5
+ # ABSTRACT
6
+
7
+ Deep learning algorithms achieve high classification accuracy at the expense of significant computation cost. To address this cost, a number of quantization schemes have been proposed - but most of these techniques focused on quantizing weights, which are relatively smaller in size compared to activations. This paper proposes a novel quantization scheme for activations during training - that enables neural networks to work well with ultra low precision weights and activations without any significant accuracy degradation. This technique, PArameterized Clipping acTivation (PACT), uses an activation clipping parameter $\alpha$ that is optimized during training to find the right quantization scale. PACT allows quantizing activations to arbitrary bit precisions, while achieving much better accuracy relative to published state-of-the-art quantization schemes. We show, for the first time, that both weights and activations can be quantized to 4-bits of precision while still achieving accuracy comparable to full precision networks across a range of popular models and datasets. We also show that exploiting these reduced-precision computational units in hardware can enable a super-linear improvement in inferencing performance due to a significant reduction in the area of accelerator compute engines coupled with the ability to retain the quantized model and activation data in on-chip memories.
8
+
9
+ # 1 INTRODUCTION
10
+
11
+ Deep Convolutional Neural Networks (CNNs) have achieved remarkable accuracy for tasks in a wide range of application domains including image processing (He et al. (2016b)), machine translation (Gehring et al. (2017)), and speech recognition (Zhang et al. (2017)). These state-of-the-art CNNs use very deep models, consuming 100s of ExaOps of computation during training and GBs of storage for model and data. This poses a tremendous challenge to widespread deployment, especially in resource constrained edge environments - leading to a plethora of explorations in compressed models that minimize memory footprint and computation while preserving model accuracy as much as possible.
12
+
13
+ Recently, a whole host of different techniques have been proposed to alleviate these computational costs. Among them, reducing the bit-precision of key CNN data structures, namely weights and activations, has gained attention due to its potential to significantly reduce both storage requirements and computational complexity. In particular, several weight quantization techniques (Li & Liu (2016) and Zhu et al. (2017)) showed significant reduction in the bit-precision of CNN weights with limited accuracy degradation. However, prior work (Hubara et al. (2016b); Zhou et al. (2016)) has shown that a straightforward extension of weight quantization schemes to activations incurs significant accuracy degradation in large-scale image classification tasks such as ImageNet (Russakovsky et al. (2015)). Recently, activation quantization schemes based on greedy layer-wise optimization were proposed (Park et al. (2017); Graham (2017); Cai et al. (2017)), but achieve limited accuracy improvement.
14
+
15
+ In this paper, we propose a novel activation quantization technique, PArameterized Clipping acTivation function (PACT), that automatically optimizes the quantization scales during model training. PACT allows significant reductions in the bit-widths needed to represent both weights and activations and opens up new opportunities for trading off hardware complexity with model accuracy.
16
+
17
+ The primary contributions of this work include:
18
+
19
+ 1) PACT: A new activation quantization scheme for finding the optimal quantization scale during training. We introduce a new parameter $\alpha$ that is used to represent the clipping level in the activation function and is learnt via back-propagation. $\alpha$ sets the quantization scale smaller than ReLU to reduce the quantization error, but larger than a conventional clipping activation function (used in previous schemes) to allow gradients to flow more effectively. In addition, regularization is applied to $\alpha$ in the loss function to enable faster convergence. We provide reasoning and analysis on the expected effectiveness of PACT in preserving model accuracy.
20
+
21
+ 3) Quantitative results demonstrating the effectiveness of PACT on a spectrum of models and datasets. Empirically, we show that: (a) for extremely low bit-precision ( $\leq 2$ -bits for weights and activations), PACT achieves the highest model accuracy compared to all published schemes and (b) 4-bit quantized CNNs based on PACT achieve accuracies similar to single-precision floating point representations.
22
+ 4) System performance analysis to demonstrate the trade-offs in hardware complexity for different bit representations vs. model accuracy. We show that a dramatic reduction in the area of the computing engines is possible and use it to estimate the achievable system-level performance gains.
23
+
24
+ The rest of the paper is organized as follows: Section 2 provides a summary of related prior work on quantized CNNs. Challenges in activation quantization are presented in Section 3. We present PACT, our proposed solution for activation quantization in Section 4. In Section 5 we demonstrate the effectiveness of PACT relative to prior schemes using experimental results on popular CNNs. Overall system performance analysis for a representative hardware system is presented in Section 6 demonstrating the observed trade-offs in hardware complexity for different bit representations.
25
+
26
+ # 2 RELATED WORK
27
+
28
+ Recently, a whole host of different techniques have been proposed to minimize CNN computation and storage costs. One of the earliest studies in weight quantization schemes (Hwang & Sung (2014) and Courbariaux et al. (2015)) show that it is indeed possible to quantize weights to 1-bit (binary) or 2-bits (ternary), enabling an entire DNN model to fit effectively in resource-constrained platforms (e.g., mobile devices). Effectiveness of weight quantization techniques has been further improved (Li & Liu (2016) and Zhu et al. (2017)), by ternarizing weights using statistical distribution of weight values or by tuning quantization scales during training. However, gain in system performance is limited when only weights are quantized while activations are left in high precision. This is particularly severe in convolutional neural networks (CNNs) since weights are relatively smaller in convolution layers in comparison to fully-connected (FC) layers.
29
+
30
+ To reduce the overhead of activations, prior work (Kim & Smaragdis (2015),Hubara et al. (2016a), and Rastegari et al. (2016)) proposed the use of fully binarized neural networks where activations are quantized using 1-bit as well. More recently, activation quantization schemes using more general selections in bit-precision (Hubara et al. (2016b); Zhou et al. (2016; 2017); Mishra et al. (2017); Mellempudi et al. (2017)) have been studied. However, these techniques show significant degradation in accuracy $( > 1 \% )$ ) for ImageNet tasks (Russakovsky et al. (2015)) when bit precision is reduced significantly $( \leq 2 - b i t s )$ . Improvements to previous logarithmic quantization schemes (Miyashita et al. (2016)) using modified base and offset based on “weighted entropy” of activations have also been studied (Park et al. (2017)). Graham (2017) recommends that normalized activation, in the process of batch normalization (Ioffe & Szegedy (2015), BatchNorm), is a good candidate for quantization. Cai et al. (2017) further exploits the statistics of activations and proposes variants of the ReLU activation function for better quantization. However, such schemes typically rely on local (and greedy) optimizations, and are therefore not adaptable or optimized effectively during training. This is further elaborated in Section 3 where we present a detailed discussion on the challenges in quantizing activations.
31
+
32
+ # 3 CHALLENGES IN ACTIVATION QUANTIZATION
33
+
34
+ Quantization of weights is equivalent to discretizing the hypothesis space of the loss function with respect to the weight variables. Therefore, it is indeed possible to compensate weight quantization errors during model training (Hwang & Sung, 2014; Courbariaux et al., 2015). Traditional activation functions, on the other hand, do not have any trainable parameters, and therefore the errors arising from quantizing activations cannot be directly compensated using back-propagation.
35
+
36
+ ![](images/94f27288f313248c619b02ff5a6b052c192ea2e5e2cc38266fe8232c7548f060.jpg)
37
+ Figure 1: (a) Training error, (b) Validation error across epochs for different activation functions (relu and clipping) with and without quantization for the ResNet20 model using the CIFAR10 dataset
38
+
39
+ Activation quantization becomes even more challenging when ReLU (the activation function most commonly used in CNNs) is used as the layer activation function (ActFn). ReLU allows gradient of activations to propagate through deep layers and therefore achieves superior accuracy relative to other activation functions (Nair & Hinton (2010)). However, as the output of the ReLU function is unbounded, the quantization after ReLU requires a high dynamic range (i.e., more bit-precision). In Fig. 1 we present the training and validation errors of ResNet20 with the CIFAR10 dataset using ReLU and show that accuracy is significantly degraded with ReLU quantizations
40
+
41
+ It has been shown that this dynamic range problem can be alleviated by using a clipping activation function, which places an upper-bound on the output (Hubara et al. (2016b); Zhou et al. (2016)). However, because of layer to layer and model to model differences - it is difficult to determine a globally optimal clipping value. In addition, as shown in Fig 1, even though the training error obtained using clipping with quantization is less than that obtained with quantized ReLU, the validation error is still noticeably higher than the baseline.
42
+
43
+ Recently, this challenge has been partially addressed by applying a half-wave Gaussian quantization scheme to activations (Cai et al. (2017)). Based on the observation that activation after BatchNorm normalization is close to a Gaussian distribution with zero mean and unit variance, they used Lloyd’s algorithm to find the optimal quantization scale for this Gaussian distribution and use that scale for every layer. However, this technique also does not fully utilize the strength of backpropagation to optimally learn the clipping level because all the quantization parameters are determined offline and remain fixed throughout the training process.
44
+
45
+ # 4 PACT: PARAMETERIZED CLIPPING ACTIVATION FUNCTION
46
+
47
+ Building on these insights, we introduce PACT, a new activation quantization scheme in which the ActFn has a parameterized clipping level, $\alpha , \alpha$ is dynamically adjusted via gradient descent-based training with the objective of minimizing the accuracy degradation arising from quantization. In PACT, the conventional ReLU activation function in CNNs is replaced with the following:
48
+
49
+ $$
50
+ y = P A C T ( x ) = 0 . 5 ( | x | - | x - \alpha | + \alpha ) = { \left\{ \begin{array} { l l } { 0 , } & { x \in ( - \infty , 0 ) } \\ { x , } & { x \in [ 0 , \alpha ) } \\ { \alpha , } & { x \in [ \alpha , + \infty ) } \end{array} \right. }
51
+ $$
52
+
53
+ where $\alpha$ limits the range of activation to $[ 0 , \alpha ]$ . The truncated activation output is then linearly quantized to $k$ bits for the dot-product computations, where
54
+
55
+ $$
56
+ y _ { q } = r o u n d ( y \cdot \frac { 2 ^ { k } - 1 } { \alpha } ) \cdot \frac { \alpha } { 2 ^ { k } - 1 }
57
+ $$
58
+
59
+ With this new activation function, $\alpha$ is a variable in the loss function, whose value can be optimized
60
+ during training. For back-propagation, gradient $\frac { \partial y _ { q } } { \partial \alpha }$ can be computed using the Straight-Through α ∂
61
+ Estimator (STE) (Bengio et al. (2013)) to estimate $\frac { \partial y _ { q } } { \partial y }$ as 1. Thus,
62
+
63
+ $$
64
+ \frac { \partial y _ { q } } { \partial \alpha } = \frac { \partial y _ { q } } { \partial y } \frac { \partial y } { \partial \alpha } = \left\{ 0 , \quad x \in ( - \infty , \alpha ) \right.
65
+ $$
66
+
67
+ The larger the $\alpha$ , the more the parameterized clipping function resembles a ReLU Actfn. To avoid large quantization errors due to a wide dynamic range, we include a L2-regularizer for $\alpha$ in the loss function. Fig. 7 illustrates how the value of $\alpha$ changes during full-precision training of CIFAR10- ResNet20 starting with an initial value of 10 and using the L2-regularizer. It can be observed that $\alpha$ converges to values much smaller than the initial value as the training epochs proceed, thereby limiting the dynamic range of activations and minimizing quantization loss.
68
+
69
+ To provide further reasoning on why PACT works, we provide in-depth analysis in Appendix A and B. In particular, we show in Appendix A that PACT is as expressive as ReLU when it is used as an activation function. Further we explain in Appendix B that PACT finds a balancing point between clipping and quantization errors to minimize their impact to classification accuracy.
70
+
71
+ ![](images/24eee5fdad3cf00cbed13cfe4398773da8eac5db0520bcc6de6f143f1c71e7d9.jpg)
72
+ Figure 2: Evolution of $\alpha$ values during training using a ResNet20 model on the CIFAR10 dataset.
73
+
74
+ 4.1 UNDERSTANDING HOW PARAMETERIZED CLIPPING WORKS
75
+
76
+ When activation is quantized, the overall behavior of network parameters is affected by the quantization error during training. To observe the impact of activation quantization during network training, we sweep the clipping parameter $\alpha$ and record the training loss with and without quantization. Figs. $^ { 3 \mathrm { ~ a , b ~ } }$ and 3c show cross-entropy and training loss (cross entropy $^ +$ regularization), respectively, over a range of $\alpha$ for the pre-trained SVHN network. The loaded network is trained with the proposed quantization scheme in which ReLU is replaced with the proposed parameterized clipping ActFn for each of its seven convolution layers. We sweep the value of $\alpha$ one layer at a time, keeping all other parameters (weight $( W )$ , bias $( b )$ , BatchNorm parameters $( \beta , \gamma )$ , and the $\alpha$ of other layers) fixed when computing the cross-entropy and training loss.
77
+
78
+ The cross-entropy computed via full-precision forward-pass of training is shown in Fig. 7b. In this case, the cross-entropy converges to a small value in many layers as $\alpha$ increases, indicating that ReLU is a good activation function when no quantization is applied. But even for the full-precision case, training clipping parameter $\alpha$ may help reduce the cross-entropy for certain layers; for example, ReLU (i.e., $\alpha = \infty$ ) is not optimal for act0 and act6 layers.
79
+
80
+ ![](images/95d536930073867b4483ea128db6a84b39449633c4e7c99fcbe77b3bfdf6f3ae.jpg)
81
+ Figure 3: Cross-entropy vs $\alpha$ for SVHN image classification.
82
+
83
+ Next, the cross-entropy computed with quantization in the forward-pass is shown in Fig. 3b. With quantization, the cross-entropy increases in most cases as $\alpha$ increases, implying that ReLU is no longer effective. We also observe that the optimal $\alpha$ has different ranges for different layers, motivating the need to "learn" the quantization scale via training. In addition, we observe plateaus of cross-entropy for the certain ranges of $\alpha$ (e.g., act6), leading to difficulties for gradient descent-based training.
84
+
85
+ Finally, in Fig. 3c, we show the total training loss including both the cross-entropy discussed above and the cost from $\alpha$ regularization. The regularization effectively gets rid of the plateaus in the training loss, thereby favoring convergence for gradient-descent based training. At the same time, $\alpha$ regularization does not perturb the global minimum point. For example, the solid circles in Fig. 3c, which are the optimal $\alpha$ extracted from the pre-trained model, are at the minimum of the training loss curves. The regularization coefficient, $\lambda _ { \alpha }$ , discussed in the next section, is an additional hyper-parameter which controls the impact of regularization on $\alpha$ .
86
+
87
+ # 4.2 EXPLORATION OF HYPER-PARAMETERS
88
+
89
+ For this new quantization approach, we studied the scope of $\alpha$ , the choice of initial values of $\alpha$ ,and the impact of regularizing $\alpha$ . We briefly summarize our findings below, and present more detailed analysis in Appendix C.
90
+
91
+ From our experiments, the best scope for $\alpha$ was to share $\alpha$ per layer. This choice also reduces hardware complexity because $\alpha$ needs to be multiplied only once after all multiply-accumulate (MAC) operations in reduced-precision in a layer are completed.
92
+
93
+ Among initialization choices for $\alpha$ , we found it to be advantageous to initialize $\alpha$ to a larger value relative to typical values of activation, and then apply regularization to reduce it during training.
94
+
95
+ Finally, we observed that applying L2-regularization for $\alpha$ with the same regularization parameter $\lambda$ used for weight works reasonably well. We also observed that, as expected, the optimal value for $\lambda _ { \alpha }$ slightly decreases when higher bit-precision is used because more quantization levels result in higher resolution for activation quantization.
96
+
97
+ Additionally, we follow the practice of many other quantized CNN studies (e.g., Hubara et al. (2016b); Zhou et al. (2016)), and do not quantize the first and last layers, as these have been reported to significantly impact accuracy.
98
+
99
+ # 5 EXPERIMENTS
100
+
101
+ We implemented PACT in Tensorflow (Abadi et al. (2015)) using Tensorpack (Zhou et al. (2016)). To demonstrate the effectiveness of PACT, we studied several well-known CNNs. The following is a summary of the Dataset-Network for the tested CNNs. More implementation details can be found in Appendix.D. Note that the baseline networks use the same hyper-parameters and ReLU activation functions as described in the references. For PACT experiments, we only replace ReLU into PACT but the same hyper-parameters are used. All the time the networks are trained from scratch.
102
+
103
+ • CIFAR10-ResNet20 (CIFAR10, Krizhevsky & Hinton (2010)): a convolution (CONV) layer followed by 3 ResNet blocks (16 CONV layers with 3x3 filter) and a final fully-connected (FC) layer.
104
+ • SVHN-SVHN (SVHN, Netzer et al. (2011)): 7 CONV layers followed by 1 FC layer.
105
+ • IMAGENET-AlexNet (AlexNet, Krizhevsky et al. (2012)): 5 parallel-CONV layers followed by 3 FC layers. BatchNorm is used before ReLU.
106
+ • IMAGENET-ResNet18 (ResNet18, He et al. (2016b)): a CONV layer followed by 8 ResNet blocks (16 CONV layers with 3x3 filter) and a final FC layer. "full pre-activation" ResNet structure (He et al. (2016a)) is employed. IMAGENET-ResNet50 (ResNet50, He et al. (2016b)): a CONV layer followed by 16 ResNet “bottleneck” blocks (total 48 CONV layers) and a final FC layer. "full pre-activation" ResNet structure (He et al. (2016a)) is employed.
107
+
108
+ For comparisons, we include accuracy results reported in the following prior work: DoReFa (Zhou et al. (2016)), BalancedQ (Zhou et al. (2017)), WRPN (Mishra et al. (2017)), FGQ (Mellempudi et al. (2017)), WEP (Park et al. (2017)), LPBN (Graham (2017)), and HWGQ (Cai et al. (2017)). Detailed experimental setting for each of these papers, as well as full comparison of accuracy (top-1 and top5) for AlexNet, ResNet18, ResNet50, can be found in Appendix E. In the following section, we present key results demonstrating the effectiveness of PACT relative to prior work.
109
+
110
+ # 5.1 ACTIVATION QUANTIZATION PERFORMANCE
111
+
112
+ We first evaluate our activation quantization scheme using various CNNs. Fig 4 shows training and validation error of PACT for the tested CNNs. Overall, the higher the bit-precision, the closer the training/validation errors are to the full-precision reference. Specifically it can be seen that training using bit-precision higher than 3-bits converges almost identically to the full-precision baseline. The final validation error has less than $1 \%$ difference relative to the full-precision validation error for all cases when the activation bit-precision is at least 4-bits.
113
+
114
+ We further compare activation quantization performance with 3 previous schemes, DoReFa, LPBN, and HWGQ. We use accuracy degradation as the quantization performance metric, which is calculated as the difference between full-precision accuracy and the accuracy for each quantization bit-precision. Fig. 4f shows accuracy degradation (top-1) for ResNet18 (left) and ResNet50 (right) for increasing activation bit-precision, when the same weight bit-precision is used for each quantization scheme (indicated within the parenthesis). Overall, we observe that accuracy degradation is reduced as we increase the bit-precision of activations. For both ResNet18 and ResNet50, PACT achieves consistently lower accuracy degradation compared to the other quantization schemes, demonstrating the robustness of PACT relative to prior quantization approaches.
115
+
116
+ # 5.2 PACT PERFORMANCE FOR QUANTIZED CNNS
117
+
118
+ In this section, we demonstrate that although PACT targets activation quantization, it does not preclude us from using weight quantization as well. We used PACT to quantize activation of CNNs, and DoReFa scheme to quantize weights. Table 1 summarizes top-1 accuracy of PACT for the tested CNNs (CIFAR10, SVHN, AlexNet, ResNet18, and ResNet50). We also show the accuracy of CNNs when both the weight and activation are quantized by DoReFa’s scheme. As can be seen, with 4 bit precision for both weights and activation, PACT achieves full-precision accuracy consistently across the networks tested. To the best of our knowledge, this is the lowest bit precision for both weights and activation ever reported, that can achieve near $( \leq 1 \% )$ ) full-precision accuracy.
119
+
120
+ We further compare the performance of PACT-based quantized CNNs with 7 previous quantization schemes (DoReFa, BalancedQ, WRPN, FGQ, WEP, LPBN, and HWGQ). Fig. 5 shows comparison of accuracy degradation (top-1) for AlexNet, ResNet18, and ResNet50. Overall, the accuracy degradation decreases as bit-precision for activation or weight increases. For example, in Fig. 5a, the accuracy degradation decreases when activation bit-precision increases given the same weight precision or when weight bit-precision increases given the same activation bit-precision. PACT outperforms other schemes for all the cases. In fact, AlexNet even achieves marginally better accuracy (i.e., negative accuracy degradation) using PACT instead of full-precision.
121
+
122
+ ![](images/c74fa0982258f08adf4cd0ebfa579b8f6b5a2d68a1a6ffc1f50a2a007794a8e9.jpg)
123
+ Figure 4: (a-e) Training and valid error with different bit-precision for various CNNs. (f) Comparison of accuracy degradation for ResNet18 (left) and ResNet50 (right). The lower the better.
124
+
125
+ # 6 SYSTEM-LEVEL PERFORMANCE GAIN
126
+
127
+ In this section, we demonstrate the gain in system performance as a result of the reduction in bit-precision achieved using PACT-CNN. To this end, as shown in Fig. 6(a), we consider a DNN accelerator system comprising of a DNN accelerator chip, comprising of multiple cores, interfaced with an external memory. Each core consists of a 2D-systolic array of fixed-point multiply-andaccumulate (MAC) processing elements on which DNN layers are executed. Each core also contains an on-chip memory, which stores the operands that are fed into the MAC processing array.
128
+
129
+ To estimate system performance at different bit precisions, we studied different versions of the DNN accelerator each comprising the same amount of on-chip memory, external memory bandwidth, and occupying iso-silicon area. First, using real hardware implementations in a state of the art technology $( 1 4 \ \mathrm { n m } \ C \mathbf { M } \mathrm { O S } )$ ), we accurately estimate the reduction in the MAC area achieved by aggressively scaling bit precision. As shown in Fig. 6(b), we achieve ${ \sim } 1 4 \times$ improvement in density when the bit-precisions of both activations and weights are uniformly reduced from 16 bits to 2 bits.
130
+
131
+ ![](images/4dcfa903472db5c7573e7800510c4cc2aa65c7ce905cc7a86488c0430497c653.jpg)
132
+ Figure 5: Comparison of accuracy degradation (Top-1) for (a) AlexNet, (b) ResNet18, and (c) ResNet50.
133
+
134
+ ![](images/e6451310701b441e3e0ae7c0969a1ceb3bbe0a91ddf267387753617ba93f9cb3.jpg)
135
+ Figure 6: (a)System architecture and parameters, (b) Variation in MAC area with bit-precision and (b) Speedup at different quantizations for inference using ResNet50 DNN
136
+
137
+ Next, to translate the reduction in area to improvement in overall performance, we built a precisionconfigurable MAC unit, whose bit precision can be modulated dynamically. The peak compute capability (FLOPs) of the MAC unit varied such that we achieve iso-area at each precision. Note that the total on-chip memory and external bandwidth remains constant at all precisions. We estimate the overall system performance using DeepMatrix, a detailed performance modelling framework for DNN accelerators (Venkataramani et al.).
138
+
139
+ Fig. 6(c) shows the gain in inference performance for the ResNet50 DNN benchmark. We study the performance improvement using different external memory bandwidths, namely, a bandwidth unconstrained system (infinite memory bandwidth) and two bandwidth constrained systems at 32 and 64 GBps. In the bandwidth unconstrained scenario, the gain in performance is limited by how amenable it is to parallelize the work. In this case, we see a near-linear increase in performance for upto 4 bits and a small drop at extreme quantization levels (2 bits).
140
+
141
+ Practical systems, whose bandwidths are constrained, (surprisingly) exhibit a super-linear growth in performance with quantization. For example, when external bandwidth is limited to 64 GBps, quantizing from 16 to 4 bits leads to a $4 \times$ increase in peak FLOPs but a $4 . 5 \times$ improvement in performance. This is because, the total amount of on-chip memory remains constant, and at very low precision some of the data-structures begin to fit within the memory present in the cores, thereby avoiding data transfers from the external memory. Consequently, in bandwidth limited systems, reducing the amount of data transferred from off-chip can provide an additional boost in system performance beyond the increase in peak FLOPs. Note that for the 4 and 2 bit precision configurations, we still used 8 bit precision to execute the first and last layers of the DNN. If we are able to quantize the first and last layers as well to 4 or 2 bits, we estimate an additional $1 . 2 4 \times$ improvement in performance, motivating the need to explore ways to quantize the first and last layers.
142
+
143
+ Table 1: Comparison of top-1 accuracy between DoReFa and PACT. Weights are quantized with DoReFa scheme, whereas activations are quantized with our scheme. Note that CNNs with 4b quantization based on our scheme achieves full-precision accuracy for all the CNNs we explored.
144
+
145
+ <table><tr><td rowspan=2 colspan=1>Network</td><td rowspan=2 colspan=1>FullPrec</td><td rowspan=1 colspan=4>DoReFa</td><td rowspan=1 colspan=4>PACT</td></tr><tr><td rowspan=1 colspan=1>2b</td><td rowspan=1 colspan=1>3b</td><td rowspan=1 colspan=1>4b</td><td rowspan=1 colspan=1>5b</td><td rowspan=1 colspan=1>2b</td><td rowspan=1 colspan=1>3b</td><td rowspan=1 colspan=1>4b</td><td rowspan=1 colspan=1>5b</td></tr><tr><td rowspan=1 colspan=1>CIFAR10</td><td rowspan=1 colspan=1>0.916</td><td rowspan=1 colspan=1>0.882</td><td rowspan=1 colspan=1>0.899</td><td rowspan=1 colspan=1>0.905</td><td rowspan=1 colspan=1>0.904</td><td rowspan=1 colspan=1>0.897</td><td rowspan=1 colspan=1>0.911</td><td rowspan=1 colspan=1>0.913</td><td rowspan=1 colspan=1>0.917</td></tr><tr><td rowspan=1 colspan=1>SVHN</td><td rowspan=1 colspan=1>0.978</td><td rowspan=1 colspan=1>0.976</td><td rowspan=1 colspan=1>0.976</td><td rowspan=1 colspan=1>0.975</td><td rowspan=1 colspan=1>0.975</td><td rowspan=1 colspan=1>0.977</td><td rowspan=1 colspan=1>0.978</td><td rowspan=1 colspan=1>0.978</td><td rowspan=1 colspan=1>0.979</td></tr><tr><td rowspan=1 colspan=1>AlexNet</td><td rowspan=1 colspan=1>0.551</td><td rowspan=1 colspan=1>0.536</td><td rowspan=1 colspan=1>0.550</td><td rowspan=1 colspan=1>0.549</td><td rowspan=1 colspan=1>0.549</td><td rowspan=1 colspan=1>0.550</td><td rowspan=1 colspan=1>0.556</td><td rowspan=1 colspan=1>0.557</td><td rowspan=1 colspan=1>0.557</td></tr><tr><td rowspan=1 colspan=1>ResNet18</td><td rowspan=1 colspan=1>0.702</td><td rowspan=1 colspan=1>0.626</td><td rowspan=1 colspan=1>0.675</td><td rowspan=1 colspan=1>0.681</td><td rowspan=1 colspan=1>0.684</td><td rowspan=1 colspan=1>0.644</td><td rowspan=1 colspan=1>0.681</td><td rowspan=1 colspan=1>0.692</td><td rowspan=1 colspan=1>0.698</td></tr><tr><td rowspan=1 colspan=1>ResNet50</td><td rowspan=1 colspan=1>0.769</td><td rowspan=1 colspan=1>0.671</td><td rowspan=1 colspan=1>0.699</td><td rowspan=1 colspan=1>0.714</td><td rowspan=1 colspan=1>0.714</td><td rowspan=1 colspan=1>0.722</td><td rowspan=1 colspan=1>0.753</td><td rowspan=1 colspan=1>0.765</td><td rowspan=1 colspan=1>0.767</td></tr></table>
146
+
147
+ # 7 CONCLUSION
148
+
149
+ In this paper, we propose a novel activation quantization scheme based on the PArameterized Clipping acTivation function (PACT). The proposed scheme replaces ReLU with an activation function with a clipping parameter, $\alpha$ , that is optimized via gradient descent based training. We provide analysis on why PACT outperforms ReLU when quantization is applied during training. Extensive empirical evaluation using several popular convolutional neural networks, such as CIFAR10, SVHN, AlexNet, ResNet18 and ResNet50, shows that PACT quantizes activations very effectively while simultaneously allowing weights to be heavily quantized. In comparison to all previous quantization schemes, we show that both weights and activations can be quantized much more aggressively (down to 4-bits) - while achieving near $( \leq 1 \% )$ ) full-precision accuracy. In addition, we have shown that the area savings from using reduced-precision MAC units enable a dramatic increase in the number of accelerator cores in the same area, thereby, significantly improving overall system-performance.
150
+
151
+ # REFERENCES
152
+
153
+ Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dan Mané, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Viégas, Oriol Vinyals, Pete Warden, Martin Wattenberg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems, 2015. URL https://www.tensorflow.org/. Software available from tensorflow.org.
154
+
155
+ Yoshua Bengio, Nicholas Léonard, and Aaron C. Courville. Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation. CoRR, abs/1308.3432, 2013.
156
+
157
+ Zhaowei Cai, Xiaodong He, Jian Sun, and Nuno Vasconcelos. Deep Learning With Low Precision by Half-Wave Gaussian Quantization. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017.
158
+
159
+ Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. BinaryConnect: Training Deep Neural Networks with binary weights during propagations. CoRR, abs/1511.00363, 2015.
160
+
161
+ Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N. Dauphin. Convolutional Sequence to Sequence Learning. CoRR, abs/1705.03122, 2017.
162
+
163
+ Benjamin Graham. Low-Precision Batch-Normalized Activations. CoRR, abs/1702.08231, 2017.
164
+
165
+ Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity Mappings in Deep Residual Networks. CoRR, abs/1603.05027, 2016a.
166
+
167
+ Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778, 2016b.
168
+
169
+ Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Binarized Neural Networks. NIPs, pp. 4107–4115, 2016a.
170
+
171
+ Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Quantized Neural Networks: Training Neural Networks with Low Precision Weights and Activations. CoRR, abs/1609.07061, 2016b.
172
+
173
+ Kyuyeon Hwang and Wonyong Sung. Fixed-point Feedforward Deep Neural Network Design Using Weights $+ 1$ , 0, and -1. In IEEE Workshop on Signal Processing Systems (SiPS), pp. 1–6, Oct. 2014.
174
+
175
+ Sergey Ioffe and Christian Szegedy. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. In Proceedings of the 32nd International Conference on Machine Learning, pp. 448–456, 2015.
176
+
177
+ Minje Kim and Paris Smaragdis. Bitwise Neural Networks. ICML Workshop on Resource-Efficient Machine Learning, 2015.
178
+
179
+ Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. International Conference on Learning Representations (ICLR), 2015.
180
+
181
+ Alex Krizhevsky and G Hinton. Convolutional deep belief networks on cifar-10. Unpublished manuscript, 40, 2010.
182
+
183
+ Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. ImageNet Classification with Deep Convolutional Neural Networks. In Advances in Neural Information Processing Systems 25 (NIPS), pp. 1097–1105, 2012.
184
+
185
+ Fengfu Li and Bin Liu. Ternary Weight Networks. CoRR, abs/1605.04711, 2016.
186
+
187
+ Naveen Mellempudi, Abhisek Kundu, Dheevatsa Mudigere, Dipankar Das, Bharat Kaul, and Pradeep Dubey. Ternary Neural Networks with Fine-Grained Quantization. CoRR, abs/1705.01462, 2017.
188
+
189
+ Asit K. Mishra, Jeffrey J. Cook, Eriko Nurvitadhi, and Debbie Marr. WRPN: Training and Inference using Wide Reduced-Precision Networks. CoRR, abs/1704.03079, 2017.
190
+
191
+ Daisuke Miyashita, Edward H. Lee, and Boris Murmann. Convolutional Neural Networks using Logarithmic Data Representation. CoRR, abs/1603.01025, 2016.
192
+
193
+ Vinod Nair and Geoffrey E. Hinton. Rectified Linear Units Improve Restricted Boltzmann Machines. 27th International Conference on Machine Learning (ICML), pp. 807–814, 2010.
194
+
195
+ Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. In NIPS workshop on deep learning and unsupervised feature learning, number 2, pp. 5, 2011.
196
+
197
+ Eunhyeok Park, Junwhan Ahn, and Sungjoo Yoo. Weighted-Entropy-Based Quantization for Deep Neural Networks. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017.
198
+
199
+ Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks. CoRR, abs/1603.05279, 2016.
200
+
201
+ Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115 (3):211–252, 2015.
202
+
203
+ Swagath Venkataramani, Jungwook Choi, Vijayalakshmi Srinivasan, Kailash Gopalakrishnan, and Leland Chang. POSTER: Design Space Exploration for Performance Optimization of Deep Neural Networks on Shared Memory Accelerators. Proc. PACT 2017.
204
+
205
+ Ying Zhang, Mohammad Pezeshki, Philemon Brakel, Saizheng Zhang, César Laurent, Yoshua Bengio, and Aaron C. Courville. Towards End-to-End Speech Recognition with Deep Convolutional Neural Networks. CoRR, abs/1701.02720, 2017.
206
+
207
+ Shuchang Zhou, Zekun Ni, Xinyu Zhou, He Wen, Yuxin Wu, and Yuheng Zou. DoReFa-Net: Training Low Bitwidth Convolutional Neural Networks with Low Bitwidth Gradients. CoRR, abs/1606.06160, 2016.
208
+
209
+ Shuchang Zhou, Yuzhi Wang, He Wen, Qinyao He, and Yuheng Zou. Balanced Quantization: An Effective and Efficient Approach to Quantized Neural Networks. CoRR, abs/1706.07145, 2017.
210
+
211
+ Chenzhuo Zhu, Song Han, Huizi Mao, and William J. Dally. Trained Ternary Quantization. International Conference on Learning Representations (ICLR), 2017.
212
+
213
+ # APPENDIX A PACT IS AS EXPRESSIVE AS RELU
214
+
215
+ When used as an activation function of the neural network, PACT is as expressive as ReLU. This is because clipping parameter, $\alpha$ , introduced in PACT, allows flexibility in adjusting the dynamic range of activation for each layer. We demonstrate in the simple example below that PACT can reach the same solution as ReLU via SGD.
216
+
217
+ Lemma A.1. Consider a single-neuron network with PACT; $x = w \cdot a ,$ , $y = P A C T ( x )$ , where a is input and w is weight. This network can be trained with SGD to find the output the network with ReLU would produce.
218
+
219
+ Proof. Consider a sample of training data $( a , y ^ { * } )$ . For illustration purposes consider mean-squareerror (MSE) as the cost function: $L = 0 . 5 \cdot ( y ^ { * } - y ) ^ { 2 }$ .
220
+
221
+ If $x \leq \alpha$ , then clearly the network with PACT behaves the same as the network with ReLU.
222
+
223
+ If $x > \alpha$ , then $y = \alpha$ and $\begin{array} { r } { \frac { \partial y } { \partial \alpha } = 1 } \end{array}$ from (1). Thus,
224
+
225
+ $$
226
+ { \frac { \partial L } { \partial \alpha } } = { \frac { \partial L } { \partial y } } \cdot { \frac { \partial y } { \partial \alpha } } = { \frac { \partial L } { \partial y } }
227
+ $$
228
+
229
+ Therefore, when $\alpha$ is updated by SGD,
230
+
231
+ $$
232
+ \alpha _ { n e w } = \alpha - \eta \frac { \partial L } { \partial \alpha } = \alpha - \eta \frac { \partial L } { \partial y }
233
+ $$
234
+
235
+ where $\eta$ is a learning rate. Note that during this update, the weight is not updated as $\begin{array} { r } { \frac { \partial L } { \partial w } = \frac { \partial L } { \partial y } \cdot \frac { \partial y } { \partial x } ( = } \end{array}$ $0 ) \cdot a = 0$ .
236
+
237
+ From MSE, $\begin{array} { r } { \frac { \partial L } { \partial y } = ( y - y ^ { * } ) } \end{array}$ . Therefore, if $y ^ { * } > x$ , $\alpha$ is increased for each update of (5) until $\alpha \geq x$ then the PACT network behaves the same as the ReLU network.
238
+
239
+ Interestingly, if $y ^ { * } \leq y$ or $y < y ^ { \ast } < x$ , $\alpha$ is decreased or increased to converge to $y ^ { * }$ . Note that in this case, ReLU would pass erroneous output $x$ to increase cost function, which needs to be fixed by updating $w$ with $\frac { \partial L } { \partial w }$ . PACT, on the other hand, ignores this erroneous output by directly adapting the dynamic range to match the target output $y ^ { * }$ . In this way, the PACT network can be trained to produce output which converges to the same target that the ReLU network would achieve via SGD.
240
+
241
+ In general cases, = Pi ∂L∂yi , and PACT considers output of neurons together to change the dynamic range. There are two options: (1) if output $x _ { i }$ is not clipped, then the network is trained via back-propagation of gradient to update weight, (2) if output $x _ { i }$ is clipped, then $\alpha$ is increased or decreased based on how close the overall output is to the target. Hence, there are configurations under which SGD would lead to a solution close to the one which the network with ReLU would achieve. Fig. 7 demonstrates that ResNet20 with PACT converges almost identical to the network with ReLU.
242
+
243
+ ![](images/920cc4cd2f37d8322b7090187413377c82f79ca044c46a90b782b8c843abefcf.jpg)
244
+ Figure 7: (a) Training error and (b) validation error of PACT for ResNet20 model on the CIFAR10 dataset. Note that the convergence curve for PACT is almost identical to ReLU, although the dynamic range via trained clipping levels are much lower than ReLU.
245
+
246
+ # APPENDIX B PACT FOR BALANCING CLIPPING AND QUANTIZATION ERRORS
247
+
248
+ In Section 3, when we briefly discussed the challenges in activation quantization, we mentioned that there is a trade-off between errors due to clipping and quantization. As the clipping level increases, larger range of activation can be passed to the next layer of the neural network causing less clipping error $( E r r C l i p _ { i } = m a x ( x _ { i } - \alpha , 0 ) )$ . However, the increased dynamic range incurs larger quantization error, since its magnitude is proportional to the clipping level $\begin{array} { r } { ( E r r Q u a n t { i } \le 0 . 5 \cdot \frac { \alpha } { 2 ^ { k } - 1 } } \end{array}$ , with $k$ -bit quantization). This imposes the challenge of finding a proper clipping level to balance between clipping and quantization errors.
249
+
250
+ This trade-off can be better observed in Fig. 8a, which shows normalized mean-square-error caused by clipping and quantization during training of the CIFAR10-ResNet20 with different clipping levels. It can be seen that activation functions with large dynamic range, such as ReLU, would suffer quantization errors whose magnitude increases exponentially as the bit-precision $k$ decreases. This explains why the network with ReLU fails to converge when the activation is quantized (Fig. 1).
251
+
252
+ PACT can find a balancing point between clipping and quantization errors. As explained in Section A, PACT adjusts dynamic range based on how close the output is to the target. As both clipping and quantization errors distort output far from the target, PACT would increase or decrease the dynamic range during training to minimize both clipping and quantization errors.
253
+
254
+ Fig. 8b shows how PACT balances the clipping and quantization errors during training. CIFAR10- ResNet20 is trained with clipping activation function with varying clipping level $\alpha$ from 1 to 16. When activation is quantized, the network trained with clipping activation shows significant accuracy degradation as $\alpha$ increases. This is consistent with the trend in quantization error we observed in Fig. 8a. In this case, PACT achieves the best accuracy one of the clipping activation could achieve, but without exhaustively sweeping over different clipping levels. In other words, PACT auto-tunes the clipping level to achieve best accuracy without incurring significant computation overhead. PACT’s auto-tuning of dynamic range is critical in efficient yet robust training of large scale quantized neural networks, especially because it does not increase the burden for hyper-parameter tuning. In fact, we used the same hyper-parameters as well as the original network structure for all the models we tested, except replacing ReLU to PACT, when we applied activation quantization.
255
+
256
+ Without quantization, there is a trend that validation error decreases as $\alpha$ increases. Surprisingly, some of the cases even outperforms the ReLU network. In this case, PACT also achieves comparable accuracy as ReLU, confirming its expressivity discussed in Section A.
257
+
258
+ # APPENDIX C EXPLORATION OF HYPER-PARAMETERS AND DESIGN CHOICES
259
+
260
+ In this section, we present details on the hyper-parameters and design choices studied for PACT.
261
+
262
+ ![](images/72a58ba6a8bc7bb2d72dd8f3432e22656c53043e2753fbe732daa5fc378656fc.jpg)
263
+ Figure 8: Experiment on CIFAR10-ResNet20 to validate that PACT balances clipping and quantization errors. (a) Trade-off between clipping and quantization error. (b) PACT achieving lowest validation error that clipping activation can achieve without exhaustive search over clipping level $\alpha$ .
264
+
265
+ # C.1 SCOPE OF $\alpha$
266
+
267
+ One of key questions is the optimal scope for $\alpha$ . In other words, determining which neuron activations should share the same $\alpha$ . We considered 3 possible choices: (a) Individual $\alpha$ for each neuron activation, (b) Shared $\alpha$ among neurons within the same output channel, and (c) Shared $\alpha$ within a layer. We empirically studied each of these choices of $\alpha$ (without quantization) using CIFAR10- ResNet20 and determined training and validation error for PACT. As shown in Fig. 9, sharing $\alpha$ per layer is the best choice in terms of accuracy. This is in fact a preferred option from the perspective of hardware complexity as well, since $\alpha$ needs to be multiplied only once after all multiply-accumulate(MAC) operations in a layer are completed.
268
+
269
+ ![](images/5064d22ae3b8e904a8a501bfd423ee93151390bdb4af2bb05f46748ee8c687c3.jpg)
270
+ Figure 9: Training and validation error of CIFAR10-ResNet20 for PACT with different scope of $\alpha$
271
+
272
+ C.2 INITIAL VALUE AND REGULARIZATION OF $\alpha$
273
+
274
+ The optimization behavior of $\alpha$ can be explained from the formulation of the parameterized clipping function. From Eq. 3 it is clear that, if $\alpha$ is initialized to a very small value, more activations fall into the range for the nonzero gradient, leading to unstable $\alpha$ in the early epochs, potentially causing accuracy degradation. On the other hand, if $\alpha$ is initialized to a very large value, the gradient becomes too small and $\alpha$ may be stuck at a large value, potentially suffering more on quantization error. Therefore, it is intuitive to start with a reasonably large value to cover a wide dynamic range and avoid unstable adaptation of $\alpha$ , but apply regularizer to reduce the value of $\alpha$ so as to alleviate quantization error.
275
+
276
+ In practice, we found that applying L2-regularization for $\alpha$ while setting its coefficient $\lambda _ { \alpha }$ the same as the L2-regularization coefficient for weight, $\lambda$ , works well. Fig. 10 shows that validation error for
277
+
278
+ PACT-quantized CIFAR10-ResNet20 does not significantly vary for a wide range of $\lambda _ { \alpha }$ . We also observed that, as expected, the optimal value for $\lambda _ { \alpha }$ slightly decreases when higher bit-precision is used because more quantization levels result in higher resolution for activation quantization.
279
+
280
+ ![](images/32d3b4919d38720aad499c650d70a85b8662e57226181fc4abc07980dfa105e8.jpg)
281
+ Figure 10: Training and validation error of quantized CIFAR10-ResNet20 for PACT with different regularization parameter $\lambda _ { \alpha }$ .
282
+
283
+ # C.3 QUANTIZATION OF FIRST AND LAST LAYERS
284
+
285
+ Many previous work (e.g., Hubara et al. (2016b); Zhou et al. (2016)) follow the convention to keep the first and last layer in full precision during training, since quantizing those layers lead to substantial accuracy degradation. We empirically studied this for the proposed quantization approach for CIFAR10-ResNet20. In Fig. 11, the only difference among the curves is whether input activation and weight of the first convolution layer or the last fully-connected layer are quantized. As can be seen from the plots, there can be noticeable accuracy degradation if the first or last layers are aggressively quantized. But computation in floating point is very expensive in hardware.
286
+
287
+ Therefore, we further studied the option of quantizing the first and last layers with higher quantization bit-precision than the bit-precision of the other layers. Table 2 shows that independent of the quantization level for the other layers, there is little accuracy degradation if the first and last layer are quantized with 8-bits. This motivates us to employ reduced precision computation even for the first/last layers.
288
+
289
+ ![](images/a1a8ea51973e630e5fdbec0577c25ee0c0451ccd68988e512a0503001a27e854.jpg)
290
+ Figure 11: Comparison of accuracy of CIFAR10-ResNet20 with and without quantization of the first and last layers.
291
+
292
+ # APPENDIX D CNN IMPLEMENTATION DETAILS
293
+
294
+ In this section, we summarize details of our CNN implementation as well as our training settings, which is based on the default networks provided by Tensorpack (Zhou et al. (2016)). Unless mentioned otherwise, ReLU following BatchNorm is used for ActFn of the convolution (CONV)
295
+
296
+ Table 2: Validation error (in $\%$ ) of CIFAR10-ResNet20 when first and last layers are quantized with different bit-precision. FL/M/FL means the first and last layers are quantized with Bit-FL bits, while the other layers are quantized with Bit-M bits. NQ represents no quantization is applied.
297
+
298
+ <table><tr><td rowspan=1 colspan=1>BIT-M (bits)</td><td rowspan=1 colspan=4>2</td><td rowspan=1 colspan=4>3</td><td rowspan=1 colspan=4>4</td><td rowspan=1 colspan=4>5</td></tr><tr><td rowspan=1 colspan=1>BIT-FL (bits)</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>8</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>8</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>8</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>8</td><td rowspan=1 colspan=1>32</td></tr><tr><td rowspan=1 colspan=1>FL/M/FL</td><td rowspan=1 colspan=1>21.0</td><td rowspan=1 colspan=1>12.9</td><td rowspan=1 colspan=1>11.1</td><td rowspan=1 colspan=1>10.9</td><td rowspan=1 colspan=1>17.4</td><td rowspan=1 colspan=1>10.0</td><td rowspan=1 colspan=1>9.4</td><td rowspan=1 colspan=1>8.9</td><td rowspan=1 colspan=1>15.9</td><td rowspan=1 colspan=1>9.7</td><td rowspan=1 colspan=1>9.2</td><td rowspan=1 colspan=1>8.9</td><td rowspan=1 colspan=1>18.2</td><td rowspan=1 colspan=1>9.0</td><td rowspan=1 colspan=1>8.4</td><td rowspan=1 colspan=1>8.5</td></tr><tr><td rowspan=1 colspan=1>FL/M/NQ</td><td rowspan=1 colspan=1>21.3</td><td rowspan=1 colspan=1>11.5</td><td rowspan=1 colspan=1>11.5</td><td rowspan=1 colspan=1>10.7</td><td rowspan=1 colspan=1>17.6</td><td rowspan=1 colspan=1>9.7</td><td rowspan=1 colspan=1>9.2</td><td rowspan=1 colspan=1>9.0</td><td rowspan=1 colspan=1>16.5</td><td rowspan=1 colspan=1>9.7</td><td rowspan=1 colspan=1>8.7</td><td rowspan=1 colspan=1>8.7</td><td rowspan=1 colspan=1>16.3</td><td rowspan=1 colspan=1>9.3</td><td rowspan=1 colspan=1>8.6</td><td rowspan=1 colspan=1>8.5</td></tr><tr><td rowspan=1 colspan=1>NQ/M/FL</td><td rowspan=1 colspan=1>12.1</td><td rowspan=1 colspan=1>11.2</td><td rowspan=1 colspan=1>11.0</td><td rowspan=1 colspan=1>11.5</td><td rowspan=1 colspan=1>9.8</td><td rowspan=1 colspan=1>8.9</td><td rowspan=1 colspan=1>9.2</td><td rowspan=1 colspan=1>9.2</td><td rowspan=1 colspan=1>8.4</td><td rowspan=1 colspan=1>8.4</td><td rowspan=1 colspan=1>8.7</td><td rowspan=1 colspan=1>8.8</td><td rowspan=1 colspan=1>8.5</td><td rowspan=1 colspan=1>9.0</td><td rowspan=1 colspan=1>8.5</td><td rowspan=1 colspan=1>8.5</td></tr></table>
299
+
300
+ layers, and Softmax is used for the fully-connected (FC) layer. Note that the baseline networks use the same hyper-parameters and ReLU activation functions as described in the references. For PACT experiments, we only replace ReLU into PACT but the same hyper-parameters are used. All the time the networks are trained from scratch.
301
+
302
+ The CIFAR10 dataset (Krizhevsky & Hinton (2010)) is an image classification benchmark containing $3 2 \times 3 2$ pixel RGB images. It consists of 50K training and 10K test image sets. We used the “standard” ResNet structure (He et al. (2016a)) which consists of a CONV layer followed by 3 ResNet blocks (16 CONV layers with 3x3 filter) and a final FC layer. We used stochastic gradient descent (SGD) with momentum of 0.9 and learning rate starting from 0.1 and scaled by 0.1 at epoch 60, 120. L2-regularizer with decay of 0.0002 is applied to weight. The mini-batch size of 128 is used, and the maximum number of epochs is 200.
303
+
304
+ The SVHN dataset (Netzer et al. (2011)) is a real-world digit recognition dataset containing photos of house numbers in Google Street View images, where the “cropped” $3 2 \times 3 2$ colored images (resized to $4 0 \times 4 0$ as input to the network) centered around a single character are used. It consists of 73257 digits for training and 26032 digits for testing. We used a CNN model which contains 7 CONV layers followed by 1 FC layer. We used ADAM(Kingma & Ba (2015)) with epsilon $1 0 ^ { - 5 }$ and learning rate starting from $1 0 ^ { - 3 }$ and scaled by 0.5 every 50 epoch. L2-regularizer with decay of $1 0 ^ { - 7 }$ is applied to weight. The mini-batch size of 128 is used, and the maximum number of epochs is 200.
305
+
306
+ The IMAGENET dataset (Russakovsky et al. (2015)) consists of 1000-categories of objects with over 1.2M training and 50K validation images. Images are first resized to 256 256 and randomly cropped to 224224 prior to being used as input to the network. We used a modified AlexNet, ResNet18 and ResNet50.
307
+
308
+ We used AlexNet network (Krizhevsky et al. (2012)) in which local contrast renormalization (RNorm) layer is replaced with BatchNorm layer. We used ADAM with epsilon $1 0 ^ { - 5 }$ and learning rate starting from $\mathrm { 1 0 ^ { - 4 } }$ and scaled by 0.2 at epoch 56 and 64. L2-regularizer with decay factor of $5 \times 1 0 ^ { - 6 }$ is applied to weight. The mini-batch size of 128 is used, and the maximum number of epochs is 100.
309
+
310
+ ResNet18 consists of a CONV layer followed by 8 ResNet blocks (16 CONV layers with 3x3 filter) and a final FC layer. "full pre-activation" ResNet structure (He et al. (2016a)) is employed. ResNet50 consists of a CONV layer followed by 16 ResNet “bottleneck” blocks (total 48 CONV layers) and a final FC layer. "full pre-activation" ResNet structure (He et al. (2016a)) is employed.
311
+
312
+ For both ResNet18 and ResNet50, we used stochastic gradient descent (SGD) with momentum of 0.9 and learning rate starting from 0.1 and scaled by 0.1 at epoch 30, 60, 85, 95. L2-regularizer with decay of $1 0 ^ { - 4 }$ is applied to weight. The mini-batch size of 256 is used, and the maximum number of epochs is 110.
313
+
314
+ # APPENDIX E COMPARISON WITH RELATED WORK
315
+
316
+ # E.1 QUANTIZATION EXPERIMENT SETTING
317
+
318
+ • DoReFa-Net (DoReFa, Zhou et al. (2016)): A general bit-precision uniform quantization schemes for weight, activation, and gradient of DNN training.We compared the experimental results of DoReFa for CIFAR10, SVHN, AlexNet and ResNet18 under the same experimental
319
+
320
+ setting as PACT. Note that a clipped absolute activation function is used for SVHN in DoReFa.
321
+ Balanced Quantization (BalancedQ, Zhou et al. (2017)): A quantization scheme based on recursive partitioning of data into balanced bins. We compared the reported top-1/top-5 validation accuracy of their quantization scheme for AlexNet and ResNet18.
322
+ Quantization using Wide Reduced-Precision Networks (WRPN, Mishra et al. (2017)): A scheme to increase the number of filter maps to increase robustness for activation quantization. We compared the reported top-1 accuracy of their quantization with various weight/activation bit-precision for AlexNet.
323
+ • Fine-grained Quantization (FGQ, Mellempudi et al. (2017)): A direct quantization scheme (i.e., little re-training needed) based on fine-grained grouping (i.e., within a small subset of filter maps). We compared the reported top-1 validation accuracy of their quantization with 2-bit weight and 4-bit activation for AlexNet and ResNet50.
324
+ • Weighted-entropy-based quantization (WEP, Park et al. (2017)): A quantization scheme that considers statistics of weight/activation. We compared the top-1/top-5 reported accuracy of their quantization with various bit-precision for AlexNet, where the first and last layers are not quantized.
325
+ Low-precision batch normalization (LPBN, Graham (2017)): A scheme for activation quantization in the process of batch normalization. We compared the top-1/top-5 reported accuracy of their quantization with 3-5 bit precision for activation. The first layer activation is not quantized.
326
+ Half-wave Gaussian quantization (HWGQ, Cai et al. (2017)): A quantization scheme that finds the scale via Lloyd search on Normal distribution. We compared the top-1/top-5 reported accuracy for their quantization with 1-bit weight and varying activation bit-precision for AlexNet, and 2-bit weight for ResNet18 and ResNet50. The first and last layers are not quantized.
327
+
328
+ # E.2 COMPARISON OF ACCURACY
329
+
330
+ In this section, we present full comparison of accuracy (top-1 and top-5) of the tested CNNs (AlexNet, ResNet18, ResNet50) for image classification on IMAGENET dataset. All the data points for PACT and DoReFa are obtained by running experiments on Tensorpack. All the other data points are accuracy reported in the corresponding papers. As can be seen, PACT achieves the best accuracy across the board for various flavors of quantization. We also observe that using PACT for activation quantization enables more aggressive weight quantization without loss in accuracy.
331
+
332
+ Table 3: Comparison of Top-1 accuracy (in $\%$ ) for AlexNet. Bold entries indicate the lowest accuracy degradation compared to single-precision reference from each work. Baseline (full-precision) accuracy for PACT is $5 5 . 1 \%$ .
333
+
334
+ <table><tr><td rowspan=1 colspan=1>BitW</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>5</td></tr><tr><td rowspan=1 colspan=1>BitA</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>5</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>5</td></tr><tr><td rowspan=1 colspan=1>WRPN</td><td rowspan=1 colspan=1>57.2</td><td rowspan=1 colspan=1>52.7</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>54.4</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>51.3</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>50.5</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>52.4</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>54.4</td><td rowspan=1 colspan=1></td></tr><tr><td rowspan=1 colspan=1>BalancedQ</td><td rowspan=1 colspan=1>57.1</td><td rowspan=1 colspan=1>56.5</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>55.7</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td></tr><tr><td rowspan=1 colspan=1>FGQ</td><td rowspan=1 colspan=1>56.8</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>49.0</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td></tr><tr><td rowspan=1 colspan=1>WEQ</td><td rowspan=1 colspan=1>57.1</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>50.6</td><td rowspan=1 colspan=1>53.7</td><td rowspan=1 colspan=1>54.4</td><td rowspan=1 colspan=1>51.8</td><td rowspan=1 colspan=1>54.9</td><td rowspan=1 colspan=1>55.5</td><td rowspan=1 colspan=1>52.3</td><td rowspan=1 colspan=1>55.1</td><td rowspan=1 colspan=1>55.9</td><td rowspan=1 colspan=1></td></tr><tr><td rowspan=1 colspan=1>DoReFa</td><td rowspan=1 colspan=1>55.1</td><td rowspan=1 colspan=1>54.1</td><td rowspan=1 colspan=1>55.1</td><td rowspan=1 colspan=1>54.8</td><td rowspan=1 colspan=1>54.9</td><td rowspan=1 colspan=1>46.4</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>45.0</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>45.1</td><td rowspan=1 colspan=1>45.1</td></tr><tr><td rowspan=1 colspan=1>PACT</td><td rowspan=1 colspan=1>55.1</td><td rowspan=1 colspan=1>54.9</td><td rowspan=1 colspan=1>55.6</td><td rowspan=1 colspan=1>55.5</td><td rowspan=1 colspan=1>55.2</td><td rowspan=1 colspan=1>55.0</td><td rowspan=1 colspan=1>55.4</td><td rowspan=1 colspan=1>55.7</td><td rowspan=1 colspan=1>54.6</td><td rowspan=1 colspan=1>55.6</td><td rowspan=1 colspan=1>55.7</td><td rowspan=1 colspan=1>54.6</td><td rowspan=1 colspan=1>55.7</td><td rowspan=1 colspan=1>55.7</td><td rowspan=1 colspan=1>55.7</td></tr></table>
335
+
336
+ Table 4: Comparison of Top-5 accuracy (in $\%$ ) for AlexNet. Bold entries indicate the lowest accuracy degradation compared to single-precision reference from each work. Baseline (full-precision) accuracy for PACT is $7 7 . 0 \%$ .
337
+
338
+ <table><tr><td rowspan=1 colspan=1>BitW</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>5</td></tr><tr><td rowspan=1 colspan=1>BitA</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>5</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>5</td></tr><tr><td rowspan=1 colspan=1>LogQuant</td><td rowspan=1 colspan=1>78.3</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>77.1</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td></tr><tr><td rowspan=1 colspan=1>WEQ</td><td rowspan=1 colspan=1>80.2</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>75.0</td><td rowspan=1 colspan=1>77.5</td><td rowspan=1 colspan=1>78.0</td><td rowspan=1 colspan=1>76.0</td><td rowspan=1 colspan=1>78.5</td><td rowspan=1 colspan=1>79.1</td><td rowspan=1 colspan=1>76.5</td><td rowspan=1 colspan=1>78.5</td><td rowspan=1 colspan=1>79.2</td><td rowspan=1 colspan=1></td></tr><tr><td rowspan=1 colspan=1>BalancedQ</td><td rowspan=1 colspan=1>79.4</td><td rowspan=1 colspan=1>79.0</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>78.0</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td></tr><tr><td rowspan=1 colspan=1>DoReFa</td><td rowspan=1 colspan=1>77.0</td><td rowspan=1 colspan=1>76.9</td><td rowspan=1 colspan=1>77.9</td><td rowspan=1 colspan=1>77.5</td><td rowspan=1 colspan=1>77.5</td><td rowspan=1 colspan=1>76.8</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>77.8</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>77.5</td><td rowspan=1 colspan=1>77.9</td></tr><tr><td rowspan=1 colspan=1>PACT</td><td rowspan=1 colspan=1>77.0</td><td rowspan=1 colspan=1>77.2</td><td rowspan=1 colspan=1>77.8</td><td rowspan=1 colspan=1>77.6</td><td rowspan=1 colspan=1>77.2</td><td rowspan=1 colspan=1>77.7</td><td rowspan=1 colspan=1>77.9</td><td rowspan=1 colspan=1>78.0</td><td rowspan=1 colspan=1>77.1</td><td rowspan=1 colspan=1>78.0</td><td rowspan=1 colspan=1>78.0</td><td rowspan=1 colspan=1>77.1</td><td rowspan=1 colspan=1>78.0</td><td rowspan=1 colspan=1>78.0</td><td rowspan=1 colspan=1>77.8</td></tr></table>
339
+
340
+ Table 5: Comparison of Top-1 accuracy $( \mathrm { i n \% } )$ ) for ResNet18. Bold entries indicate the lowest accuracy degradation compared to single-precision reference from each work. Baseline (full-precision) accuracy for PACT is ${ \bar { 7 } } 0 . 4 \%$ .
341
+
342
+ <table><tr><td rowspan=1 colspan=1>BitW</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>1</td><td rowspan=1 colspan=1>1</td><td rowspan=1 colspan=1>1</td><td rowspan=1 colspan=1>1</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>5</td></tr><tr><td rowspan=1 colspan=1>BitA</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>5</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>5</td></tr><tr><td rowspan=1 colspan=1>BalancedQ</td><td rowspan=1 colspan=1>68.2</td><td rowspan=1 colspan=1>62.1</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>59.4</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td></tr><tr><td rowspan=1 colspan=1>LPBN</td><td rowspan=1 colspan=1>69.6</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>63.6</td><td rowspan=1 colspan=1>66.7</td><td rowspan=1 colspan=1>69.3</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td></tr><tr><td rowspan=1 colspan=1>HWGQ</td><td rowspan=1 colspan=1>69.6</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>61.3</td><td rowspan=1 colspan=1>57.6</td><td rowspan=1 colspan=1>60.3</td><td rowspan=1 colspan=1>60.8</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td></tr><tr><td rowspan=1 colspan=1>DoReFa</td><td rowspan=1 colspan=1>70.4</td><td rowspan=1 colspan=1>66.9</td><td rowspan=1 colspan=1>68.3</td><td rowspan=1 colspan=1>68.5</td><td rowspan=1 colspan=1>68.7</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>62.6</td><td rowspan=1 colspan=1>67.5</td><td rowspan=1 colspan=1>68.1</td><td rowspan=1 colspan=1>68.4</td></tr><tr><td rowspan=1 colspan=1>PACT</td><td rowspan=1 colspan=1>70.4</td><td rowspan=1 colspan=1>67.5</td><td rowspan=1 colspan=1>69.2</td><td rowspan=1 colspan=1>70.0</td><td rowspan=1 colspan=1>70.0</td><td rowspan=1 colspan=1>65.8</td><td rowspan=1 colspan=1>62.9</td><td rowspan=1 colspan=1>65.3</td><td rowspan=1 colspan=1>65.0</td><td rowspan=1 colspan=1>64.4</td><td rowspan=1 colspan=1>68.1</td><td rowspan=1 colspan=1>69.2</td><td rowspan=1 colspan=1>69.8</td></tr></table>
343
+
344
+ Table 6: Comparison of Top-5 accuracy $( \mathrm { i n \% } )$ ) for ResNet18. Bold entries indicate the lowest accuracy degradation compared to single-precision reference from each work. Baseline (full-precision) accuracy for PACT is ${ \bar { 8 } } 9 . 6 \%$ .
345
+
346
+ <table><tr><td rowspan=1 colspan=1>BitW</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>1</td><td rowspan=1 colspan=1>1</td><td rowspan=1 colspan=1>1</td><td rowspan=1 colspan=1>1</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>5</td></tr><tr><td rowspan=1 colspan=1>BitA</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>5</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>5</td></tr><tr><td rowspan=1 colspan=1>BalancedQ</td><td rowspan=1 colspan=1>87.5</td><td rowspan=1 colspan=1>82.7</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>82.0</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td></tr><tr><td rowspan=1 colspan=1>LPBN</td><td rowspan=1 colspan=1>89.2</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>85.2</td><td rowspan=1 colspan=1>87.5</td><td rowspan=1 colspan=1>88.8</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td></tr><tr><td rowspan=1 colspan=1>HWGQ</td><td rowspan=1 colspan=1>89.2</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>83.6</td><td rowspan=1 colspan=1>81.0</td><td rowspan=1 colspan=1>82.8</td><td rowspan=1 colspan=1>83.4</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td></tr><tr><td rowspan=1 colspan=1>DoReFa</td><td rowspan=1 colspan=1>89.6</td><td rowspan=1 colspan=1>87.3</td><td rowspan=1 colspan=1>88.2</td><td rowspan=1 colspan=1>88.5</td><td rowspan=1 colspan=1>88.6</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>84.4</td><td rowspan=1 colspan=1>87.6</td><td rowspan=1 colspan=1>88.1</td><td rowspan=1 colspan=1>88.3</td></tr><tr><td rowspan=1 colspan=1>PACT</td><td rowspan=1 colspan=1>89.6</td><td rowspan=1 colspan=1>87.6</td><td rowspan=1 colspan=1>88.9</td><td rowspan=1 colspan=1>89.3</td><td rowspan=1 colspan=1>89.3</td><td rowspan=1 colspan=1>86.7</td><td rowspan=1 colspan=1>84.7</td><td rowspan=1 colspan=1>85.9</td><td rowspan=1 colspan=1>85.9</td><td rowspan=1 colspan=1>85.6</td><td rowspan=1 colspan=1>88.2</td><td rowspan=1 colspan=1>89.0</td><td rowspan=1 colspan=1>89.3</td></tr></table>
347
+
348
+ Table 7: Comparison of Top-1 accuracy $( \mathrm { i n \% } )$ ) for ResNet50. Bold entries indicate the lowest accuracy degradation compared to single-precision reference from each work. Baseline (full-precision) accuracy for PACT is $7 6 . 9 \%$ .
349
+
350
+ Table 8: Comparison of Top-5 accuracy $( \mathrm { i n \% } )$ ) for ResNet50. Bold entries indicate the lowest accuracy degradation compared to single-precision reference from each work. Baseline (full-precision) accuracy for PACT is $9 3 . 1 \%$ .
351
+
352
+ <table><tr><td rowspan=1 colspan=1>BitW</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>1</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>5</td></tr><tr><td rowspan=1 colspan=1>BitA</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>5</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>5</td></tr><tr><td rowspan=1 colspan=1>FGQ</td><td rowspan=1 colspan=1>75.1</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>68.4</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td></tr><tr><td rowspan=1 colspan=1>LPBN</td><td rowspan=1 colspan=1>76.0</td><td rowspan=1 colspan=1>56.1</td><td rowspan=1 colspan=1>73.8</td><td rowspan=1 colspan=1>75.6</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td></tr><tr><td rowspan=1 colspan=1>HWGQ</td><td rowspan=1 colspan=1>76.0</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>64.6</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td></tr><tr><td rowspan=1 colspan=1>DoReFa</td><td rowspan=1 colspan=1>76.9</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>67.1</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>69.9</td><td rowspan=1 colspan=1>71.4</td><td rowspan=1 colspan=1>71.4</td></tr><tr><td rowspan=1 colspan=1>PACT</td><td rowspan=1 colspan=1>76.9</td><td rowspan=1 colspan=1>75.5</td><td rowspan=1 colspan=1>75.9</td><td rowspan=1 colspan=1>76.0</td><td rowspan=1 colspan=1>67.8</td><td rowspan=1 colspan=1>72.2</td><td rowspan=1 colspan=1>74.5</td><td rowspan=1 colspan=1>75.3</td><td rowspan=1 colspan=1>76.5</td><td rowspan=1 colspan=1>76.7</td></tr></table>
353
+
354
+ <table><tr><td rowspan=1 colspan=1>BitW</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>1</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>5</td></tr><tr><td rowspan=1 colspan=1>BitA</td><td rowspan=1 colspan=1>32</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>5</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>2</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>3</td><td rowspan=1 colspan=1>4</td><td rowspan=1 colspan=1>5</td></tr><tr><td rowspan=1 colspan=1>LPBN</td><td rowspan=1 colspan=1>93.0</td><td rowspan=1 colspan=1>79.6</td><td rowspan=1 colspan=1>91.8</td><td rowspan=1 colspan=1>92.6</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td></tr><tr><td rowspan=1 colspan=1>HWGQ</td><td rowspan=1 colspan=1>93.0</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>85.9</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td></tr><tr><td rowspan=1 colspan=1>DoReFa</td><td rowspan=1 colspan=1>93.1</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>87.3</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>89.2</td><td rowspan=1 colspan=1>89.8</td><td rowspan=1 colspan=1>93.3</td></tr><tr><td rowspan=1 colspan=1>PACT</td><td rowspan=1 colspan=1>93.1</td><td rowspan=1 colspan=1>92.6</td><td rowspan=1 colspan=1>92.9</td><td rowspan=1 colspan=1>92.9</td><td rowspan=1 colspan=1>87.9</td><td rowspan=1 colspan=1>90.5</td><td rowspan=1 colspan=1>91.9</td><td rowspan=1 colspan=1>92.6</td><td rowspan=1 colspan=1>93.2</td><td rowspan=1 colspan=1>93.3</td></tr></table>
parse/train/By5ugjyCb/By5ugjyCb_content_list.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/By5ugjyCb/By5ugjyCb_middle.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/By5ugjyCb/By5ugjyCb_model.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/BylJUTEKvB/BylJUTEKvB.md ADDED
@@ -0,0 +1,324 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # CROSS-ITERATION BATCH NORMALIZATION
2
+
3
+ Anonymous authors Paper under double-blind review
4
+
5
+ # ABSTRACT
6
+
7
+ A well-known issue of Batch Normalization is its significantly reduced effectiveness in the case of small mini-batch sizes. When a mini-batch contains few examples, the statistics upon which the normalization is defined cannot be reliably estimated from it during a training iteration. To address this problem, we present Cross-Iteration Batch Normalization (CBN), in which examples from multiple recent iterations are jointly utilized to enhance estimation quality. A challenge of computing statistics over multiple iterations is that the network activations from different iterations are not comparable to each other due to changes in network weights. We thus compensate for the network weight changes via a proposed technique based on Taylor polynomials, so that the statistics can be accurately estimated and batch normalization can be effectively applied. On object detection and image classification with small mini-batch sizes, CBN is found to outperform the original batch normalization and a direct calculation of statistics over previous iterations without the proposed compensation technique.
8
+
9
+ # 1 INTRODUCTION
10
+
11
+ Batch Normalization (BN) (Ioffe & Szegedy, 2015) has played a significant role in the success of deep neural networks. It was introduced to address the issue of internal covariate shift, where the distribution of network activations changes during training iterations due to the updates of network parameters. This shift is commonly believed to be disruptive to network training, and BN alleviates this problem through normalization of the network activations by their mean and variance, computed over the examples within the mini-batch at each iteration. With this normalization, network training can be performed at much higher learning rates and with less sensitivity to weight initialization.
12
+
13
+ In BN, it is assumed that the distribution statistics for the examples within each mini-batch reflect the statistics over the full training set. While this assumption is generally valid for large batch sizes, it breaks down in the small batch size regime (Peng et al., 2018; Wu & He, 2018; Ioffe, 2017), where noisy statistics computed from small sets of examples can lead to a dramatic drop in performance. This problem hinders the application of BN to memory-consuming tasks such as object detection (Ren et al., 2015; Dai et al., 2017), semantic segmentation (Long et al., 2015; Chen et al., 2017) and action recognition (Wang et al., 2018b), where batch sizes are limited due to memory constraints.
14
+
15
+ Towards improving estimation of statistics in the small batch size regime, alternative normalizers have been proposed. Several of them, including Layer Normalization (LN) (Ba et al., 2016), Instance Normalization (IN) (Ulyanov et al., 2016), and Group Normalization (GN) (Wu & He, 2018), compute the mean and variance over the channel dimension, independent of batch size. Different channel-wise normalization techniques, however, tend to be suitable for different tasks, depending on the set of channels involved. On the other hand, synchronized BN (SyncBN) (Peng et al., 2018) yields consistent improvements by processing larger batch sizes across multiple GPUs. These gains in performance come at the cost of additional overhead needed for synchronization across the devices.
16
+
17
+ A seldom explored direction for estimating better statistics is to compute them over the examples from multiple recent training iterations, instead of from only the current iteration as done in previous techniques. This can substantially enlarge the pool of data from which the mean and variance are obtained. However, there exists an obvious drawback to this approach, in that the activation values from different iterations are not comparable to each other due to the changes in network weights. As shown in Figure 1, directly calculating the statistics over multiple iterations, which we refer to as Naive CBN, results in lower accuracy.
18
+
19
+ ![](images/026cbf38422de395fdbe5421774471d5517eefca106ac21df3b09ef02e25ece1.jpg)
20
+ Figure 1: Top-1 classification accuracy vs. batch sizes per iteration. The base model is a ResNet-18 (He et al., 2016) trained on ImageNet (Russakovsky et al., 2015). BN (Ioffe & Szegedy, 2015)’s accuracy drops rapidly when the batch size is reduced. GN (Wu & He, 2018) exhibits stable performance but underperforms BN on adequate batch sizes. CBN compensates for the reduced batch size per GPU by exploiting approximated statistics from recent iterations (Temporal window size denotes how many recent iters are utilized for statistics computation). CBN shows relatively stable performance over different batch sizes. Naive CBN does not work well, which directly calculates statistics from recent iterations without compensation.
21
+
22
+ In this paper, we present a method that compensates for the network weight changes among iterations, so that examples from preceding iterations can be effectively used to improve batch normalization. Our method, called Cross-Iteration Batch Normalization (CBN), is motivated by the observation that network weights change gradually, instead of abruptly, between consecutive training iterations, thanks to the iterative nature of Stochastic Gradient Descent (SGD). As a result, the mean and variance of examples from recent iterations can be well approximated for the current network weights via a low-order Taylor polynomial, defined on gradients of the statistics with respect to the network weights. The compensated means and variances from multiple recent iterations are averaged with those of the current iteration to produce better estimates of the statistics.
23
+
24
+ In the small batch size regime, CBN leads to appreciable performance improvements over the original BN, as exhibited in Figure 1. The superiority of our proposed approach is further demonstrated through more extensive experiments on ImageNet classification and object detection on COCO. These gains are obtained with negligible overhead, as the statistics from previous iterations have already been computed and Taylor polynomials are simple to evaluate. With this work, it is shown that cues for batch normalization can successfully be extracted along the time dimension, opening a new direction for investigation.
25
+
26
+ # 2 RELATED WORK
27
+
28
+ The importance of normalization in training neural networks has been recognized for decades (LeCun et al., 1998). In general, normalization can be performed on three components: input data, hidden activations, and network parameters. Among them, input data normalization is used most commonly because of its simplicity and effectiveness (Sola & Sevilla, 1997; LeCun et al., 1998).
29
+
30
+ After the introduction of Batch Normalization (Ioffe & Szegedy, 2015), the normalization of activations has become nearly as prevalent. By normalizing hidden activations by their statistics within each mini-batch, BN effectively alleviates the vanishing gradient problem and significantly speeds up the training of deep networks. To mitigate the mini-batch size dependency of BN, a number of variants have been proposed, including Layer Normalization (LN) (Ba et al., 2016), Instance Normalization (IN) (Ulyanov et al., 2016), Group Normalization (GN) (Wu & He, 2018), and Batch Instance Normalization (BIN) (Nam & Kim, 2018). The motivation of LN is to explore more suitable statistics for sequential models, while IN performs normalization in a manner similar to BN but with statistics only for each instance. GN achieves a balance between IN and LN, by dividing features into multiple groups along the channel dimension and computing the mean and variance within each group for normalization. BIN introduces a learnable method for automatically switching between normalizing and maintaining style information, enjoying the advantages of both BN and IN on style transfer tasks. Cross-GPU Batch Normalization (CGBN or SyncBN) (Peng et al., 2018) extends BN across multiple GPUs for the purpose of increasing the effective batch size. Though providing higher accuracy, it introduces synchronization overhead to the training process. Kalman Normalization (KN) (Wang et al., 2018a) presents a Kalman filtering procedure for estimating the statistics for a network layer from the layer’s observed statistics and the computed statistics of previous layers.
31
+
32
+ Batch Renormalization (BRN) (Ioffe, 2017) is the first attempt to utilize the statistics of recent iterations for normalization. It does not compensate for the statistics from recent iterations, but rather it down-weights the importance of statistics from distant iterations. This down-weighting heuristic, however, does not make the resulting statistics “correct", as the statistics from recent iterations are not of the current network weights. BRN can be deemed as a special version of our Naive CBN baseline (without Taylor polynomial approximation), where distant iterations are down-weighted.
33
+
34
+ ![](images/94ad9ce1a1693e6654fb5403358b512982b490d0b827eb1b63dbb0684ac9829b.jpg)
35
+ Figure 2: Illustration of BN and the proposed Cross-Iteration Batch Normalization (CBN).
36
+
37
+ Recent work have also investigated the normalization of network parameters. In Weight Normalization (WN) (Salimans & Kingma, 2016), the optimization of network weights is improved through a reparameterization of weight vectors into their length and direction. Weight Standardization (WS) (Qiao et al., 2019) instead reparameterizes weights based on their first and second moments for the purpose of smoothing the loss landscape of the optimization problem. To combine the advantages of multiple normalization techniques, Switchable Normalization (SN) (Luo et al., 2018) and Sparse Switchable Normalization (SSN) (Shao et al., 2019) make use of differentiable learning to switch among different normalization methods.
38
+
39
+ The proposed CBN takes an activation normalization approach that aims to mitigate the mini-batch dependency of BN. Different from existing techniques, it provides a way to effectively aggregate statistics across multiple training iterations.
40
+
41
+ # 3 METHOD
42
+
43
+ # 3.1 REVISITING BATCH NORMALIZATION
44
+
45
+ The original batch normalization (BN) (Ioffe & Szegedy, 2015) whitens the activations of each layer by the statistics computed within a mini-batch. Denote $\theta _ { t }$ and $x _ { t , i } ( \theta _ { t } )$ as the network weights and the feature response of a certain layer for the $i$ -th example in the $t$ -th mini-batch. With these values, BN conducts the following normalization:
46
+
47
+ $$
48
+ \hat { x } _ { t , i } ( \theta _ { t } ) = \frac { x _ { t , i } ( \theta _ { t } ) - \mu _ { t } ( \theta _ { t } ) } { \sqrt { \sigma _ { t } ( \theta _ { t } ) ^ { 2 } + \varepsilon } } ,
49
+ $$
50
+
51
+ where $\hat { x } _ { t , i } ( \theta _ { t } )$ is the whitened activation with zero mean and unit variance, $\varepsilon$ is a small constant added for numerical stability, and $\mu _ { t } ( \theta _ { t } )$ and $\sigma _ { t } ( \theta _ { t } )$ are the mean and variance computed for all the examples from the current mini-batch, i.e.,
52
+
53
+ $$
54
+ \begin{array} { l } { \displaystyle \mu _ { t } ( \boldsymbol { \theta } _ { t } ) = \frac { 1 } { m } \sum _ { i = 1 } ^ { m } \boldsymbol { x } _ { t , i } ( \boldsymbol { \theta } _ { t } ) , } \\ { \displaystyle \sigma _ { t } ( \boldsymbol { \theta } _ { t } ) = \sqrt { \frac { 1 } { m } \sum _ { i = 1 } ^ { m } ( \boldsymbol { x } _ { t , i } ( \boldsymbol { \theta } _ { t } ) - \boldsymbol { \mu } _ { t } ( \boldsymbol { \theta } _ { t } ) ) ^ { 2 } } = \sqrt { \boldsymbol { \nu } _ { t } ( \boldsymbol { \theta } _ { t } ) - \boldsymbol { \mu } _ { t } ( \boldsymbol { \theta } _ { t } ) ^ { 2 } } , } \end{array}
55
+ $$
56
+
57
+ where $\begin{array} { r } { \nu _ { t } ( \theta _ { t } ) = \frac { 1 } { m } \sum _ { i = 1 } ^ { m } x _ { t , i } ( \theta _ { t } ) ^ { 2 } } \end{array}$ , and $m$ denotes the number of examples in the current mini-batch. The whitened activation $\hat { x } _ { t , i } ( \theta _ { t } )$ further undergoes a linear transform with learnable weights, to increase its expressive power:
58
+
59
+ $$
60
+ y _ { t , i } ( \theta _ { t } ) = \gamma \hat { x } _ { t , i } ( \theta _ { t } ) + \beta ,
61
+ $$
62
+
63
+ where $\gamma$ and $\beta$ are the learnable parameters (initialized to $\gamma = 1$ and $\beta = 0$ in this work).
64
+
65
+ When the batch size $m$ is small, the statistics $\mu _ { t } ( \theta _ { t } )$ and $\sigma _ { t } ( \theta _ { t } )$ become noisy estimates of the training set statistics, thus degrading the effects of batch normalization. In the ImageNet classification task for which the BN module was originally designed, a batch size of 32 is typical. However, for other tasks requiring larger models and/or higher image resolution, such as object detection, semantic segmentation and video recognition, the typical batch size may be as small as 1 or 2 due to GPU memory limitations. The original BN becomes considerably less effective in such cases.
66
+
67
+ # 3.2 LEVERAGING STATISTICS FROM PREVIOUS ITERATIONS
68
+
69
+ To address the issue of BN with small mini-batches, a naive approach is to compute the mean and variance over the current and previous iterations. However, the statistics $\mu _ { t - \tau } ( \theta _ { t - \tau } )$ and $\nu _ { t - \tau } ( \theta _ { t - \tau } )$ of the $\left( t - \tau \right)$ -th iteration are computed under the network weights $\theta _ { t - \tau }$ , making them obsolete for the current iteration. As a consequence, directly aggregating statistics from multiple iterations produces inaccurate estimates of the mean and variance, leading to significantly worse performance.
70
+
71
+ We observe that the network weights change smoothly between consecutive iterations, due to the nature of gradient-based training. This allows us to approximate $\mu _ { t - \tau } ( \theta _ { t } )$ and $\nu _ { t - \tau } ( \theta _ { t } )$ from the readily available $\mu _ { t - \tau } ( \theta _ { t - \tau } )$ and $\nu _ { t - \tau } ( \theta _ { t - \tau } )$ via a Taylor polynomial, i.e.,
72
+
73
+ $$
74
+ \begin{array} { r } { \mu _ { t - \tau } ( \theta _ { t } ) = \mu _ { t - \tau } ( \theta _ { t - \tau } ) + \frac { \partial \mu _ { t - \tau } ( \theta _ { t - \tau } ) } { \partial \theta _ { t - \tau } } ( \theta _ { t } - \theta _ { t - \tau } ) + \mathbf { O } ( | | \theta _ { t } - \theta _ { t - \tau } | | ^ { 2 } ) , } \\ { \nu _ { t - \tau } ( \theta _ { t } ) = \nu _ { t - \tau } ( \theta _ { t - \tau } ) + \frac { \partial \nu _ { t - \tau } ( \theta _ { t - \tau } ) } { \partial \theta _ { t - \tau } } ( \theta _ { t } - \theta _ { t - \tau } ) + \mathbf { O } ( | | \theta _ { t } - \theta _ { t - \tau } | | ^ { 2 } ) , } \end{array}
75
+ $$
76
+
77
+ where $\partial \mu _ { t - \tau } ( \theta _ { t - \tau } ) / \partial \theta _ { t - \tau }$ and $\partial \nu _ { t - \tau } ( \theta _ { t - \tau } ) / \partial \theta _ { t - \tau }$ are gradients of the statistics with respect to the network weights, and $\mathbf { O } ( | | \theta _ { t } - \theta _ { t - \tau } | | ^ { 2 } )$ denotes higher-order terms of the Taylor polynomial, which can be omitted since the first-order term dominates when $\left( \theta _ { t } - \theta _ { t - \tau } \right)$ is small.
78
+
79
+ In Eq. (5) and Eq. (6), the gradients $\partial \mu _ { t - \tau } ( \theta _ { t - \tau } ) / \partial \theta _ { t - \tau }$ and $\partial \nu _ { t - \tau } ( \theta _ { t - \tau } ) / \partial \theta _ { t - \tau }$ cannot be precisely determined at a negligible cost because the statistics $\mu _ { t - \tau } ^ { l } ( \theta _ { t - \tau } )$ and $\nu _ { t - \tau } ^ { l } ( \theta _ { t - \tau } )$ for a node at the $l$ -th network layer depend on all the network weights prior to the $l$ -th layer, i.e., $\partial \mu _ { t - \tau } ^ { l } ( \theta _ { t - \tau } ) / \partial \theta _ { t - \tau } ^ { r } \neq 0$ and $\partial { \nu _ { t - \tau } ^ { l } } ( \theta _ { t - \tau } ) / \partial \theta _ { t - \tau } ^ { r } \ne 0$ for $r \leq l$ , where $\theta _ { t - \tau } ^ { r }$ denotes the network weights at the $r$ -th layer. Only when $r = l$ can these gradients be derived in closed form efficiently.
80
+
81
+ Empirically, we find that as the layer index $r$ decreases $( r \leq l )$ , the partial gradients $\frac { \partial \mu _ { t } ^ { l } ( \theta _ { t } ) } { \theta _ { t } ^ { r } }$ and $\frac { \partial \nu _ { t } ^ { l } ( \theta _ { t } ) } { \theta _ { t } ^ { r } }$ rapidly diminish. These reduced effects of network weight changes at earlier layers on the activation distributions in later layers may perhaps be explained by the reduced internal covariate shift of BN. Motivated by this phenomenon, which is studied in Appendix C, we propose to truncate these partial gradients at layer $l$ .
82
+
83
+ Thus, we further approximate Eq. (5) and Eq. (6) by
84
+
85
+ $$
86
+ \mu _ { t - \tau } ^ { l } ( \boldsymbol { \theta } _ { t } ) \approx \mu _ { t - \tau } ^ { l } ( \boldsymbol { \theta } _ { t - \tau } ) + \frac { \partial \mu _ { t - \tau } ^ { l } ( \boldsymbol { \theta } _ { t - \tau } ) } { \partial \boldsymbol { \theta } _ { t - \tau } ^ { l } } ( \boldsymbol { \theta } _ { t } ^ { l } - \boldsymbol { \theta } _ { t - \tau } ^ { l } ) ,
87
+ $$
88
+
89
+ $$
90
+ \boldsymbol { v } _ { t - \tau } ^ { l } ( \boldsymbol { \theta } _ { t } ) \approx \boldsymbol { v } _ { t - \tau } ^ { l } ( \boldsymbol { \theta } _ { t - \tau } ) + \frac { \partial \boldsymbol { v } _ { t - \tau } ^ { l } ( \boldsymbol { \theta } _ { t - \tau } ) } { \partial \boldsymbol { \theta } _ { t - \tau } ^ { l } } ( \boldsymbol { \theta } _ { t } ^ { l } - \boldsymbol { \theta } _ { t - \tau } ^ { l } ) .
91
+ $$
92
+
93
+ A naive implementation of $\partial \mu _ { t - \tau } ^ { l } ( \theta _ { t - \tau } ) / \partial \theta _ { t - \tau } ^ { l }$ and $\partial \nu _ { t - \tau } ^ { l } ( \theta _ { t - \tau } ) / \partial \theta _ { t - \tau } ^ { l }$ involves computational overhead of $O ( C ^ { l } \times C ^ { l } \times C ^ { l - 1 } \times K )$ , where $C ^ { l }$ and $C ^ { l - 1 }$ denote the channel dimension of the $l$ -th layer and the $( l - 1 )$ -th layer, respectively, and $K$ denotes the kernel size of $\theta _ { t - \tau } ^ { l }$ . Here we find that the operation can be implemented efficiently in $O ( C ^ { l } \times C ^ { l - 1 } \times K )$ , thanks to the averaging over feature responses of $\mu$ and $\nu$ . See Appendix $\mathbf { B }$ for the details.
94
+
95
+ # 3.3 CROSS-ITERATION BATCH NORMALIZATION
96
+
97
+ After compensating for network weight changes, we aggregate the statistics of the $k - 1$ most recent iterations with those of the current iteration $t$ to obtain the statistics used in CBN:
98
+
99
+ $$
100
+ \begin{array} { r l } & { \bar { \mu } _ { t , k } ^ { l } ( \boldsymbol { \theta } _ { t } ) = \displaystyle \frac { 1 } { k } \sum _ { \tau = 0 } ^ { k - 1 } \mu _ { t - \tau } ^ { l } ( \boldsymbol { \theta } _ { t } ) , } \\ & { \bar { \nu } _ { t , k } ^ { l } ( \boldsymbol { \theta } _ { t } ) = \displaystyle \frac { 1 } { k } \sum _ { \tau = 0 } ^ { k - 1 } \operatorname* { m a x } \left[ \nu _ { t - \tau } ^ { l } ( \boldsymbol { \theta } _ { t } ) , \mu _ { t - \tau } ^ { l } ( \boldsymbol { \theta } _ { t } ) ^ { 2 } \right] , } \end{array}
101
+ $$
102
+
103
+ <table><tr><td></td><td>batch size per iter</td><td>#examples for statistics</td><td>Norm axis</td></tr><tr><td>IN</td><td>#bs/GPU * #GPU</td><td>1</td><td>(spatial)</td></tr><tr><td>LN</td><td>#bs/GPU * #GPU</td><td>1</td><td>(channel, spatial)</td></tr><tr><td>GN</td><td>#bs/GPU * #GPU</td><td>1</td><td>(channel group,spatial)</td></tr><tr><td>BN</td><td>#bs/GPU * #GPU</td><td>#bs/GPU</td><td>(batch,spatial)</td></tr><tr><td>syncBN</td><td>#bs/GPU * #GPU</td><td>#bs/GPU *#GPU</td><td>(batch,spatial, GPU)</td></tr><tr><td>CBN</td><td>#bs/GPU * #GPU</td><td>#bs/GPU J* temporal window</td><td>(batch,spatial, iteration)</td></tr></table>
104
+
105
+ Table 1: Comparison of different feature normalization methods. #bs/GPU denotes batch size per GPU.
106
+
107
+ $$
108
+ \bar { \sigma } _ { t , k } ^ { l } ( \theta _ { t } ) = \sqrt { \bar { \nu } _ { t , k } ^ { l } ( \theta _ { t } ) - \bar { \mu } _ { t , k } ^ { l } ( \theta _ { t } ) ^ { 2 } } ,
109
+ $$
110
+
111
+ where $\mu _ { t - \tau } ^ { l } ( \theta _ { t } )$ and $\nu _ { t - \tau } ^ { l } ( \theta _ { t } )$ are computed from Eq. (7) and Eq. (8). In Eq. (10), $\bar { \nu } _ { t , k } ^ { l } ( \theta _ { t } )$ is determined from the maximum of $\nu _ { t - \tau } ^ { l } ( \theta _ { t } )$ and $\mu _ { t - \tau } ^ { l } ( \theta _ { t } ) ^ { 2 }$ in each iteration because $\nu _ { t - \tau } ^ { l } ( \theta _ { t } ) \geq \mu _ { t - \tau } ^ { l } ( \theta _ { t } ) ^ { 2 }$ should hold for valid statistics but may be violated by Taylor polynomial approximations in Eq. (7) and Eq. (8). Finally, $\bar { \mu } _ { t , k } ^ { l } ( \theta _ { t } )$ and $\bar { \sigma } _ { t , k } ^ { \tilde { l } } ( \theta _ { t } )$ are applied to normalize the corresponding feature responses $\{ x _ { t , i } ^ { l } ( \theta _ { t } ) \} _ { i = 1 } ^ { m }$ at the current iteration:
112
+
113
+ $$
114
+ \hat { x } _ { t , i } ^ { l } ( \theta _ { t } ) = \frac { x _ { t , i } ^ { l } ( \theta _ { t } ) - \bar { \mu } _ { t , k } ^ { l } ( \theta _ { t } ) } { \sqrt { \bar { \sigma } _ { t , k } ^ { l } ( \theta _ { t } ) ^ { 2 } + \varepsilon } } .
115
+ $$
116
+
117
+ With CBN, the effective number of examples used to compute the statistics for the current iteration is $k$ times as large as that for the original BN. In training, the loss gradients are backpropagated to the network weights and activations at the current iteration, i.e., $\mathbf { \bar { \boldsymbol { \theta } } } _ { t } ^ { l }$ and $x _ { t , i } ^ { l } ( \theta _ { t } )$ . Those of the previous iterations are fixed and do not receive gradients. Hence, the computation cost of CBN in back-propagation is the same as that of BN.
118
+
119
+ Replacing the BN modules in a network by CBN leads to only minor increases in computational overhead and memory footprint. For computation, the additional overhead mainly comes from computing the partial derivatives $\partial \mu _ { t - \tau } ( \theta _ { t - \tau } ) / \partial \theta _ { t - \tau } ^ { l }$ and $\partial \nu _ { t - \tau } ( \theta _ { t - \tau } ) / \partial \theta _ { t - \tau } ^ { l }$ , which is insignificant in relation tothe statistics $( \{ \mu _ { t - \tau } ^ { l } ( \theta _ { t - \tau } ) \} _ { \tau = 1 } ^ { k - 1 }$ theand $\{ \nu _ { t - \tau } ^ { l } ( \theta _ { t - \tau } ) \} _ { \tau = 1 } ^ { k - 1 } )$ For memory, the and the gradients $( \{ \partial \mu _ { t - \tau } ( \stackrel { \cdot } { \theta _ { t - \tau } } ) / \partial \theta _ { t - \tau } ^ { l } \} _ { \tau = 1 } ^ { k - 1 }$ and $\{ \partial \nu _ { t - \tau } ( \theta _ { t - \tau } ) / \partial \theta _ { t - \tau } ^ { l } \} _ { \tau = 1 } ^ { k - 1 } )$ τ τ=1 computed for the most recent τ τ=1 iterations, which is also minor compared to the rest of the memory consumed in processing the input examples. The additional computation and memory of CBN is reported for our experiments in Table 6.
120
+
121
+ A key hyper-parameter in the proposed CBN is the temporal window size, $k$ , of recent iterations used for statistics estimation. A broader window enlarges the set of examples, but the example quality becomes increasingly lower for more distant iterations, since the differences in network parameters $\theta _ { t }$ and $\theta _ { t - \tau }$ become more significant and are compensated less well using a low-order Taylor polynomial. Empirically, we found that CBN is effective with a window size up to $k = 8$ in a variety of settings and tasks. The only trick is that the window size should be kept small at the beginning of training, when the network weights change quickly. Thus, we introduce a burn-in period of length $T _ { \mathrm { b u r n } }$ -in for the window size, where $k = 1$ and CBN degenerates to the original BN. In our experiments, the burn-in period is set to 25 epochs on ImageNet image classification and 3 epochs on COCO object detection by default. Ablations on this parameter are presented in the Appendix.
122
+
123
+ Table 1 compares CBN with other feature normalization methods. The key difference among these approaches is the axis along which the statistics are counted and the features are normalized. The previous techniques are all designed to exploit examples from the same iteration. By contrast, CBN explores the aggregation of examples along the temporal dimension. As the data utilized by CBN lies in a direction orthogonal to that of previous methods, the proposed CBN could potentially be combined with other feature normalization approaches to further enhance statistics estimation in certain challenging applications.
124
+
125
+ # 4 EXPERIMENTS
126
+
127
+ # 4.1 IMAGE CLASSIFICATION ON IMAGENET
128
+
129
+ Experimental settings. ImageNet (Russakovsky et al., 2015) is a benchmark dataset for image classification, containing 1.28M training images and 50K validation images from 1000 classes. We follow the standard setting in (He et al., 2015) to train deep networks on the training set and report the single-crop top-1 accuracy on the validation set. Our preprocessing and augmentation strategy strictly follows the GN baseline (Wu & He, 2018). We use a weight decay of 0.0001 for all weight layers, including $\gamma$ and $\beta$ . We train standard ResNet-18 for 100 epochs on 4 GPUs, and decrease the learning rate by the cosine decay strategy (He et al., 2019). We use the average over 5 trials for all results. All hyper-parameters, e.g. group size of GN, are carefully tuned via cross-validation. ResNet-18 with BN is our base model. To compare with other normalization methods, we directly replace BN with IN, LN, GN, BRN, and our proposed CBN.
130
+
131
+ Comparison of feature normalization methods. We compare the performance of each normalization method with a normal batch size, 32, in Table 2. With sufficient data for reliable statistics, BN easily reaches the highest top-1 accuracy. Similar as the results in previous papers
132
+
133
+ <table><tr><td></td><td>IN</td><td>LN</td><td>GN</td><td>BN</td><td>CBN</td></tr><tr><td>Top-1 accuracy</td><td>64.4</td><td>67.9</td><td>68.9</td><td>70.2</td><td>70.2</td></tr></table>
134
+
135
+ Table 2: Top-1 accuracy of feature normalization methods using ResNet-18 on ImageNet.
136
+
137
+ (Wu & He, 2018), IN and LN achieve significantly worse performance than BN. GN works well on image classification, but still has a small degradation of $1 . 2 \%$ compared with BN. Over all the methods, our CBN is the only one that is able to achieve comparable accuracy with BN, as it converges to the procedure of BN as the batch size becomes larger.
138
+
139
+ Sensitivity to batch size. We compare the behavior of CBN, original BN (Ioffe & Szegedy, 2015), GN (Wu & He, 2018), and BRN (Ioffe, 2017) at the same number of images per GPU on ImageNet classification. For CBN, the recent iterations are utilized so as to ensure that the number of effective examples is no fewer than 16. For BRN, the settings strictly follow the original paper. We adopt a learning rate of 0.1 for the batch size of 32, and linearly scale the learning rate by $N / 3 2$ for a batch size of $N$ .
140
+
141
+ Table 3: Top-1 accuracy of feature normalization methods with different batch sizes using ResNet-18 as the base model on ImageNet.
142
+
143
+ <table><tr><td>batch size per GPU</td><td>32</td><td>16</td><td>8</td><td>4</td><td>2</td></tr><tr><td>BN</td><td>70.2</td><td>70.2</td><td>68.4</td><td>65.1</td><td>55.9</td></tr><tr><td>GN</td><td>68.9</td><td>69.0</td><td>68.9</td><td>69.0</td><td>69.1</td></tr><tr><td>BRN</td><td>70.1</td><td>68.5</td><td>68.2</td><td>67.9</td><td>60.3</td></tr><tr><td>CBN</td><td>70.2</td><td>70.2</td><td>70.1</td><td>69.8</td><td>69.3</td></tr></table>
144
+
145
+ The results are shown in Table 3. For the original BN, its accuracy drops noticeably as the number of images per GPU is reduced from 32 to 2. BRN suffers a significant performance drop as well. GN maintains its accuracy by utilizing the channel dimension but not batch dimension. For CBN, its accuracy holds by exploiting the examples of recent iterations. Also, CBN outperforms GN by $0 . 9 \%$ on average top-1 accuracy with different batch sizes. This is reasonable, because the statistics computation of CBN introduces uncertainty caused by the stochastic batch sampling like in BN, but this uncertainty is missing in GN which results in some loss of regularization ability.
146
+
147
+ # 4.2 OBJECT DETECTION AND INSTANCE SEGMENTATION ON COCO
148
+
149
+ Experimental settings. COCO (Lin et al., 2014) is chosen as the benchmark for object detection and instance segmentation. Models are trained on the COCO 2017 train split with $1 1 8 \mathrm { k }$ images, and evaluated on the COCO 2017 validation split with 5k images. Following the standard protocol in (Lin et al., 2014), the object detection and instance segmentation accuracies are measured by the mean average precision (mAP) scores at different intersection-over-union (IoU) overlaps at the box and the mask levels, respectively.
150
+
151
+ Following (Wu & He, 2018), Faster R-CNN (Ren et al., 2015) and Mask R-CNN (He et al., 2017) with FPN (Lin et al., 2017) are chosen as the baselines for object detection and instance segmentation, respectively. For both, the 2fc box head is replaced by a 4conv1fc head for better use of the normalization mechanism (Wu & He, 2018). The backbone networks are ImageNet pretrained ResNet-50 (default) or ResNet-101, with specific normalization. Finetuning is performed on the COCO train set for 12 epochs on 4 GPUs by SGD, where each GPU processes 4 images (default). Note that the mean and variance statistics in CBN are computed within each GPU. The learning rate is initialized to be $0 . 0 2 * N / 1 6$ for a batch size per GPU of $N$ , and is decayed by a factor of 10 at the 9-th and the 11-th epochs. The weight decay and momentum parameters are set to 0.0001 and 0.9, respectively. We use the average over 5 trials for all results. All hyper-parameters, e.g. group size of GN, are carefully tuned via cross-validation.
152
+
153
+ Table 4: Results of feature normalization methods on Faster R-CNN with FPN and ResNet50 on COCO.
154
+
155
+ <table><tr><td>backbone</td><td>box head</td><td>Apbbox</td><td>APbbox 50</td><td></td><td>APbbox S</td><td>APbbox M</td><td>APbbox</td></tr><tr><td>fixed BN</td><td>1</td><td>36.9</td><td>58.2</td><td>39.9</td><td>21.2</td><td>40.8</td><td>46.9</td></tr><tr><td>fixed BN</td><td>BN</td><td>36.3</td><td>57.3</td><td>39.2</td><td>20.8</td><td>39.7</td><td>47.3</td></tr><tr><td>fixed BN</td><td>syncBN</td><td>37.7</td><td>58.5</td><td>41.1</td><td>22.0</td><td>40.9</td><td>49.0</td></tr><tr><td>fixed BN</td><td>GN</td><td>37.8</td><td>59.0</td><td>40.8</td><td>22.3</td><td>41.2</td><td>48.4</td></tr><tr><td>fixed BN</td><td>CBN</td><td>37.7</td><td>59.0</td><td>40.7</td><td>22.1</td><td>40.9</td><td>48.8</td></tr><tr><td>BN</td><td>BN</td><td>35.5</td><td>56.4</td><td>38.7</td><td>19.7</td><td>38.8</td><td>47.3</td></tr><tr><td>syncBN</td><td>syncBN</td><td>37.9</td><td>58.5</td><td>41.1</td><td>21.7</td><td>41.5</td><td>49.7</td></tr><tr><td>GN</td><td>GN</td><td>37.8</td><td>59.1</td><td>40.9</td><td>22.4</td><td>41.2</td><td>49.0</td></tr><tr><td>CBN</td><td>CBN</td><td>37.3</td><td>57.7</td><td>39.3</td><td>21.9</td><td>40.8</td><td>48.2</td></tr></table>
156
+
157
+ <table><tr><td>Backbone</td><td>method</td><td>norm</td><td>Apbbox</td><td>APbox 50</td><td>AP5ox</td><td>APbbox S</td><td>APbox M</td><td>APbbox</td></tr><tr><td rowspan="3">R50+FPN</td><td rowspan="3">Faster RCNN</td><td>GN</td><td>37.8</td><td>59.0</td><td>40.8</td><td>22.3</td><td>41.2</td><td>48.4</td></tr><tr><td>syncBN</td><td>37.7</td><td>58.5</td><td>41.1</td><td>22.0</td><td>40.9</td><td>49.0</td></tr><tr><td>CBN</td><td>37.7</td><td>59.0</td><td>40.7</td><td>22.1</td><td>40.9</td><td>48.8</td></tr><tr><td rowspan="2">R101+FPN</td><td rowspan="2">Faster RCNN</td><td rowspan="2">GN syncBN CBN</td><td>39.3 39.2</td><td>60.6</td><td>42.7 43.0</td><td>22.5</td><td>42.5</td><td>51.3</td></tr><tr><td>39.2</td><td>59.8 60.0</td><td>42.6</td><td>22.2 22.3</td><td>42.9 42.6</td><td>51.6 51.1</td></tr><tr><td rowspan="3">R50+FPN</td><td rowspan="3">Mask RCNN</td><td rowspan="3">GN syncBN</td><td>Apbbox</td><td>APo</td><td>AP5box</td><td>Apmask</td><td>APsk</td><td>APmsk</td></tr><tr><td>38.6</td><td>59.8</td><td>41.9</td><td>35.0</td><td>56.7</td><td>37.3</td></tr><tr><td>38.5</td><td>58.9</td><td>42.3</td><td>34.7</td><td>56.3</td><td>36.8</td></tr><tr><td rowspan="3">R101+FPN</td><td rowspan="3">Mask RCNN</td><td>CBN</td><td>38.5</td><td>59.2</td><td>42.1</td><td>34.6</td><td>56.4</td><td>36.6</td></tr><tr><td>GN</td><td>40.3</td><td>61.2</td><td>44.2</td><td>36.6</td><td>58.5</td><td>39.2</td></tr><tr><td>syncBN</td><td>40.3</td><td></td><td>44.2</td><td></td><td></td><td>38.6</td></tr><tr><td rowspan="2"></td><td rowspan="2"></td><td rowspan="2">CBN</td><td></td><td>60.8</td><td></td><td>36.0</td><td>57.7</td><td></td></tr><tr><td>40.1</td><td>60.5</td><td>44.1</td><td>35.8</td><td>57.3</td><td>38.5</td></tr></table>
158
+
159
+ Table 5: Results with stronger backbones on COCO object detection and instance segmentation.
160
+
161
+ As done in (Wu & He, 2018), we experiment with two settings where the normalizers are activated only at the task-specific heads with frozen BN at the backbone (default), or the normalizers are activated at all the layers except for the early conv1 and conv2 stages in ResNet.
162
+
163
+ Normalizers at backbone and task-specific heads. We further study the effect of different normalizers on the backbone network and task-specific heads for object detection on COCO. CBN, original BN, syncBN, and GN are included in the comparison.
164
+
165
+ Table 4 presents the results. When BN is frozen in the backbone and no normalizer is applied at the head, the $\mathsf { A P } ^ { \mathsf { b b o x } }$ score is $3 6 . 9 \%$ . When the original BN is applied at the head only and at both the backbone and the head, the accuracy drops to $3 6 . 3 \%$ and $3 5 . 5 \%$ , respectively. For CBN, the accuracy is $3 7 . 7 \%$ and $3 7 . 3 \%$ at these two settings, respectively. Without any synchronization across GPUs, CBN can achieve comparable performance with syncBN and GN, showing the superiority of the proposed approach. Unfortunately, due to the accumulation of approximation error, CBN observes a $0 . 4 \%$ decrease in $\mathsf { A P } ^ { \mathsf { b b o x } }$ when replacing frozen BN with CBN in the backbone. Even so, CBN still outperforms the variant with unfrozen BN in backbone by $1 . 8 \%$ .
166
+
167
+ Instance segmentation and stronger backbones. Results of object detection (Faster R-CNN (Ren et al., 2015)) and instance segmentation (Mask R-CNN (He et al., 2017)) with ResNet-50 and ResNet-101 are presented in Table 5. We can observe that our proposed CBN achieves performance comparable to syncBN and GN with R50 and R101 as the backbone on both Faster R-CNN and Mask R-CNN, which demonstrates that CBN is robust and versatile to various deep models and tasks.
168
+
169
+ # 4.3 ABLATION STUDY
170
+
171
+ Effect of temporal window size $k$ . We conduct this ablation on ImageNet image classification and COCO object detection, with each GPU processing 4 images. Figure 3 presents the results. When $k = 1$ , only the batch from the current iteration is utilized; therefore, CBN degenerates to the original BN. The accuracy suffers due to the noisy statistics on small batch sizes. As the window size $k$ gradually increases, more examples from recent iterations are utilized for statistics estimation, leading to greater accuracy. Accuracy saturates at $k = 8$ and even drops slightly. For more distant iterations, the network weights differ more substantially and the Taylor polynomial approximation becomes less accurate.
172
+
173
+ On the other hand, it is empirically observed that the original BN saturates at a batch size of 16 or 32 for numerous applications (Peng et al., 2018; Wu & He, 2018), indicating that the computed statistics become accurate. Thus, a temporal window size of $k = \mathrm { m i n } \big ( \lceil \frac { 1 6 } { \mathrm { b s } \mathrm { p e r G P U } } \rceil , 8 \big )$ is suggested.
174
+
175
+ Effect of compensation. To study this, we compare CBN with 1) a naive baseline where statistics from recent iterations are directly aggregated without compensation via Taylor polynomial, referred to as Naive CBN; and 2) the original BN applied with the same
176
+
177
+ ![](images/013821ca568156ca9f1fddd1d10c2de88ac1d29cd606a0a18f9f54a4deb2eb4a.jpg)
178
+ Figure 3: The effect of temporal window size ${ \bf ( k ) }$ on ImageNet (ResNet-18) and COCO (Faster R-CNN with ResNet-50 and FPN) with $\# \mathrm { b s / G P U } = 4$ for CBN and Naive CBN. Naive CBN directly utilizes statistics from recent iterations, while BN uses the equivalent #examples as CBN for statistics computation.
179
+
180
+ effective example number as CBN (i.e., its batch size per GPU is set to the product of the batch size per GPU and the temporal window size of CBN), which does not require any compensation and serves as an upper performance bound.
181
+
182
+ The experimental results are also presented in Figure 3. CBN clearly surpasses Naive CBN when the previous iterations are included. Actually, Naive CBN fails when the temporal window size grows to $k = 8$ as shown in Figure 3(a), demonstrating the necessity of compensating for changing network weights over iterations. Compared with the original BN upper bound, CBN achieves similar accuracy at the same effective example number. This result indicates that the compensation using a low-order Taylor polynomial by CBN is effective.
183
+
184
+ Figure 4 presents the train and test curves of CBN, Naive CBN, BN-bs4, and BN-bs16 on ImageNet, with 4 images per GPU and a temporal window size of 4 for CBN, Naive CBN, and BN-bs4, and 16 images per GPU for BN-bs16. The train curve of CBN is close to BN-bs4 at the beginning, and approaches BN-bs16 at the end. The reason is that we adopt a burn-in period to avoid the disadvantage of rapid statistics change at beginning of training. The gap between the train curve of Naive CBN and CBN shows that
185
+
186
+ ![](images/83d4820668c41205f69b82a3ef196a731be69a297c6b32808fefac5c7c359924.jpg)
187
+ Figure 4: Training and test curves for CBN, Naive CBN, and BN on ImageNet, with batch size per GPU of 4 and temporal window size $k = \overset { \cdot } { 4 }$ for CBN, Naive CBN, and BN-bs4, and batch size per GPU of 16 for BN-bs16. Thus, the plot of BN-bs16 is the ideal bound.
188
+
189
+ Naive CBN cannot even reach a good convergence on the training set. The test curve of CBN is close to BN-bs16 at the end, while Naive CBN exhibits considerable jitter. All these phenomena indicate the effectiveness of our proposed Taylor compensation.
190
+
191
+ Additional computational overhead and memory footprint. As the inference stage of CBN is the same as BN, we only need to compare the computational overhead and memory footprint at the training time, shown in Table 6. The extra computational overhead mainly includes calculations of the statistics’ respective gradients, Taylor compensations, and averaging operations. For the extra memory, the statistics ( $\mu$ and $\nu$ ), their respective gradients, and the network parameters $( \theta _ { t - 1 } \cdot \cdot \cdot \theta _ { t - ( k - 1 ) } )$ of previous iterations are all stored when applying CBN.
192
+
193
+ <table><tr><td rowspan=2 colspan=1></td><td rowspan=1 colspan=1>ImageNet</td><td rowspan=1 colspan=1>COCO</td></tr><tr><td rowspan=1 colspan=1>BN CBN</td><td rowspan=1 colspan=1>BN CBN</td></tr><tr><td rowspan=1 colspan=1>Base modelGFLOPs Taylor expansionTotal</td><td rowspan=1 colspan=1>5.965.96- 0.215.966.17</td><td rowspan=1 colspan=1>5155.15809.7- 654.65155.15809.7</td></tr><tr><td rowspan=3 colspan=1>MemoryFeature mapNet params(GB) Total</td><td rowspan=1 colspan=1>0.150.45</td><td rowspan=1 colspan=1>14.1 15.1</td></tr><tr><td rowspan=1 colspan=1>0.090.21</td><td rowspan=2 colspan=1>0.3 0.614.4 15.7</td></tr><tr><td rowspan=1 colspan=1>0.240.66</td></tr></table>
194
+
195
+ Table 6: Comparison of theoretical FLOPs and memory footprint between CBN and original BN in both forward and backward passes at training time.
196
+
197
+ From these results, the additional overhead of CBN is seen to be minor.
198
+
199
+ # REFERENCES
200
+
201
+ Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016.
202
+
203
+ Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834–848, 2017.
204
+
205
+ Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, pp. 764–773, 2017.
206
+
207
+ Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. arXiv preprint arXiv:1512.03385, 2015.
208
+
209
+ Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778, 2016.
210
+
211
+ Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pp. 2961–2969, 2017.
212
+
213
+ Tong He, Zhi Zhang, Hang Zhang, Zhongyue Zhang, Junyuan Xie, and Mu Li. Bag of tricks for image classification with convolutional neural networks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019.
214
+
215
+ Sergey Ioffe. Batch renormalization: Towards reducing minibatch dependence in batch-normalized models. In Advances in Neural Information Processing Systems, pp. 1945–1953, 2017.
216
+
217
+ Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International Conference on Machine Learning, pp. 448–456, 2015.
218
+
219
+ Yann LeCun, Léon Bottou, Genevieve B Orr, and Klaus-Robert Müller. Efficient backprop. In Neural Networks: Tricks of the Trade, this book is an outgrowth of a 1996 NIPS workshop, pp. 9–50. Springer-Verlag, 1998.
220
+
221
+ Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pp. 740–755. Springer, 2014.
222
+
223
+ Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017.
224
+
225
+ Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 3431–3440, 2015.
226
+
227
+ Ping Luo, Jiamin Ren, and Zhanglin Peng. Differentiable learning-to-normalize via switchable normalization. arXiv preprint arXiv:1806.10779, 2018.
228
+
229
+ Hyeonseob Nam and Hyo-Eun Kim. Batch-instance normalization for adaptively style-invariant neural networks. In Advances in Neural Information Processing Systems, pp. 2563–2572, 2018.
230
+
231
+ Chao Peng, Tete Xiao, Zeming Li, Yuning Jiang, Xiangyu Zhang, Kai Jia, Gang Yu, and Jian Sun. Megdet: A large mini-batch object detector. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 6181–6189, 2018.
232
+
233
+ Siyuan Qiao, Huiyu Wang, Chenxi Liu, Wei Shen, and Alan Yuille. Weight standardization. arXiv preprint arXiv:1903.10520, 2019.
234
+
235
+ Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in neural information processing systems, pp. 91–99, 2015.
236
+
237
+ Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211–252, 2015.
238
+
239
+ Tim Salimans and Diederik P Kingma. Weight normalization: A simple reparameterization to accelerate training of deep neural networks. In Advances in Neural Information Processing Systems, pp. 901–909, 2016.
240
+
241
+ Wenqi Shao, Tianjian Meng, Jingyu Li, Ruimao Zhang, Yudian Li, Xiaogang Wang, and Ping Luo. Ssn: Learning sparse switchable normalization via sparsestmax. arXiv preprint arXiv:1903.03793, 2019.
242
+
243
+ J Sola and Joaquin Sevilla. Importance of input data normalization for the application of neural networks to complex industrial problems. IEEE Transactions on nuclear science, 44(3):1464–1468, 1997.
244
+
245
+ Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016.
246
+
247
+ Guangrun Wang, Ping Luo, Xinjiang Wang, Liang Lin, et al. Kalman normalization: Normalizing internal representations across network layers. In Advances in Neural Information Processing Systems, pp. 21–31, 2018a.
248
+
249
+ Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 7794–7803, 2018b.
250
+
251
+ Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European Conference on Computer Vision (ECCV), pp. 3–19, 2018.
252
+
253
+ # A ALGORITHM OUTLINE
254
+
255
+ Algorithm 1 presents an outline of our proposed Cross-Iteration Batch Normalization (CBN).
256
+
257
+ # Algorithm 1: Cross-Iteration Batch Normalization(CBN)
258
+
259
+ Input: Feature responses of a network node of the $l$ -th layer at the $t$ -th iteration $\{ x _ { t , i } ^ { l } ( \theta _ { t } ) \} _ { i = 1 } ^ { m }$ , network weights $\{ \theta _ { t - \tau } ^ { l } \} _ { \tau = 0 } ^ { k - 1 }$ , statistics $\{ \mu _ { t - \tau } ^ { l } ( \theta _ { t - \tau } ) \} _ { \tau = 1 } ^ { k - 1 }$ and $\{ \nu _ { t - \tau } ^ { l } ( \theta _ { t - \tau } ) \} _ { \tau = 1 } ^ { k - 1 }$ , and gradients $\{ \partial \mu _ { t - \tau } ( \theta _ { t - \tau } ) / \partial \theta _ { t - \tau } ^ { l } \} _ { \tau = 1 } ^ { k - 1 }$ =0 and $\{ \partial \nu _ { t - \tau } ( \theta _ { t - \tau } ) / \partial \theta _ { t - \tau } ^ { l } \} _ { \tau = 1 } ^ { k - 1 }$ τ τfrom most recent $k - 1$ iterations Output: $\{ y _ { t , i } ^ { l } ( \theta _ { t } ) = \mathbf { C B N } ( x _ { t , i } ^ { l } ( \theta _ { t } ) ) \}$
260
+ 1 $\begin{array} { r } { \mu _ { t } ( \theta _ { t } ) \gets \frac { 1 } { m } \sum _ { i = 1 } ^ { m } x _ { t , i } ( \theta _ { t } ) , \nu _ { t } ( \theta _ { t } ) \gets \frac { 1 } { m } \sum _ { i = 1 } ^ { m } x _ { t , i } ^ { 2 } ( \theta _ { t } ) } \end{array}$ //statistics on the current iteration
261
+ 2 for $\tau \in \{ 1 , \ldots , k \}$ do
262
+ 3 $\begin{array} { r l } & { \mu _ { t - \tau } ^ { l } ( \boldsymbol { \theta } _ { t } ) \dot { \mu } _ { t - \tau } ^ { l } ( \boldsymbol { \theta } _ { t - \tau } ) + \frac { \partial \mu _ { t - \tau } ^ { l } ( \boldsymbol { \theta } _ { t - \tau } ) } { \partial \boldsymbol { \theta } _ { t - \tau } ^ { l } } ( \boldsymbol { \theta } _ { t } ^ { l } - \boldsymbol { \theta } _ { t - \tau } ^ { l } ) } \\ & { \nu _ { t - \tau } ^ { l } ( \boldsymbol { \theta } _ { t } ) \nu _ { t - \tau } ^ { l } ( \boldsymbol { \theta } _ { t - \tau } ) + \frac { \partial \nu _ { t - \tau } ^ { l } ( \boldsymbol { \theta } _ { t - \tau } ) } { \partial \boldsymbol { \theta } _ { t - \tau } ^ { l } } ( \boldsymbol { \theta } _ { t } ^ { l } - \boldsymbol { \theta } _ { t - \tau } ^ { l } ) } \end{array}$ //approximation from recent iterations
263
+ 4 //approximation from recent iterations
264
+ 5 end
265
+ 6 $\begin{array} { r } { \bar { \mu } _ { t , k } ^ { l } ( \theta _ { t } ) \gets \frac { 1 } { k } \sum _ { \tau = 0 } ^ { k - 1 } \mu _ { t - \tau } ^ { l } ( \theta _ { t } ) } \end{array}$ //averaging over recent iterations
266
+ 7 $\begin{array} { r } { \bar { \nu } _ { t , k } ^ { l } ( \theta _ { t } ) \gets \frac { 1 } { k } \sum _ { \tau = 0 } ^ { k - 1 } \operatorname* { m a x } \left[ \nu _ { t - \tau } ^ { l } ( \theta _ { t } ) , \mu _ { t - \tau } ^ { l } ( \theta _ { t } ) ^ { 2 } \right] } \end{array}$ //validation and averaging over recent iterations
267
+ 8 $\bar { \sigma } _ { t , k } ^ { l } ( \theta _ { t } ) ^ { 2 } \gets \bar { \nu } _ { t , k } ^ { l } ( \theta _ { t } ) - \bar { \mu } _ { t , k } ^ { l } ( \theta _ { t } ) ^ { 2 }$ $\begin{array} { r } { \hat { x } _ { t , i } ^ { l } ( \theta _ { t } ) = \frac { x _ { t , i } ^ { l } ( \theta _ { t } ) - \bar { \mu } _ { t , k } ^ { l } ( \theta _ { t } ) } { \sqrt { \bar { \sigma } _ { t , k } ^ { l } ( \theta _ { t } ) ^ { 2 } + \varepsilon } } } \end{array}$ //normalize
268
+ 10 $y _ { t , i } ^ { l } ( \theta _ { t } ) \gets \gamma \hat { x } _ { t , i } ^ { l } ( \theta _ { t } ) + \beta$ //scale and shift
269
+
270
+ B EFFICIENT IMPLEMENTATION OF $\partial \mu _ { t - \tau } ^ { l } ( \theta _ { t - \tau } ) / \partial \theta _ { t - \tau } ^ { l } \ \cdot$ ND $\partial \nu _ { t - \tau } ^ { l } ( \theta _ { t - \tau } ) / \partial \theta _ { t - \tau } ^ { l }$
271
+
272
+ Let $C ^ { l }$ and $C ^ { l - 1 }$ denote the channel dimension of the $l$ -th layer and the $( l - 1 )$ -th layer, respectively, and $K$ denotes the kernel size of $\theta _ { t - \tau } ^ { l }$ . $\mu _ { t - \tau } ^ { l }$ and $\nu _ { t - \tau } ^ { l }$ are thus of $C ^ { l }$ dimensions in channels, and $\theta _ { t - \tau } ^ { l }$ is a $C ^ { l } \times C ^ { l - 1 } \times K$ dimensional tensor. A naive implementation of $\partial \mu _ { t - \tau } ^ { l } ( \theta _ { t - \tau } ) / \partial \theta _ { t - \tau } ^ { l }$ and $\partial \nu _ { t - \tau } ^ { l } ( \theta _ { t - \tau } ) / \partial \theta _ { t - \tau } ^ { l }$ involves computational overhead of $O ( C ^ { l } \times C ^ { l } \times C ^ { l - 1 } \times K )$ . Here we find that the operations of $\mu$ and $\nu$ can be implemented efficiently in $O ( C ^ { l - 1 } \times K )$ and $O ( C ^ { l } \times C ^ { l - 1 } \times K )$ , respectively, thanks to the averaging of feature responses in $\mu$ and $\nu$ .
273
+
274
+ Here we derive the efficient implementation of $\partial \mu _ { t - \tau } ^ { l } ( \theta _ { t - \tau } ) / \partial \theta _ { t - \tau } ^ { l }$ . That of $\partial \nu _ { t - \tau } ^ { l } ( \theta _ { t - \tau } ) / \partial \theta _ { t - \tau } ^ { l }$ is about the same. Let us first simplify the notations a bit. Let $\mu ^ { l }$ and $\theta ^ { l }$ denote $\mu _ { t - \tau } ^ { l } ( \theta _ { t - \tau } )$ and $\theta _ { t - \tau } ^ { l }$ respectively, by removing the irrelevant notations for iterations. The element-wise computation in the forward pass can be computed as
275
+
276
+ $$
277
+ \mu _ { j } ^ { l } = \frac { 1 } { m } \sum _ { i = 1 } ^ { m } x _ { i , j } ^ { l } ,
278
+ $$
279
+
280
+ where $\mu _ { j } ^ { l }$ denotes the $j$ -th channel in $\mu ^ { l }$ , and $x _ { i , j } ^ { l }$ denotes the $j$ -th channel in the $i$ -th example. $x _ { i , j } ^ { l }$ is computed as
281
+
282
+ $$
283
+ x _ { i , j } ^ { l } = \sum _ { n = 1 } ^ { C ^ { l - 1 } } \sum _ { k = 1 } ^ { K } \theta _ { j , n , k } ^ { l } \cdot y _ { i + \mathrm { o f f s e t } ( k ) , n } ^ { l - 1 } ,
284
+ $$
285
+
286
+ where $n$ and $k$ enumerate the input feature dimension and the convolution kernel index, respectively, offset $( k )$ denotes the spatial offset in applying the $k$ -th kernel, and $y ^ { l - 1 }$ is the output of the $\bar { ( } l - 1 )$ -th layer.
287
+
288
+ The element-wise calculation of ${ \partial \mu ^ { l } } / { \partial \theta ^ { l } } \in \mathbb { R } ^ { C ^ { l } \times C ^ { l } \times C ^ { l - 1 } \times K }$ is as follows, taking Eq. (13) and Eq. (14) into consideration:
289
+
290
+ $$
291
+ \begin{array} { l } { \displaystyle { | \frac { \partial \mu ^ { l } } { \partial \theta ^ { l } } | _ { j , q , p , \eta } = \frac { \partial \mu _ { j } ^ { l } } { \partial \theta _ { q , p , \eta } ^ { l } } } } \\ { \displaystyle { \ } = \frac { \partial \frac { 1 } { m } \sum _ { i = 1 } ^ { m } x _ { i , j } ^ { l } } { \partial \theta _ { i , p , \eta } ^ { l } } } \\ { \displaystyle { \ } = \frac { \partial \frac { 1 } { m } \sum _ { i = 1 } ^ { m } \sum _ { n = 1 } ^ { c ^ { l - 1 } } \sum _ { k = 1 } ^ { K } \theta _ { j , n , k } ^ { l } \cdot y _ { i + \mathrm { o f i s e t } ( k ) , n } ^ { l - 1 } } { \partial \theta _ { q , p , \eta } ^ { l } } } \\ { \displaystyle { \ } = \left\{ \begin{array} { l l } { \frac { 1 } { m } \sum _ { i = 1 } ^ { m } y _ { i + \mathrm { o f i s e t } ( \eta ) , p } ^ { l - 1 } } & { , j = q } \\ { 0 } & { , j \neq q } \end{array} \right. . } \end{array}
292
+ $$
293
+
294
+ Thus, $\lbrack { \frac { \partial \mu ^ { l } } { \partial \theta ^ { l } } } \rbrack _ { j , q , p , \eta }$ takes non-zero values only when $j = q$ . This operation can be implemented efficiently in $O ( C ^ { l - 1 } \times K )$ . Similarly, the calculation of $\partial \nu ^ { l } / \partial \theta ^ { l }$ can be obtained in $O ( C ^ { l } \times C ^ { l - 1 } \times K )$ .
295
+
296
+ # C ADDITIONAL EXPERIMENTS
297
+
298
+ CIFAR-10 is selected for the experiments in this section. It consists of 50k training images and 10k test images from 10 classes. We train the standard ResNet-18 for 160 epochs on one GPU by SGD. The momentum and weight decay parameters are set to 0.9 and 0.0001, respectively. We experiment with batch sizes of 32, 16, 8, 4, and 2 images per GPU. The learning rate is scaled linearly to different batch sizes, following the practice in (Peng et al., 2018). The initial learning rate is $0 . 0 2 5 * N / 3 2$ for a batch size per
299
+
300
+ <table><tr><td>Trials</td><td>1 2</td><td>3</td><td>4</td><td></td><td>overall</td></tr><tr><td>BN-bs16</td><td>95.3</td><td>95.0 95.4</td><td>95.3</td><td></td><td>95.395.26±0.14</td></tr><tr><td>BN-bs4</td><td>93.6</td><td>93.5 93.6</td><td>93.6</td><td></td><td>93.8 93.62±0.11</td></tr><tr><td>BRN</td><td>94.4</td><td>94.7 94.5</td><td>94.3</td><td></td><td>94.4 94.46±0.15</td></tr><tr><td>GN</td><td>94.3</td><td>94.0 94.4</td><td>94.2</td><td></td><td>94.4 94.26±0.17</td></tr><tr><td>CBN</td><td>95.095.0</td><td>94.6</td><td></td><td></td><td>95.0 95.2 94.96±0.22</td></tr></table>
301
+
302
+ Table 7: Top-1 accuracy of ResNet-18 with different trials on CIFAR-10. The batch size per GPU is 16 and 4 for BN-bs16 and the other methods, respectively.
303
+
304
+ iteration of $N$ . The learning rate is divided by 10 at epochs 80 and 120. The images are of $3 2 \times 3 2$ pixels with per-image standardization in both training and inference. Random flipping is applied in training.
305
+
306
+ We report the results of BN, BRN, GN, CBN with five trials on CIFAR-10, as shown in Table 7. CBN has the smallest gap with BN-bs16 compared to BN-bs4, BRN, and GN. This result is consistent with previous experiments on ImageNet and COCO. Also the std is tiny, indicating that performance on CIFAR-10 is stable enough for some empirical studies.
307
+
308
+ ![](images/4c51de780bc1b1aafd921d0195e94e264c7efedfdbef719949f0085fb325a4b2.jpg)
309
+ Figure 5: Results of different burn-in periods (in epochs) on CBN, with batch size per iteration of 4, on CIFAR-10 and COCO.
310
+
311
+ On the burn-in period length $T _ { \mathbf { b u r n - i n } }$ . We further study the influence of varying the burn-in period length $T _ { \mathrm { b u r n - i n } }$ , at 4 images per GPU on both CIFAR-10 image classification (ResNet-18) and COCO object detection (Faster R-CNN with FPN and ResNet-50).
312
+
313
+ Figure 5(a) and 5(b) present the results. When the burn-in period is too short, the accuracy suffers. This is because at the beginning of training, the network weights change rapidly, causing the compensation across iterations to be less effective. On the other hand, the accuracy is stable for a wide range of burn-in periods $T _ { \mathrm { b u r n - i n } }$ that are not too short.
314
+
315
+ On the effect of using more than one layer. The efficient implementation is no longer applicable when more than one layer of compensation is adopted. Therefore, we only conduct a two-layer experiment of ResNet-18 on CIFAR-10 in consideration of the heavy extra memory and computational overhead. CBN using two layers for compensation achieves 95.0 on CIFAR-10 (batch size $^ { = 4 }$ , $\mathrm { k } { = } 4$ ), which is comparable to CBN using only one layer. As using more layers does not further improve performance but consumes more FLOPs, we adopt one-layer compensation on CBN in practice.
316
+
317
+ ![](images/90be16eff10751e7954b755905278b75e3f3bb0cef4c82697b14302e2d9bb5b5.jpg)
318
+ Figure 6: Comparison of gradients of statistics w.r.t. current layer vs. that w.r.t. previous layers on CIFAR-10.
319
+
320
+ On the gradients from different layers. The key assumption in Eq. (7) and Eq. (8) is that for a node at the $l$ -th layer, the gradient of its statistics with respect to the network weights at the l-th layer is larger than that of weights from the prior layers, i.e., || ∂ µlt−τ (θt−τ )∂ θ l || $| | \frac { \partial \mu _ { t - \tau } ^ { l } ( \theta _ { t - \tau } ) } { \partial \theta _ { t - \tau } ^ { l } } | | _ { F } \gg | | \frac { \partial \bar { \mu _ { t - \tau } ^ { l } } ( \theta _ { t - \tau } ) } { \partial \theta _ { t - \tau } ^ { r } } | | _ { F }$ and $| | \frac { \partial \boldsymbol { v } _ { t - \tau } ^ { l } ( \theta _ { t - \tau } ) } { \partial \theta _ { t - \tau } ^ { l } } | | _ { F } \gg | | \frac { \partial \boldsymbol { v } _ { t - \tau } ^ { l } ( \theta _ { t - \tau } ) } { \partial \theta _ { t - \tau } ^ { r } } | | _ { F }$ || ∂ νlt−τ (θt−τ )∂ θ r− ||F for r < l, where || · ||F denotes the Frobenius norm. Here we τexamine this assumption empirically for networks trained on CIFAR-10 image recognition.
321
+
322
+ Figure 6 presents the computed ratio of $| | \frac { \partial \mu _ { t - \tau } ^ { l } ( \theta _ { t - \tau } ) } { \partial \theta _ { t - \tau } ^ { r } } | | _ { F } / | | \frac { \partial \mu _ { t - \tau } ^ { l } ( \theta _ { t - \tau } ) } { \partial \theta _ { t - \tau } ^ { l } } | | _ { F }$ and $| | \frac { \partial \nu _ { t - \tau } ^ { l } ( \theta _ { t - \tau } ) } { \partial \theta _ { t - \tau } ^ { r } } | | _ { F } / | | \frac { \partial \nu _ { t - \tau } ^ { l } ( \theta _ { t - \tau } ) } { \partial \theta _ { t - \tau } ^ { l } } | | _ { F }$ || ∂ νlt−τ (θt−τ )l ||F for r ≤ l, at different training epochs. The results suggest that $| | \frac { \partial \boldsymbol { \mu } _ { t - \tau } ^ { l } ( \boldsymbol { \theta } _ { t - \tau } ) } { \partial \boldsymbol { \theta } _ { t - \tau } ^ { l } } | | _ { F } \gg | | \frac { \partial \boldsymbol { \mu } _ { t - \tau } ^ { l } ( \boldsymbol { \theta } _ { t - \tau } ) } { \partial \boldsymbol { \theta } _ { t - \tau } ^ { r } } | | _ { F }$ || ∂ µlt−τ (θt−τ )r ||F and | $| | \frac { \partial \boldsymbol { v } _ { t - \tau } ^ { l } ( \boldsymbol { \theta } _ { t - \tau } ) } { \partial \boldsymbol { \theta } _ { t - \tau } ^ { l } } | | _ { F } \gg | | \frac { \partial \boldsymbol { v } _ { t - \tau } ^ { l } ( \boldsymbol { \theta } _ { t - \tau } ) } { \partial \boldsymbol { \theta } _ { t - \tau } ^ { r } } | | _ { F }$ || ∂ νlt−τ (θt−τ )r ||F hold for r < l, thus validating the approximation in Eq. (7) and Eq. (8).
323
+
324
+ We also study the gradients of non-ResNet models. $| | \frac { \partial \mu _ { t - \tau } ^ { l } ( \theta _ { t - \tau } ) } { \partial \theta _ { t - \tau } ^ { l - 2 } } | | _ { F } / | | \frac { \partial \mu _ { t - \tau } ^ { l } ( \theta _ { t - \tau } ) } { \partial \theta _ { t - \tau } ^ { l } } | | _ { F }$ and $| | \frac { \partial \nu _ { t - \tau } ^ { l } ( \theta _ { t - \tau } ) } { \partial \theta _ { t - \tau } ^ { l - 2 } } | | _ { F } / | | \frac { \partial \nu _ { t - \tau } ^ { l } ( \theta _ { t - \tau } ) } { \partial \theta _ { t - \tau } ^ { l } } | | _ { F }$ || ∂ νlt−τ (θt−τ )l ||F on VGG-16 and InceptionV3 are (0.22 and 0.46) and (0.17 and 0.38), respectively, which is similar to ResNet-18 (0.13 and 0.40), indicating that the assumption should also hold for VGG and the Inception series.
parse/train/BylJUTEKvB/BylJUTEKvB_content_list.json ADDED
@@ -0,0 +1,1630 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "type": "text",
4
+ "text": "CROSS-ITERATION BATCH NORMALIZATION ",
5
+ "text_level": 1,
6
+ "bbox": [
7
+ 176,
8
+ 98,
9
+ 705,
10
+ 121
11
+ ],
12
+ "page_idx": 0
13
+ },
14
+ {
15
+ "type": "text",
16
+ "text": "Anonymous authors Paper under double-blind review ",
17
+ "bbox": [
18
+ 183,
19
+ 145,
20
+ 400,
21
+ 172
22
+ ],
23
+ "page_idx": 0
24
+ },
25
+ {
26
+ "type": "text",
27
+ "text": "ABSTRACT ",
28
+ "text_level": 1,
29
+ "bbox": [
30
+ 454,
31
+ 210,
32
+ 544,
33
+ 224
34
+ ],
35
+ "page_idx": 0
36
+ },
37
+ {
38
+ "type": "text",
39
+ "text": "A well-known issue of Batch Normalization is its significantly reduced effectiveness in the case of small mini-batch sizes. When a mini-batch contains few examples, the statistics upon which the normalization is defined cannot be reliably estimated from it during a training iteration. To address this problem, we present Cross-Iteration Batch Normalization (CBN), in which examples from multiple recent iterations are jointly utilized to enhance estimation quality. A challenge of computing statistics over multiple iterations is that the network activations from different iterations are not comparable to each other due to changes in network weights. We thus compensate for the network weight changes via a proposed technique based on Taylor polynomials, so that the statistics can be accurately estimated and batch normalization can be effectively applied. On object detection and image classification with small mini-batch sizes, CBN is found to outperform the original batch normalization and a direct calculation of statistics over previous iterations without the proposed compensation technique. ",
40
+ "bbox": [
41
+ 233,
42
+ 242,
43
+ 766,
44
+ 436
45
+ ],
46
+ "page_idx": 0
47
+ },
48
+ {
49
+ "type": "text",
50
+ "text": "1 INTRODUCTION ",
51
+ "text_level": 1,
52
+ "bbox": [
53
+ 176,
54
+ 465,
55
+ 336,
56
+ 482
57
+ ],
58
+ "page_idx": 0
59
+ },
60
+ {
61
+ "type": "text",
62
+ "text": "Batch Normalization (BN) (Ioffe & Szegedy, 2015) has played a significant role in the success of deep neural networks. It was introduced to address the issue of internal covariate shift, where the distribution of network activations changes during training iterations due to the updates of network parameters. This shift is commonly believed to be disruptive to network training, and BN alleviates this problem through normalization of the network activations by their mean and variance, computed over the examples within the mini-batch at each iteration. With this normalization, network training can be performed at much higher learning rates and with less sensitivity to weight initialization. ",
63
+ "bbox": [
64
+ 174,
65
+ 498,
66
+ 825,
67
+ 597
68
+ ],
69
+ "page_idx": 0
70
+ },
71
+ {
72
+ "type": "text",
73
+ "text": "In BN, it is assumed that the distribution statistics for the examples within each mini-batch reflect the statistics over the full training set. While this assumption is generally valid for large batch sizes, it breaks down in the small batch size regime (Peng et al., 2018; Wu & He, 2018; Ioffe, 2017), where noisy statistics computed from small sets of examples can lead to a dramatic drop in performance. This problem hinders the application of BN to memory-consuming tasks such as object detection (Ren et al., 2015; Dai et al., 2017), semantic segmentation (Long et al., 2015; Chen et al., 2017) and action recognition (Wang et al., 2018b), where batch sizes are limited due to memory constraints. ",
74
+ "bbox": [
75
+ 174,
76
+ 603,
77
+ 825,
78
+ 702
79
+ ],
80
+ "page_idx": 0
81
+ },
82
+ {
83
+ "type": "text",
84
+ "text": "Towards improving estimation of statistics in the small batch size regime, alternative normalizers have been proposed. Several of them, including Layer Normalization (LN) (Ba et al., 2016), Instance Normalization (IN) (Ulyanov et al., 2016), and Group Normalization (GN) (Wu & He, 2018), compute the mean and variance over the channel dimension, independent of batch size. Different channel-wise normalization techniques, however, tend to be suitable for different tasks, depending on the set of channels involved. On the other hand, synchronized BN (SyncBN) (Peng et al., 2018) yields consistent improvements by processing larger batch sizes across multiple GPUs. These gains in performance come at the cost of additional overhead needed for synchronization across the devices. ",
85
+ "bbox": [
86
+ 174,
87
+ 708,
88
+ 823,
89
+ 819
90
+ ],
91
+ "page_idx": 0
92
+ },
93
+ {
94
+ "type": "text",
95
+ "text": "A seldom explored direction for estimating better statistics is to compute them over the examples from multiple recent training iterations, instead of from only the current iteration as done in previous techniques. This can substantially enlarge the pool of data from which the mean and variance are obtained. However, there exists an obvious drawback to this approach, in that the activation values from different iterations are not comparable to each other due to the changes in network weights. As shown in Figure 1, directly calculating the statistics over multiple iterations, which we refer to as Naive CBN, results in lower accuracy. ",
96
+ "bbox": [
97
+ 174,
98
+ 825,
99
+ 825,
100
+ 924
101
+ ],
102
+ "page_idx": 0
103
+ },
104
+ {
105
+ "type": "image",
106
+ "img_path": "images/026cbf38422de395fdbe5421774471d5517eefca106ac21df3b09ef02e25ece1.jpg",
107
+ "image_caption": [
108
+ "Figure 1: Top-1 classification accuracy vs. batch sizes per iteration. The base model is a ResNet-18 (He et al., 2016) trained on ImageNet (Russakovsky et al., 2015). BN (Ioffe & Szegedy, 2015)’s accuracy drops rapidly when the batch size is reduced. GN (Wu & He, 2018) exhibits stable performance but underperforms BN on adequate batch sizes. CBN compensates for the reduced batch size per GPU by exploiting approximated statistics from recent iterations (Temporal window size denotes how many recent iters are utilized for statistics computation). CBN shows relatively stable performance over different batch sizes. Naive CBN does not work well, which directly calculates statistics from recent iterations without compensation. "
109
+ ],
110
+ "image_footnote": [],
111
+ "bbox": [
112
+ 173,
113
+ 103,
114
+ 375,
115
+ 239
116
+ ],
117
+ "page_idx": 1
118
+ },
119
+ {
120
+ "type": "text",
121
+ "text": "In this paper, we present a method that compensates for the network weight changes among iterations, so that examples from preceding iterations can be effectively used to improve batch normalization. Our method, called Cross-Iteration Batch Normalization (CBN), is motivated by the observation that network weights change gradually, instead of abruptly, between consecutive training iterations, thanks to the iterative nature of Stochastic Gradient Descent (SGD). As a result, the mean and variance of examples from recent iterations can be well approximated for the current network weights via a low-order Taylor polynomial, defined on gradients of the statistics with respect to the network weights. The compensated means and variances from multiple recent iterations are averaged with those of the current iteration to produce better estimates of the statistics. ",
122
+ "bbox": [
123
+ 174,
124
+ 252,
125
+ 825,
126
+ 376
127
+ ],
128
+ "page_idx": 1
129
+ },
130
+ {
131
+ "type": "text",
132
+ "text": "In the small batch size regime, CBN leads to appreciable performance improvements over the original BN, as exhibited in Figure 1. The superiority of our proposed approach is further demonstrated through more extensive experiments on ImageNet classification and object detection on COCO. These gains are obtained with negligible overhead, as the statistics from previous iterations have already been computed and Taylor polynomials are simple to evaluate. With this work, it is shown that cues for batch normalization can successfully be extracted along the time dimension, opening a new direction for investigation. ",
133
+ "bbox": [
134
+ 174,
135
+ 383,
136
+ 825,
137
+ 481
138
+ ],
139
+ "page_idx": 1
140
+ },
141
+ {
142
+ "type": "text",
143
+ "text": "2 RELATED WORK ",
144
+ "text_level": 1,
145
+ "bbox": [
146
+ 176,
147
+ 502,
148
+ 344,
149
+ 518
150
+ ],
151
+ "page_idx": 1
152
+ },
153
+ {
154
+ "type": "text",
155
+ "text": "The importance of normalization in training neural networks has been recognized for decades (LeCun et al., 1998). In general, normalization can be performed on three components: input data, hidden activations, and network parameters. Among them, input data normalization is used most commonly because of its simplicity and effectiveness (Sola & Sevilla, 1997; LeCun et al., 1998). ",
156
+ "bbox": [
157
+ 174,
158
+ 535,
159
+ 823,
160
+ 590
161
+ ],
162
+ "page_idx": 1
163
+ },
164
+ {
165
+ "type": "text",
166
+ "text": "After the introduction of Batch Normalization (Ioffe & Szegedy, 2015), the normalization of activations has become nearly as prevalent. By normalizing hidden activations by their statistics within each mini-batch, BN effectively alleviates the vanishing gradient problem and significantly speeds up the training of deep networks. To mitigate the mini-batch size dependency of BN, a number of variants have been proposed, including Layer Normalization (LN) (Ba et al., 2016), Instance Normalization (IN) (Ulyanov et al., 2016), Group Normalization (GN) (Wu & He, 2018), and Batch Instance Normalization (BIN) (Nam & Kim, 2018). The motivation of LN is to explore more suitable statistics for sequential models, while IN performs normalization in a manner similar to BN but with statistics only for each instance. GN achieves a balance between IN and LN, by dividing features into multiple groups along the channel dimension and computing the mean and variance within each group for normalization. BIN introduces a learnable method for automatically switching between normalizing and maintaining style information, enjoying the advantages of both BN and IN on style transfer tasks. Cross-GPU Batch Normalization (CGBN or SyncBN) (Peng et al., 2018) extends BN across multiple GPUs for the purpose of increasing the effective batch size. Though providing higher accuracy, it introduces synchronization overhead to the training process. Kalman Normalization (KN) (Wang et al., 2018a) presents a Kalman filtering procedure for estimating the statistics for a network layer from the layer’s observed statistics and the computed statistics of previous layers. ",
167
+ "bbox": [
168
+ 174,
169
+ 597,
170
+ 825,
171
+ 833
172
+ ],
173
+ "page_idx": 1
174
+ },
175
+ {
176
+ "type": "text",
177
+ "text": "Batch Renormalization (BRN) (Ioffe, 2017) is the first attempt to utilize the statistics of recent iterations for normalization. It does not compensate for the statistics from recent iterations, but rather it down-weights the importance of statistics from distant iterations. This down-weighting heuristic, however, does not make the resulting statistics “correct\", as the statistics from recent iterations are not of the current network weights. BRN can be deemed as a special version of our Naive CBN baseline (without Taylor polynomial approximation), where distant iterations are down-weighted. ",
178
+ "bbox": [
179
+ 174,
180
+ 840,
181
+ 825,
182
+ 924
183
+ ],
184
+ "page_idx": 1
185
+ },
186
+ {
187
+ "type": "image",
188
+ "img_path": "images/94ad9ce1a1693e6654fb5403358b512982b490d0b827eb1b63dbb0684ac9829b.jpg",
189
+ "image_caption": [
190
+ "Figure 2: Illustration of BN and the proposed Cross-Iteration Batch Normalization (CBN). "
191
+ ],
192
+ "image_footnote": [],
193
+ "bbox": [
194
+ 241,
195
+ 99,
196
+ 761,
197
+ 299
198
+ ],
199
+ "page_idx": 2
200
+ },
201
+ {
202
+ "type": "text",
203
+ "text": "Recent work have also investigated the normalization of network parameters. In Weight Normalization (WN) (Salimans & Kingma, 2016), the optimization of network weights is improved through a reparameterization of weight vectors into their length and direction. Weight Standardization (WS) (Qiao et al., 2019) instead reparameterizes weights based on their first and second moments for the purpose of smoothing the loss landscape of the optimization problem. To combine the advantages of multiple normalization techniques, Switchable Normalization (SN) (Luo et al., 2018) and Sparse Switchable Normalization (SSN) (Shao et al., 2019) make use of differentiable learning to switch among different normalization methods. ",
204
+ "bbox": [
205
+ 173,
206
+ 335,
207
+ 825,
208
+ 446
209
+ ],
210
+ "page_idx": 2
211
+ },
212
+ {
213
+ "type": "text",
214
+ "text": "The proposed CBN takes an activation normalization approach that aims to mitigate the mini-batch dependency of BN. Different from existing techniques, it provides a way to effectively aggregate statistics across multiple training iterations. ",
215
+ "bbox": [
216
+ 176,
217
+ 453,
218
+ 823,
219
+ 496
220
+ ],
221
+ "page_idx": 2
222
+ },
223
+ {
224
+ "type": "text",
225
+ "text": "3 METHOD ",
226
+ "text_level": 1,
227
+ "bbox": [
228
+ 174,
229
+ 515,
230
+ 282,
231
+ 531
232
+ ],
233
+ "page_idx": 2
234
+ },
235
+ {
236
+ "type": "text",
237
+ "text": "3.1 REVISITING BATCH NORMALIZATION ",
238
+ "text_level": 1,
239
+ "bbox": [
240
+ 176,
241
+ 546,
242
+ 475,
243
+ 561
244
+ ],
245
+ "page_idx": 2
246
+ },
247
+ {
248
+ "type": "text",
249
+ "text": "The original batch normalization (BN) (Ioffe & Szegedy, 2015) whitens the activations of each layer by the statistics computed within a mini-batch. Denote $\\theta _ { t }$ and $x _ { t , i } ( \\theta _ { t } )$ as the network weights and the feature response of a certain layer for the $i$ -th example in the $t$ -th mini-batch. With these values, BN conducts the following normalization: ",
250
+ "bbox": [
251
+ 173,
252
+ 571,
253
+ 825,
254
+ 628
255
+ ],
256
+ "page_idx": 2
257
+ },
258
+ {
259
+ "type": "equation",
260
+ "img_path": "images/a507543df0a1735b1e1ad9e97abcc060b7cf57f165b44a0bb5fca09daf514f81.jpg",
261
+ "text": "$$\n\\hat { x } _ { t , i } ( \\theta _ { t } ) = \\frac { x _ { t , i } ( \\theta _ { t } ) - \\mu _ { t } ( \\theta _ { t } ) } { \\sqrt { \\sigma _ { t } ( \\theta _ { t } ) ^ { 2 } + \\varepsilon } } ,\n$$",
262
+ "text_format": "latex",
263
+ "bbox": [
264
+ 408,
265
+ 627,
266
+ 588,
267
+ 665
268
+ ],
269
+ "page_idx": 2
270
+ },
271
+ {
272
+ "type": "text",
273
+ "text": "where $\\hat { x } _ { t , i } ( \\theta _ { t } )$ is the whitened activation with zero mean and unit variance, $\\varepsilon$ is a small constant added for numerical stability, and $\\mu _ { t } ( \\theta _ { t } )$ and $\\sigma _ { t } ( \\theta _ { t } )$ are the mean and variance computed for all the examples from the current mini-batch, i.e., ",
274
+ "bbox": [
275
+ 174,
276
+ 669,
277
+ 825,
278
+ 712
279
+ ],
280
+ "page_idx": 2
281
+ },
282
+ {
283
+ "type": "equation",
284
+ "img_path": "images/e559527b5f7ad46a205d3a941485692044ca21f28e212dddae122152625f67f6.jpg",
285
+ "text": "$$\n\\begin{array} { l } { \\displaystyle \\mu _ { t } ( \\boldsymbol { \\theta } _ { t } ) = \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\boldsymbol { x } _ { t , i } ( \\boldsymbol { \\theta } _ { t } ) , } \\\\ { \\displaystyle \\sigma _ { t } ( \\boldsymbol { \\theta } _ { t } ) = \\sqrt { \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } ( \\boldsymbol { x } _ { t , i } ( \\boldsymbol { \\theta } _ { t } ) - \\boldsymbol { \\mu } _ { t } ( \\boldsymbol { \\theta } _ { t } ) ) ^ { 2 } } = \\sqrt { \\boldsymbol { \\nu } _ { t } ( \\boldsymbol { \\theta } _ { t } ) - \\boldsymbol { \\mu } _ { t } ( \\boldsymbol { \\theta } _ { t } ) ^ { 2 } } , } \\end{array}\n$$",
286
+ "text_format": "latex",
287
+ "bbox": [
288
+ 307,
289
+ 715,
290
+ 691,
291
+ 797
292
+ ],
293
+ "page_idx": 2
294
+ },
295
+ {
296
+ "type": "text",
297
+ "text": "where $\\begin{array} { r } { \\nu _ { t } ( \\theta _ { t } ) = \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } x _ { t , i } ( \\theta _ { t } ) ^ { 2 } } \\end{array}$ , and $m$ denotes the number of examples in the current mini-batch. The whitened activation $\\hat { x } _ { t , i } ( \\theta _ { t } )$ further undergoes a linear transform with learnable weights, to increase its expressive power: ",
298
+ "bbox": [
299
+ 173,
300
+ 801,
301
+ 826,
302
+ 847
303
+ ],
304
+ "page_idx": 2
305
+ },
306
+ {
307
+ "type": "equation",
308
+ "img_path": "images/47d316b23d869dd42bba650813068539bfae9a807ec081a4059d459f7b1c6700.jpg",
309
+ "text": "$$\ny _ { t , i } ( \\theta _ { t } ) = \\gamma \\hat { x } _ { t , i } ( \\theta _ { t } ) + \\beta ,\n$$",
310
+ "text_format": "latex",
311
+ "bbox": [
312
+ 421,
313
+ 852,
314
+ 575,
315
+ 869
316
+ ],
317
+ "page_idx": 2
318
+ },
319
+ {
320
+ "type": "text",
321
+ "text": "where $\\gamma$ and $\\beta$ are the learnable parameters (initialized to $\\gamma = 1$ and $\\beta = 0$ in this work). ",
322
+ "bbox": [
323
+ 174,
324
+ 873,
325
+ 750,
326
+ 890
327
+ ],
328
+ "page_idx": 2
329
+ },
330
+ {
331
+ "type": "text",
332
+ "text": "When the batch size $m$ is small, the statistics $\\mu _ { t } ( \\theta _ { t } )$ and $\\sigma _ { t } ( \\theta _ { t } )$ become noisy estimates of the training set statistics, thus degrading the effects of batch normalization. In the ImageNet classification task for which the BN module was originally designed, a batch size of 32 is typical. However, for other tasks requiring larger models and/or higher image resolution, such as object detection, semantic segmentation and video recognition, the typical batch size may be as small as 1 or 2 due to GPU memory limitations. The original BN becomes considerably less effective in such cases. ",
333
+ "bbox": [
334
+ 173,
335
+ 895,
336
+ 830,
337
+ 924
338
+ ],
339
+ "page_idx": 2
340
+ },
341
+ {
342
+ "type": "text",
343
+ "text": "",
344
+ "bbox": [
345
+ 174,
346
+ 103,
347
+ 825,
348
+ 160
349
+ ],
350
+ "page_idx": 3
351
+ },
352
+ {
353
+ "type": "text",
354
+ "text": "3.2 LEVERAGING STATISTICS FROM PREVIOUS ITERATIONS ",
355
+ "text_level": 1,
356
+ "bbox": [
357
+ 174,
358
+ 175,
359
+ 601,
360
+ 190
361
+ ],
362
+ "page_idx": 3
363
+ },
364
+ {
365
+ "type": "text",
366
+ "text": "To address the issue of BN with small mini-batches, a naive approach is to compute the mean and variance over the current and previous iterations. However, the statistics $\\mu _ { t - \\tau } ( \\theta _ { t - \\tau } )$ and $\\nu _ { t - \\tau } ( \\theta _ { t - \\tau } )$ of the $\\left( t - \\tau \\right)$ -th iteration are computed under the network weights $\\theta _ { t - \\tau }$ , making them obsolete for the current iteration. As a consequence, directly aggregating statistics from multiple iterations produces inaccurate estimates of the mean and variance, leading to significantly worse performance. ",
367
+ "bbox": [
368
+ 173,
369
+ 200,
370
+ 825,
371
+ 272
372
+ ],
373
+ "page_idx": 3
374
+ },
375
+ {
376
+ "type": "text",
377
+ "text": "We observe that the network weights change smoothly between consecutive iterations, due to the nature of gradient-based training. This allows us to approximate $\\mu _ { t - \\tau } ( \\theta _ { t } )$ and $\\nu _ { t - \\tau } ( \\theta _ { t } )$ from the readily available $\\mu _ { t - \\tau } ( \\theta _ { t - \\tau } )$ and $\\nu _ { t - \\tau } ( \\theta _ { t - \\tau } )$ via a Taylor polynomial, i.e., ",
378
+ "bbox": [
379
+ 174,
380
+ 277,
381
+ 821,
382
+ 321
383
+ ],
384
+ "page_idx": 3
385
+ },
386
+ {
387
+ "type": "equation",
388
+ "img_path": "images/55666f84e994c6bdb6e2651889747a3461f5f6edf57620db14eb7b49a472cb7b.jpg",
389
+ "text": "$$\n\\begin{array} { r } { \\mu _ { t - \\tau } ( \\theta _ { t } ) = \\mu _ { t - \\tau } ( \\theta _ { t - \\tau } ) + \\frac { \\partial \\mu _ { t - \\tau } ( \\theta _ { t - \\tau } ) } { \\partial \\theta _ { t - \\tau } } ( \\theta _ { t } - \\theta _ { t - \\tau } ) + \\mathbf { O } ( | | \\theta _ { t } - \\theta _ { t - \\tau } | | ^ { 2 } ) , } \\\\ { \\nu _ { t - \\tau } ( \\theta _ { t } ) = \\nu _ { t - \\tau } ( \\theta _ { t - \\tau } ) + \\frac { \\partial \\nu _ { t - \\tau } ( \\theta _ { t - \\tau } ) } { \\partial \\theta _ { t - \\tau } } ( \\theta _ { t } - \\theta _ { t - \\tau } ) + \\mathbf { O } ( | | \\theta _ { t } - \\theta _ { t - \\tau } | | ^ { 2 } ) , } \\end{array}\n$$",
390
+ "text_format": "latex",
391
+ "bbox": [
392
+ 269,
393
+ 321,
394
+ 727,
395
+ 390
396
+ ],
397
+ "page_idx": 3
398
+ },
399
+ {
400
+ "type": "text",
401
+ "text": "where $\\partial \\mu _ { t - \\tau } ( \\theta _ { t - \\tau } ) / \\partial \\theta _ { t - \\tau }$ and $\\partial \\nu _ { t - \\tau } ( \\theta _ { t - \\tau } ) / \\partial \\theta _ { t - \\tau }$ are gradients of the statistics with respect to the network weights, and $\\mathbf { O } ( | | \\theta _ { t } - \\theta _ { t - \\tau } | | ^ { 2 } )$ denotes higher-order terms of the Taylor polynomial, which can be omitted since the first-order term dominates when $\\left( \\theta _ { t } - \\theta _ { t - \\tau } \\right)$ is small. ",
402
+ "bbox": [
403
+ 176,
404
+ 391,
405
+ 823,
406
+ 435
407
+ ],
408
+ "page_idx": 3
409
+ },
410
+ {
411
+ "type": "text",
412
+ "text": "In Eq. (5) and Eq. (6), the gradients $\\partial \\mu _ { t - \\tau } ( \\theta _ { t - \\tau } ) / \\partial \\theta _ { t - \\tau }$ and $\\partial \\nu _ { t - \\tau } ( \\theta _ { t - \\tau } ) / \\partial \\theta _ { t - \\tau }$ cannot be precisely determined at a negligible cost because the statistics $\\mu _ { t - \\tau } ^ { l } ( \\theta _ { t - \\tau } )$ and $\\nu _ { t - \\tau } ^ { l } ( \\theta _ { t - \\tau } )$ for a node at the $l$ -th network layer depend on all the network weights prior to the $l$ -th layer, i.e., $\\partial \\mu _ { t - \\tau } ^ { l } ( \\theta _ { t - \\tau } ) / \\partial \\theta _ { t - \\tau } ^ { r } \\neq 0$ and $\\partial { \\nu _ { t - \\tau } ^ { l } } ( \\theta _ { t - \\tau } ) / \\partial \\theta _ { t - \\tau } ^ { r } \\ne 0$ for $r \\leq l$ , where $\\theta _ { t - \\tau } ^ { r }$ denotes the network weights at the $r$ -th layer. Only when $r = l$ can these gradients be derived in closed form efficiently. ",
413
+ "bbox": [
414
+ 173,
415
+ 440,
416
+ 825,
417
+ 518
418
+ ],
419
+ "page_idx": 3
420
+ },
421
+ {
422
+ "type": "text",
423
+ "text": "Empirically, we find that as the layer index $r$ decreases $( r \\leq l )$ , the partial gradients $\\frac { \\partial \\mu _ { t } ^ { l } ( \\theta _ { t } ) } { \\theta _ { t } ^ { r } }$ and $\\frac { \\partial \\nu _ { t } ^ { l } ( \\theta _ { t } ) } { \\theta _ { t } ^ { r } }$ rapidly diminish. These reduced effects of network weight changes at earlier layers on the activation distributions in later layers may perhaps be explained by the reduced internal covariate shift of BN. Motivated by this phenomenon, which is studied in Appendix C, we propose to truncate these partial gradients at layer $l$ . ",
424
+ "bbox": [
425
+ 173,
426
+ 523,
427
+ 825,
428
+ 603
429
+ ],
430
+ "page_idx": 3
431
+ },
432
+ {
433
+ "type": "text",
434
+ "text": "Thus, we further approximate Eq. (5) and Eq. (6) by ",
435
+ "bbox": [
436
+ 176,
437
+ 608,
438
+ 517,
439
+ 623
440
+ ],
441
+ "page_idx": 3
442
+ },
443
+ {
444
+ "type": "equation",
445
+ "img_path": "images/99c2307be99e2111252c6c2bdea6e883d47b0cbe9f990b1d3cfdda1074c07e51.jpg",
446
+ "text": "$$\n\\mu _ { t - \\tau } ^ { l } ( \\boldsymbol { \\theta } _ { t } ) \\approx \\mu _ { t - \\tau } ^ { l } ( \\boldsymbol { \\theta } _ { t - \\tau } ) + \\frac { \\partial \\mu _ { t - \\tau } ^ { l } ( \\boldsymbol { \\theta } _ { t - \\tau } ) } { \\partial \\boldsymbol { \\theta } _ { t - \\tau } ^ { l } } ( \\boldsymbol { \\theta } _ { t } ^ { l } - \\boldsymbol { \\theta } _ { t - \\tau } ^ { l } ) ,\n$$",
447
+ "text_format": "latex",
448
+ "bbox": [
449
+ 331,
450
+ 619,
451
+ 665,
452
+ 655
453
+ ],
454
+ "page_idx": 3
455
+ },
456
+ {
457
+ "type": "equation",
458
+ "img_path": "images/d4e6feb917c7355c7913903a23b50287dc9f16145af1b81e87ed66d5bc97f500.jpg",
459
+ "text": "$$\n\\boldsymbol { v } _ { t - \\tau } ^ { l } ( \\boldsymbol { \\theta } _ { t } ) \\approx \\boldsymbol { v } _ { t - \\tau } ^ { l } ( \\boldsymbol { \\theta } _ { t - \\tau } ) + \\frac { \\partial \\boldsymbol { v } _ { t - \\tau } ^ { l } ( \\boldsymbol { \\theta } _ { t - \\tau } ) } { \\partial \\boldsymbol { \\theta } _ { t - \\tau } ^ { l } } ( \\boldsymbol { \\theta } _ { t } ^ { l } - \\boldsymbol { \\theta } _ { t - \\tau } ^ { l } ) .\n$$",
460
+ "text_format": "latex",
461
+ "bbox": [
462
+ 334,
463
+ 657,
464
+ 665,
465
+ 694
466
+ ],
467
+ "page_idx": 3
468
+ },
469
+ {
470
+ "type": "text",
471
+ "text": "A naive implementation of $\\partial \\mu _ { t - \\tau } ^ { l } ( \\theta _ { t - \\tau } ) / \\partial \\theta _ { t - \\tau } ^ { l }$ and $\\partial \\nu _ { t - \\tau } ^ { l } ( \\theta _ { t - \\tau } ) / \\partial \\theta _ { t - \\tau } ^ { l }$ involves computational overhead of $O ( C ^ { l } \\times C ^ { l } \\times C ^ { l - 1 } \\times K )$ , where $C ^ { l }$ and $C ^ { l - 1 }$ denote the channel dimension of the $l$ -th layer and the $( l - 1 )$ -th layer, respectively, and $K$ denotes the kernel size of $\\theta _ { t - \\tau } ^ { l }$ . Here we find that the operation can be implemented efficiently in $O ( C ^ { l } \\times C ^ { l - 1 } \\times K )$ , thanks to the averaging over feature responses of $\\mu$ and $\\nu$ . See Appendix $\\mathbf { B }$ for the details. ",
472
+ "bbox": [
473
+ 173,
474
+ 695,
475
+ 826,
476
+ 773
477
+ ],
478
+ "page_idx": 3
479
+ },
480
+ {
481
+ "type": "text",
482
+ "text": "3.3 CROSS-ITERATION BATCH NORMALIZATION ",
483
+ "text_level": 1,
484
+ "bbox": [
485
+ 174,
486
+ 787,
487
+ 522,
488
+ 803
489
+ ],
490
+ "page_idx": 3
491
+ },
492
+ {
493
+ "type": "text",
494
+ "text": "After compensating for network weight changes, we aggregate the statistics of the $k - 1$ most recent iterations with those of the current iteration $t$ to obtain the statistics used in CBN: ",
495
+ "bbox": [
496
+ 173,
497
+ 814,
498
+ 823,
499
+ 843
500
+ ],
501
+ "page_idx": 3
502
+ },
503
+ {
504
+ "type": "equation",
505
+ "img_path": "images/892fa270a7f4b96234d9aee034ee9888f41f840a6d96abd92e4a8da5af41253c.jpg",
506
+ "text": "$$\n\\begin{array} { r l } & { \\bar { \\mu } _ { t , k } ^ { l } ( \\boldsymbol { \\theta } _ { t } ) = \\displaystyle \\frac { 1 } { k } \\sum _ { \\tau = 0 } ^ { k - 1 } \\mu _ { t - \\tau } ^ { l } ( \\boldsymbol { \\theta } _ { t } ) , } \\\\ & { \\bar { \\nu } _ { t , k } ^ { l } ( \\boldsymbol { \\theta } _ { t } ) = \\displaystyle \\frac { 1 } { k } \\sum _ { \\tau = 0 } ^ { k - 1 } \\operatorname* { m a x } \\left[ \\nu _ { t - \\tau } ^ { l } ( \\boldsymbol { \\theta } _ { t } ) , \\mu _ { t - \\tau } ^ { l } ( \\boldsymbol { \\theta } _ { t } ) ^ { 2 } \\right] , } \\end{array}\n$$",
507
+ "text_format": "latex",
508
+ "bbox": [
509
+ 357,
510
+ 843,
511
+ 638,
512
+ 922
513
+ ],
514
+ "page_idx": 3
515
+ },
516
+ {
517
+ "type": "table",
518
+ "img_path": "images/3628fc2d05f1da7c5b916aa17559f1ff9a5ea17cc6a47b29fb13d4f6b07074c5.jpg",
519
+ "table_caption": [],
520
+ "table_footnote": [
521
+ "Table 1: Comparison of different feature normalization methods. #bs/GPU denotes batch size per GPU. "
522
+ ],
523
+ "table_body": "<table><tr><td></td><td>batch size per iter</td><td>#examples for statistics</td><td>Norm axis</td></tr><tr><td>IN</td><td>#bs/GPU * #GPU</td><td>1</td><td>(spatial)</td></tr><tr><td>LN</td><td>#bs/GPU * #GPU</td><td>1</td><td>(channel, spatial)</td></tr><tr><td>GN</td><td>#bs/GPU * #GPU</td><td>1</td><td>(channel group,spatial)</td></tr><tr><td>BN</td><td>#bs/GPU * #GPU</td><td>#bs/GPU</td><td>(batch,spatial)</td></tr><tr><td>syncBN</td><td>#bs/GPU * #GPU</td><td>#bs/GPU *#GPU</td><td>(batch,spatial, GPU)</td></tr><tr><td>CBN</td><td>#bs/GPU * #GPU</td><td>#bs/GPU J* temporal window</td><td>(batch,spatial, iteration)</td></tr></table>",
524
+ "bbox": [
525
+ 215,
526
+ 101,
527
+ 790,
528
+ 193
529
+ ],
530
+ "page_idx": 4
531
+ },
532
+ {
533
+ "type": "equation",
534
+ "img_path": "images/5358b4a4406e1a6aaa6b035ff456d59a934df10213f0cfede58c6499f1e3830e.jpg",
535
+ "text": "$$\n\\bar { \\sigma } _ { t , k } ^ { l } ( \\theta _ { t } ) = \\sqrt { \\bar { \\nu } _ { t , k } ^ { l } ( \\theta _ { t } ) - \\bar { \\mu } _ { t , k } ^ { l } ( \\theta _ { t } ) ^ { 2 } } ,\n$$",
536
+ "text_format": "latex",
537
+ "bbox": [
538
+ 392,
539
+ 212,
540
+ 606,
541
+ 239
542
+ ],
543
+ "page_idx": 4
544
+ },
545
+ {
546
+ "type": "text",
547
+ "text": "where $\\mu _ { t - \\tau } ^ { l } ( \\theta _ { t } )$ and $\\nu _ { t - \\tau } ^ { l } ( \\theta _ { t } )$ are computed from Eq. (7) and Eq. (8). In Eq. (10), $\\bar { \\nu } _ { t , k } ^ { l } ( \\theta _ { t } )$ is determined from the maximum of $\\nu _ { t - \\tau } ^ { l } ( \\theta _ { t } )$ and $\\mu _ { t - \\tau } ^ { l } ( \\theta _ { t } ) ^ { 2 }$ in each iteration because $\\nu _ { t - \\tau } ^ { l } ( \\theta _ { t } ) \\geq \\mu _ { t - \\tau } ^ { l } ( \\theta _ { t } ) ^ { 2 }$ should hold for valid statistics but may be violated by Taylor polynomial approximations in Eq. (7) and Eq. (8). Finally, $\\bar { \\mu } _ { t , k } ^ { l } ( \\theta _ { t } )$ and $\\bar { \\sigma } _ { t , k } ^ { \\tilde { l } } ( \\theta _ { t } )$ are applied to normalize the corresponding feature responses $\\{ x _ { t , i } ^ { l } ( \\theta _ { t } ) \\} _ { i = 1 } ^ { m }$ at the current iteration: ",
548
+ "bbox": [
549
+ 173,
550
+ 242,
551
+ 825,
552
+ 320
553
+ ],
554
+ "page_idx": 4
555
+ },
556
+ {
557
+ "type": "equation",
558
+ "img_path": "images/17e7bab410d67d81a88329c9e290a41d8922c68c4f620aff57068cf8e510a33d.jpg",
559
+ "text": "$$\n\\hat { x } _ { t , i } ^ { l } ( \\theta _ { t } ) = \\frac { x _ { t , i } ^ { l } ( \\theta _ { t } ) - \\bar { \\mu } _ { t , k } ^ { l } ( \\theta _ { t } ) } { \\sqrt { \\bar { \\sigma } _ { t , k } ^ { l } ( \\theta _ { t } ) ^ { 2 } + \\varepsilon } } .\n$$",
560
+ "text_format": "latex",
561
+ "bbox": [
562
+ 403,
563
+ 318,
564
+ 593,
565
+ 367
566
+ ],
567
+ "page_idx": 4
568
+ },
569
+ {
570
+ "type": "text",
571
+ "text": "With CBN, the effective number of examples used to compute the statistics for the current iteration is $k$ times as large as that for the original BN. In training, the loss gradients are backpropagated to the network weights and activations at the current iteration, i.e., $\\mathbf { \\bar { \\boldsymbol { \\theta } } } _ { t } ^ { l }$ and $x _ { t , i } ^ { l } ( \\theta _ { t } )$ . Those of the previous iterations are fixed and do not receive gradients. Hence, the computation cost of CBN in back-propagation is the same as that of BN. ",
572
+ "bbox": [
573
+ 173,
574
+ 375,
575
+ 823,
576
+ 446
577
+ ],
578
+ "page_idx": 4
579
+ },
580
+ {
581
+ "type": "text",
582
+ "text": "Replacing the BN modules in a network by CBN leads to only minor increases in computational overhead and memory footprint. For computation, the additional overhead mainly comes from computing the partial derivatives $\\partial \\mu _ { t - \\tau } ( \\theta _ { t - \\tau } ) / \\partial \\theta _ { t - \\tau } ^ { l }$ and $\\partial \\nu _ { t - \\tau } ( \\theta _ { t - \\tau } ) / \\partial \\theta _ { t - \\tau } ^ { l }$ , which is insignificant in relation tothe statistics $( \\{ \\mu _ { t - \\tau } ^ { l } ( \\theta _ { t - \\tau } ) \\} _ { \\tau = 1 } ^ { k - 1 }$ theand $\\{ \\nu _ { t - \\tau } ^ { l } ( \\theta _ { t - \\tau } ) \\} _ { \\tau = 1 } ^ { k - 1 } )$ For memory, the and the gradients $( \\{ \\partial \\mu _ { t - \\tau } ( \\stackrel { \\cdot } { \\theta _ { t - \\tau } } ) / \\partial \\theta _ { t - \\tau } ^ { l } \\} _ { \\tau = 1 } ^ { k - 1 }$ and $\\{ \\partial \\nu _ { t - \\tau } ( \\theta _ { t - \\tau } ) / \\partial \\theta _ { t - \\tau } ^ { l } \\} _ { \\tau = 1 } ^ { k - 1 } )$ τ τ=1 computed for the most recent τ τ=1 iterations, which is also minor compared to the rest of the memory consumed in processing the input examples. The additional computation and memory of CBN is reported for our experiments in Table 6. ",
583
+ "bbox": [
584
+ 173,
585
+ 452,
586
+ 825,
587
+ 569
588
+ ],
589
+ "page_idx": 4
590
+ },
591
+ {
592
+ "type": "text",
593
+ "text": "A key hyper-parameter in the proposed CBN is the temporal window size, $k$ , of recent iterations used for statistics estimation. A broader window enlarges the set of examples, but the example quality becomes increasingly lower for more distant iterations, since the differences in network parameters $\\theta _ { t }$ and $\\theta _ { t - \\tau }$ become more significant and are compensated less well using a low-order Taylor polynomial. Empirically, we found that CBN is effective with a window size up to $k = 8$ in a variety of settings and tasks. The only trick is that the window size should be kept small at the beginning of training, when the network weights change quickly. Thus, we introduce a burn-in period of length $T _ { \\mathrm { b u r n } }$ -in for the window size, where $k = 1$ and CBN degenerates to the original BN. In our experiments, the burn-in period is set to 25 epochs on ImageNet image classification and 3 epochs on COCO object detection by default. Ablations on this parameter are presented in the Appendix. ",
594
+ "bbox": [
595
+ 173,
596
+ 575,
597
+ 825,
598
+ 715
599
+ ],
600
+ "page_idx": 4
601
+ },
602
+ {
603
+ "type": "text",
604
+ "text": "Table 1 compares CBN with other feature normalization methods. The key difference among these approaches is the axis along which the statistics are counted and the features are normalized. The previous techniques are all designed to exploit examples from the same iteration. By contrast, CBN explores the aggregation of examples along the temporal dimension. As the data utilized by CBN lies in a direction orthogonal to that of previous methods, the proposed CBN could potentially be combined with other feature normalization approaches to further enhance statistics estimation in certain challenging applications. ",
605
+ "bbox": [
606
+ 174,
607
+ 722,
608
+ 825,
609
+ 820
610
+ ],
611
+ "page_idx": 4
612
+ },
613
+ {
614
+ "type": "text",
615
+ "text": "4 EXPERIMENTS ",
616
+ "text_level": 1,
617
+ "bbox": [
618
+ 176,
619
+ 838,
620
+ 326,
621
+ 854
622
+ ],
623
+ "page_idx": 4
624
+ },
625
+ {
626
+ "type": "text",
627
+ "text": "4.1 IMAGE CLASSIFICATION ON IMAGENET ",
628
+ "text_level": 1,
629
+ "bbox": [
630
+ 174,
631
+ 869,
632
+ 488,
633
+ 883
634
+ ],
635
+ "page_idx": 4
636
+ },
637
+ {
638
+ "type": "text",
639
+ "text": "Experimental settings. ImageNet (Russakovsky et al., 2015) is a benchmark dataset for image classification, containing 1.28M training images and 50K validation images from 1000 classes. We follow the standard setting in (He et al., 2015) to train deep networks on the training set and report the single-crop top-1 accuracy on the validation set. Our preprocessing and augmentation strategy strictly follows the GN baseline (Wu & He, 2018). We use a weight decay of 0.0001 for all weight layers, including $\\gamma$ and $\\beta$ . We train standard ResNet-18 for 100 epochs on 4 GPUs, and decrease the learning rate by the cosine decay strategy (He et al., 2019). We use the average over 5 trials for all results. All hyper-parameters, e.g. group size of GN, are carefully tuned via cross-validation. ResNet-18 with BN is our base model. To compare with other normalization methods, we directly replace BN with IN, LN, GN, BRN, and our proposed CBN. ",
640
+ "bbox": [
641
+ 176,
642
+ 895,
643
+ 823,
644
+ 924
645
+ ],
646
+ "page_idx": 4
647
+ },
648
+ {
649
+ "type": "text",
650
+ "text": "",
651
+ "bbox": [
652
+ 173,
653
+ 103,
654
+ 825,
655
+ 215
656
+ ],
657
+ "page_idx": 5
658
+ },
659
+ {
660
+ "type": "text",
661
+ "text": "Comparison of feature normalization methods. We compare the performance of each normalization method with a normal batch size, 32, in Table 2. With sufficient data for reliable statistics, BN easily reaches the highest top-1 accuracy. Similar as the results in previous papers ",
662
+ "bbox": [
663
+ 174,
664
+ 222,
665
+ 486,
666
+ 305
667
+ ],
668
+ "page_idx": 5
669
+ },
670
+ {
671
+ "type": "table",
672
+ "img_path": "images/162e0c508e38b2730c98377634c0075cdbae000f5dcfb50b50712605142ce885.jpg",
673
+ "table_caption": [],
674
+ "table_footnote": [
675
+ "Table 2: Top-1 accuracy of feature normalization methods using ResNet-18 on ImageNet. "
676
+ ],
677
+ "table_body": "<table><tr><td></td><td>IN</td><td>LN</td><td>GN</td><td>BN</td><td>CBN</td></tr><tr><td>Top-1 accuracy</td><td>64.4</td><td>67.9</td><td>68.9</td><td>70.2</td><td>70.2</td></tr></table>",
678
+ "bbox": [
679
+ 501,
680
+ 239,
681
+ 818,
682
+ 270
683
+ ],
684
+ "page_idx": 5
685
+ },
686
+ {
687
+ "type": "text",
688
+ "text": "(Wu & He, 2018), IN and LN achieve significantly worse performance than BN. GN works well on image classification, but still has a small degradation of $1 . 2 \\%$ compared with BN. Over all the methods, our CBN is the only one that is able to achieve comparable accuracy with BN, as it converges to the procedure of BN as the batch size becomes larger. ",
689
+ "bbox": [
690
+ 174,
691
+ 306,
692
+ 825,
693
+ 361
694
+ ],
695
+ "page_idx": 5
696
+ },
697
+ {
698
+ "type": "text",
699
+ "text": "Sensitivity to batch size. We compare the behavior of CBN, original BN (Ioffe & Szegedy, 2015), GN (Wu & He, 2018), and BRN (Ioffe, 2017) at the same number of images per GPU on ImageNet classification. For CBN, the recent iterations are utilized so as to ensure that the number of effective examples is no fewer than 16. For BRN, the settings strictly follow the original paper. We adopt a learning rate of 0.1 for the batch size of 32, and linearly scale the learning rate by $N / 3 2$ for a batch size of $N$ . ",
700
+ "bbox": [
701
+ 174,
702
+ 367,
703
+ 485,
704
+ 520
705
+ ],
706
+ "page_idx": 5
707
+ },
708
+ {
709
+ "type": "table",
710
+ "img_path": "images/5cd334bf792ff90fada45bf5c3ea47f073476e75357c6d49014e6025be6c1942.jpg",
711
+ "table_caption": [
712
+ "Table 3: Top-1 accuracy of feature normalization methods with different batch sizes using ResNet-18 as the base model on ImageNet. "
713
+ ],
714
+ "table_footnote": [],
715
+ "table_body": "<table><tr><td>batch size per GPU</td><td>32</td><td>16</td><td>8</td><td>4</td><td>2</td></tr><tr><td>BN</td><td>70.2</td><td>70.2</td><td>68.4</td><td>65.1</td><td>55.9</td></tr><tr><td>GN</td><td>68.9</td><td>69.0</td><td>68.9</td><td>69.0</td><td>69.1</td></tr><tr><td>BRN</td><td>70.1</td><td>68.5</td><td>68.2</td><td>67.9</td><td>60.3</td></tr><tr><td>CBN</td><td>70.2</td><td>70.2</td><td>70.1</td><td>69.8</td><td>69.3</td></tr></table>",
716
+ "bbox": [
717
+ 501,
718
+ 385,
719
+ 818,
720
+ 453
721
+ ],
722
+ "page_idx": 5
723
+ },
724
+ {
725
+ "type": "text",
726
+ "text": "The results are shown in Table 3. For the original BN, its accuracy drops noticeably as the number of images per GPU is reduced from 32 to 2. BRN suffers a significant performance drop as well. GN maintains its accuracy by utilizing the channel dimension but not batch dimension. For CBN, its accuracy holds by exploiting the examples of recent iterations. Also, CBN outperforms GN by $0 . 9 \\%$ on average top-1 accuracy with different batch sizes. This is reasonable, because the statistics computation of CBN introduces uncertainty caused by the stochastic batch sampling like in BN, but this uncertainty is missing in GN which results in some loss of regularization ability. ",
727
+ "bbox": [
728
+ 173,
729
+ 527,
730
+ 825,
731
+ 625
732
+ ],
733
+ "page_idx": 5
734
+ },
735
+ {
736
+ "type": "text",
737
+ "text": "4.2 OBJECT DETECTION AND INSTANCE SEGMENTATION ON COCO",
738
+ "text_level": 1,
739
+ "bbox": [
740
+ 176,
741
+ 651,
742
+ 658,
743
+ 665
744
+ ],
745
+ "page_idx": 5
746
+ },
747
+ {
748
+ "type": "text",
749
+ "text": "Experimental settings. COCO (Lin et al., 2014) is chosen as the benchmark for object detection and instance segmentation. Models are trained on the COCO 2017 train split with $1 1 8 \\mathrm { k }$ images, and evaluated on the COCO 2017 validation split with 5k images. Following the standard protocol in (Lin et al., 2014), the object detection and instance segmentation accuracies are measured by the mean average precision (mAP) scores at different intersection-over-union (IoU) overlaps at the box and the mask levels, respectively. ",
750
+ "bbox": [
751
+ 174,
752
+ 679,
753
+ 825,
754
+ 763
755
+ ],
756
+ "page_idx": 5
757
+ },
758
+ {
759
+ "type": "text",
760
+ "text": "Following (Wu & He, 2018), Faster R-CNN (Ren et al., 2015) and Mask R-CNN (He et al., 2017) with FPN (Lin et al., 2017) are chosen as the baselines for object detection and instance segmentation, respectively. For both, the 2fc box head is replaced by a 4conv1fc head for better use of the normalization mechanism (Wu & He, 2018). The backbone networks are ImageNet pretrained ResNet-50 (default) or ResNet-101, with specific normalization. Finetuning is performed on the COCO train set for 12 epochs on 4 GPUs by SGD, where each GPU processes 4 images (default). Note that the mean and variance statistics in CBN are computed within each GPU. The learning rate is initialized to be $0 . 0 2 * N / 1 6$ for a batch size per GPU of $N$ , and is decayed by a factor of 10 at the 9-th and the 11-th epochs. The weight decay and momentum parameters are set to 0.0001 and 0.9, respectively. We use the average over 5 trials for all results. All hyper-parameters, e.g. group size of GN, are carefully tuned via cross-validation. ",
761
+ "bbox": [
762
+ 173,
763
+ 770,
764
+ 826,
765
+ 924
766
+ ],
767
+ "page_idx": 5
768
+ },
769
+ {
770
+ "type": "table",
771
+ "img_path": "images/a37620c195e2589e1a2569c73f9e7c64da5ab202e63c1b7f5fe08a3301fb83e8.jpg",
772
+ "table_caption": [
773
+ "Table 4: Results of feature normalization methods on Faster R-CNN with FPN and ResNet50 on COCO. "
774
+ ],
775
+ "table_footnote": [],
776
+ "table_body": "<table><tr><td>backbone</td><td>box head</td><td>Apbbox</td><td>APbbox 50</td><td></td><td>APbbox S</td><td>APbbox M</td><td>APbbox</td></tr><tr><td>fixed BN</td><td>1</td><td>36.9</td><td>58.2</td><td>39.9</td><td>21.2</td><td>40.8</td><td>46.9</td></tr><tr><td>fixed BN</td><td>BN</td><td>36.3</td><td>57.3</td><td>39.2</td><td>20.8</td><td>39.7</td><td>47.3</td></tr><tr><td>fixed BN</td><td>syncBN</td><td>37.7</td><td>58.5</td><td>41.1</td><td>22.0</td><td>40.9</td><td>49.0</td></tr><tr><td>fixed BN</td><td>GN</td><td>37.8</td><td>59.0</td><td>40.8</td><td>22.3</td><td>41.2</td><td>48.4</td></tr><tr><td>fixed BN</td><td>CBN</td><td>37.7</td><td>59.0</td><td>40.7</td><td>22.1</td><td>40.9</td><td>48.8</td></tr><tr><td>BN</td><td>BN</td><td>35.5</td><td>56.4</td><td>38.7</td><td>19.7</td><td>38.8</td><td>47.3</td></tr><tr><td>syncBN</td><td>syncBN</td><td>37.9</td><td>58.5</td><td>41.1</td><td>21.7</td><td>41.5</td><td>49.7</td></tr><tr><td>GN</td><td>GN</td><td>37.8</td><td>59.1</td><td>40.9</td><td>22.4</td><td>41.2</td><td>49.0</td></tr><tr><td>CBN</td><td>CBN</td><td>37.3</td><td>57.7</td><td>39.3</td><td>21.9</td><td>40.8</td><td>48.2</td></tr></table>",
777
+ "bbox": [
778
+ 253,
779
+ 101,
780
+ 740,
781
+ 232
782
+ ],
783
+ "page_idx": 6
784
+ },
785
+ {
786
+ "type": "table",
787
+ "img_path": "images/56de46ea1af735f260ef9b3cc235f9a6fd3935f0f0f55e97a351d5b26f5e747d.jpg",
788
+ "table_caption": [],
789
+ "table_footnote": [
790
+ "Table 5: Results with stronger backbones on COCO object detection and instance segmentation. "
791
+ ],
792
+ "table_body": "<table><tr><td>Backbone</td><td>method</td><td>norm</td><td>Apbbox</td><td>APbox 50</td><td>AP5ox</td><td>APbbox S</td><td>APbox M</td><td>APbbox</td></tr><tr><td rowspan=\"3\">R50+FPN</td><td rowspan=\"3\">Faster RCNN</td><td>GN</td><td>37.8</td><td>59.0</td><td>40.8</td><td>22.3</td><td>41.2</td><td>48.4</td></tr><tr><td>syncBN</td><td>37.7</td><td>58.5</td><td>41.1</td><td>22.0</td><td>40.9</td><td>49.0</td></tr><tr><td>CBN</td><td>37.7</td><td>59.0</td><td>40.7</td><td>22.1</td><td>40.9</td><td>48.8</td></tr><tr><td rowspan=\"2\">R101+FPN</td><td rowspan=\"2\">Faster RCNN</td><td rowspan=\"2\">GN syncBN CBN</td><td>39.3 39.2</td><td>60.6</td><td>42.7 43.0</td><td>22.5</td><td>42.5</td><td>51.3</td></tr><tr><td>39.2</td><td>59.8 60.0</td><td>42.6</td><td>22.2 22.3</td><td>42.9 42.6</td><td>51.6 51.1</td></tr><tr><td rowspan=\"3\">R50+FPN</td><td rowspan=\"3\">Mask RCNN</td><td rowspan=\"3\">GN syncBN</td><td>Apbbox</td><td>APo</td><td>AP5box</td><td>Apmask</td><td>APsk</td><td>APmsk</td></tr><tr><td>38.6</td><td>59.8</td><td>41.9</td><td>35.0</td><td>56.7</td><td>37.3</td></tr><tr><td>38.5</td><td>58.9</td><td>42.3</td><td>34.7</td><td>56.3</td><td>36.8</td></tr><tr><td rowspan=\"3\">R101+FPN</td><td rowspan=\"3\">Mask RCNN</td><td>CBN</td><td>38.5</td><td>59.2</td><td>42.1</td><td>34.6</td><td>56.4</td><td>36.6</td></tr><tr><td>GN</td><td>40.3</td><td>61.2</td><td>44.2</td><td>36.6</td><td>58.5</td><td>39.2</td></tr><tr><td>syncBN</td><td>40.3</td><td></td><td>44.2</td><td></td><td></td><td>38.6</td></tr><tr><td rowspan=\"2\"></td><td rowspan=\"2\"></td><td rowspan=\"2\">CBN</td><td></td><td>60.8</td><td></td><td>36.0</td><td>57.7</td><td></td></tr><tr><td>40.1</td><td>60.5</td><td>44.1</td><td>35.8</td><td>57.3</td><td>38.5</td></tr></table>",
793
+ "bbox": [
794
+ 202,
795
+ 258,
796
+ 792,
797
+ 444
798
+ ],
799
+ "page_idx": 6
800
+ },
801
+ {
802
+ "type": "text",
803
+ "text": "As done in (Wu & He, 2018), we experiment with two settings where the normalizers are activated only at the task-specific heads with frozen BN at the backbone (default), or the normalizers are activated at all the layers except for the early conv1 and conv2 stages in ResNet. ",
804
+ "bbox": [
805
+ 174,
806
+ 469,
807
+ 825,
808
+ 511
809
+ ],
810
+ "page_idx": 6
811
+ },
812
+ {
813
+ "type": "text",
814
+ "text": "Normalizers at backbone and task-specific heads. We further study the effect of different normalizers on the backbone network and task-specific heads for object detection on COCO. CBN, original BN, syncBN, and GN are included in the comparison. ",
815
+ "bbox": [
816
+ 178,
817
+ 518,
818
+ 821,
819
+ 560
820
+ ],
821
+ "page_idx": 6
822
+ },
823
+ {
824
+ "type": "text",
825
+ "text": "Table 4 presents the results. When BN is frozen in the backbone and no normalizer is applied at the head, the $\\mathsf { A P } ^ { \\mathsf { b b o x } }$ score is $3 6 . 9 \\%$ . When the original BN is applied at the head only and at both the backbone and the head, the accuracy drops to $3 6 . 3 \\%$ and $3 5 . 5 \\%$ , respectively. For CBN, the accuracy is $3 7 . 7 \\%$ and $3 7 . 3 \\%$ at these two settings, respectively. Without any synchronization across GPUs, CBN can achieve comparable performance with syncBN and GN, showing the superiority of the proposed approach. Unfortunately, due to the accumulation of approximation error, CBN observes a $0 . 4 \\%$ decrease in $\\mathsf { A P } ^ { \\mathsf { b b o x } }$ when replacing frozen BN with CBN in the backbone. Even so, CBN still outperforms the variant with unfrozen BN in backbone by $1 . 8 \\%$ . ",
826
+ "bbox": [
827
+ 173,
828
+ 566,
829
+ 825,
830
+ 679
831
+ ],
832
+ "page_idx": 6
833
+ },
834
+ {
835
+ "type": "text",
836
+ "text": "Instance segmentation and stronger backbones. Results of object detection (Faster R-CNN (Ren et al., 2015)) and instance segmentation (Mask R-CNN (He et al., 2017)) with ResNet-50 and ResNet-101 are presented in Table 5. We can observe that our proposed CBN achieves performance comparable to syncBN and GN with R50 and R101 as the backbone on both Faster R-CNN and Mask R-CNN, which demonstrates that CBN is robust and versatile to various deep models and tasks. ",
837
+ "bbox": [
838
+ 174,
839
+ 685,
840
+ 825,
841
+ 755
842
+ ],
843
+ "page_idx": 6
844
+ },
845
+ {
846
+ "type": "text",
847
+ "text": "4.3 ABLATION STUDY ",
848
+ "text_level": 1,
849
+ "bbox": [
850
+ 174,
851
+ 782,
852
+ 339,
853
+ 796
854
+ ],
855
+ "page_idx": 6
856
+ },
857
+ {
858
+ "type": "text",
859
+ "text": "Effect of temporal window size $k$ . We conduct this ablation on ImageNet image classification and COCO object detection, with each GPU processing 4 images. Figure 3 presents the results. When $k = 1$ , only the batch from the current iteration is utilized; therefore, CBN degenerates to the original BN. The accuracy suffers due to the noisy statistics on small batch sizes. As the window size $k$ gradually increases, more examples from recent iterations are utilized for statistics estimation, leading to greater accuracy. Accuracy saturates at $k = 8$ and even drops slightly. For more distant iterations, the network weights differ more substantially and the Taylor polynomial approximation becomes less accurate. ",
860
+ "bbox": [
861
+ 173,
862
+ 811,
863
+ 825,
864
+ 924
865
+ ],
866
+ "page_idx": 6
867
+ },
868
+ {
869
+ "type": "text",
870
+ "text": "On the other hand, it is empirically observed that the original BN saturates at a batch size of 16 or 32 for numerous applications (Peng et al., 2018; Wu & He, 2018), indicating that the computed statistics become accurate. Thus, a temporal window size of $k = \\mathrm { m i n } \\big ( \\lceil \\frac { 1 6 } { \\mathrm { b s } \\mathrm { p e r G P U } } \\rceil , 8 \\big )$ is suggested. ",
871
+ "bbox": [
872
+ 173,
873
+ 103,
874
+ 421,
875
+ 219
876
+ ],
877
+ "page_idx": 7
878
+ },
879
+ {
880
+ "type": "text",
881
+ "text": "Effect of compensation. To study this, we compare CBN with 1) a naive baseline where statistics from recent iterations are directly aggregated without compensation via Taylor polynomial, referred to as Naive CBN; and 2) the original BN applied with the same ",
882
+ "bbox": [
883
+ 174,
884
+ 227,
885
+ 419,
886
+ 323
887
+ ],
888
+ "page_idx": 7
889
+ },
890
+ {
891
+ "type": "image",
892
+ "img_path": "images/013821ca568156ca9f1fddd1d10c2de88ac1d29cd606a0a18f9f54a4deb2eb4a.jpg",
893
+ "image_caption": [
894
+ "Figure 3: The effect of temporal window size ${ \\bf ( k ) }$ on ImageNet (ResNet-18) and COCO (Faster R-CNN with ResNet-50 and FPN) with $\\# \\mathrm { b s / G P U } = 4$ for CBN and Naive CBN. Naive CBN directly utilizes statistics from recent iterations, while BN uses the equivalent #examples as CBN for statistics computation. "
895
+ ],
896
+ "image_footnote": [],
897
+ "bbox": [
898
+ 436,
899
+ 106,
900
+ 813,
901
+ 231
902
+ ],
903
+ "page_idx": 7
904
+ },
905
+ {
906
+ "type": "text",
907
+ "text": "effective example number as CBN (i.e., its batch size per GPU is set to the product of the batch size per GPU and the temporal window size of CBN), which does not require any compensation and serves as an upper performance bound. ",
908
+ "bbox": [
909
+ 174,
910
+ 323,
911
+ 825,
912
+ 364
913
+ ],
914
+ "page_idx": 7
915
+ },
916
+ {
917
+ "type": "text",
918
+ "text": "The experimental results are also presented in Figure 3. CBN clearly surpasses Naive CBN when the previous iterations are included. Actually, Naive CBN fails when the temporal window size grows to $k = 8$ as shown in Figure 3(a), demonstrating the necessity of compensating for changing network weights over iterations. Compared with the original BN upper bound, CBN achieves similar accuracy at the same effective example number. This result indicates that the compensation using a low-order Taylor polynomial by CBN is effective. ",
919
+ "bbox": [
920
+ 173,
921
+ 372,
922
+ 825,
923
+ 455
924
+ ],
925
+ "page_idx": 7
926
+ },
927
+ {
928
+ "type": "text",
929
+ "text": "Figure 4 presents the train and test curves of CBN, Naive CBN, BN-bs4, and BN-bs16 on ImageNet, with 4 images per GPU and a temporal window size of 4 for CBN, Naive CBN, and BN-bs4, and 16 images per GPU for BN-bs16. The train curve of CBN is close to BN-bs4 at the beginning, and approaches BN-bs16 at the end. The reason is that we adopt a burn-in period to avoid the disadvantage of rapid statistics change at beginning of training. The gap between the train curve of Naive CBN and CBN shows that ",
930
+ "bbox": [
931
+ 174,
932
+ 463,
933
+ 419,
934
+ 655
935
+ ],
936
+ "page_idx": 7
937
+ },
938
+ {
939
+ "type": "image",
940
+ "img_path": "images/83d4820668c41205f69b82a3ef196a731be69a297c6b32808fefac5c7c359924.jpg",
941
+ "image_caption": [
942
+ "Figure 4: Training and test curves for CBN, Naive CBN, and BN on ImageNet, with batch size per GPU of 4 and temporal window size $k = \\overset { \\cdot } { 4 }$ for CBN, Naive CBN, and BN-bs4, and batch size per GPU of 16 for BN-bs16. Thus, the plot of BN-bs16 is the ideal bound. "
943
+ ],
944
+ "image_footnote": [],
945
+ "bbox": [
946
+ 437,
947
+ 468,
948
+ 810,
949
+ 571
950
+ ],
951
+ "page_idx": 7
952
+ },
953
+ {
954
+ "type": "text",
955
+ "text": "Naive CBN cannot even reach a good convergence on the training set. The test curve of CBN is close to BN-bs16 at the end, while Naive CBN exhibits considerable jitter. All these phenomena indicate the effectiveness of our proposed Taylor compensation. ",
956
+ "bbox": [
957
+ 174,
958
+ 656,
959
+ 825,
960
+ 698
961
+ ],
962
+ "page_idx": 7
963
+ },
964
+ {
965
+ "type": "text",
966
+ "text": "Additional computational overhead and memory footprint. As the inference stage of CBN is the same as BN, we only need to compare the computational overhead and memory footprint at the training time, shown in Table 6. The extra computational overhead mainly includes calculations of the statistics’ respective gradients, Taylor compensations, and averaging operations. For the extra memory, the statistics ( $\\mu$ and $\\nu$ ), their respective gradients, and the network parameters $( \\theta _ { t - 1 } \\cdot \\cdot \\cdot \\theta _ { t - ( k - 1 ) } )$ of previous iterations are all stored when applying CBN. ",
967
+ "bbox": [
968
+ 174,
969
+ 705,
970
+ 485,
971
+ 871
972
+ ],
973
+ "page_idx": 7
974
+ },
975
+ {
976
+ "type": "table",
977
+ "img_path": "images/30c50c522e2b24de58d9426aeb1c3eccb7bb2406346d59aeb511bd634d4cb82a.jpg",
978
+ "table_caption": [],
979
+ "table_footnote": [
980
+ "Table 6: Comparison of theoretical FLOPs and memory footprint between CBN and original BN in both forward and backward passes at training time. "
981
+ ],
982
+ "table_body": "<table><tr><td rowspan=2 colspan=1></td><td rowspan=1 colspan=1>ImageNet</td><td rowspan=1 colspan=1>COCO</td></tr><tr><td rowspan=1 colspan=1>BN CBN</td><td rowspan=1 colspan=1>BN CBN</td></tr><tr><td rowspan=1 colspan=1>Base modelGFLOPs Taylor expansionTotal</td><td rowspan=1 colspan=1>5.965.96- 0.215.966.17</td><td rowspan=1 colspan=1>5155.15809.7- 654.65155.15809.7</td></tr><tr><td rowspan=3 colspan=1>MemoryFeature mapNet params(GB) Total</td><td rowspan=1 colspan=1>0.150.45</td><td rowspan=1 colspan=1>14.1 15.1</td></tr><tr><td rowspan=1 colspan=1>0.090.21</td><td rowspan=2 colspan=1>0.3 0.614.4 15.7</td></tr><tr><td rowspan=1 colspan=1>0.240.66</td></tr></table>",
983
+ "bbox": [
984
+ 496,
985
+ 709,
986
+ 821,
987
+ 816
988
+ ],
989
+ "page_idx": 7
990
+ },
991
+ {
992
+ "type": "text",
993
+ "text": "From these results, the additional overhead of CBN is seen to be minor. ",
994
+ "bbox": [
995
+ 174,
996
+ 872,
997
+ 640,
998
+ 886
999
+ ],
1000
+ "page_idx": 7
1001
+ },
1002
+ {
1003
+ "type": "text",
1004
+ "text": "REFERENCES ",
1005
+ "text_level": 1,
1006
+ "bbox": [
1007
+ 176,
1008
+ 102,
1009
+ 287,
1010
+ 118
1011
+ ],
1012
+ "page_idx": 8
1013
+ },
1014
+ {
1015
+ "type": "text",
1016
+ "text": "Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. ",
1017
+ "bbox": [
1018
+ 171,
1019
+ 126,
1020
+ 823,
1021
+ 155
1022
+ ],
1023
+ "page_idx": 8
1024
+ },
1025
+ {
1026
+ "type": "text",
1027
+ "text": "Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834–848, 2017. ",
1028
+ "bbox": [
1029
+ 173,
1030
+ 165,
1031
+ 825,
1032
+ 220
1033
+ ],
1034
+ "page_idx": 8
1035
+ },
1036
+ {
1037
+ "type": "text",
1038
+ "text": "Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, pp. 764–773, 2017. ",
1039
+ "bbox": [
1040
+ 173,
1041
+ 229,
1042
+ 825,
1043
+ 273
1044
+ ],
1045
+ "page_idx": 8
1046
+ },
1047
+ {
1048
+ "type": "text",
1049
+ "text": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. arXiv preprint arXiv:1512.03385, 2015. ",
1050
+ "bbox": [
1051
+ 173,
1052
+ 284,
1053
+ 823,
1054
+ 313
1055
+ ],
1056
+ "page_idx": 8
1057
+ },
1058
+ {
1059
+ "type": "text",
1060
+ "text": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778, 2016. ",
1061
+ "bbox": [
1062
+ 174,
1063
+ 320,
1064
+ 826,
1065
+ 364
1066
+ ],
1067
+ "page_idx": 8
1068
+ },
1069
+ {
1070
+ "type": "text",
1071
+ "text": "Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pp. 2961–2969, 2017. ",
1072
+ "bbox": [
1073
+ 173,
1074
+ 373,
1075
+ 825,
1076
+ 404
1077
+ ],
1078
+ "page_idx": 8
1079
+ },
1080
+ {
1081
+ "type": "text",
1082
+ "text": "Tong He, Zhi Zhang, Hang Zhang, Zhongyue Zhang, Junyuan Xie, and Mu Li. Bag of tricks for image classification with convolutional neural networks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. ",
1083
+ "bbox": [
1084
+ 174,
1085
+ 412,
1086
+ 825,
1087
+ 455
1088
+ ],
1089
+ "page_idx": 8
1090
+ },
1091
+ {
1092
+ "type": "text",
1093
+ "text": "Sergey Ioffe. Batch renormalization: Towards reducing minibatch dependence in batch-normalized models. In Advances in Neural Information Processing Systems, pp. 1945–1953, 2017. ",
1094
+ "bbox": [
1095
+ 173,
1096
+ 464,
1097
+ 825,
1098
+ 494
1099
+ ],
1100
+ "page_idx": 8
1101
+ },
1102
+ {
1103
+ "type": "text",
1104
+ "text": "Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International Conference on Machine Learning, pp. 448–456, 2015. ",
1105
+ "bbox": [
1106
+ 176,
1107
+ 503,
1108
+ 823,
1109
+ 546
1110
+ ],
1111
+ "page_idx": 8
1112
+ },
1113
+ {
1114
+ "type": "text",
1115
+ "text": "Yann LeCun, Léon Bottou, Genevieve B Orr, and Klaus-Robert Müller. Efficient backprop. In Neural Networks: Tricks of the Trade, this book is an outgrowth of a 1996 NIPS workshop, pp. 9–50. Springer-Verlag, 1998. ",
1116
+ "bbox": [
1117
+ 174,
1118
+ 555,
1119
+ 826,
1120
+ 599
1121
+ ],
1122
+ "page_idx": 8
1123
+ },
1124
+ {
1125
+ "type": "text",
1126
+ "text": "Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pp. 740–755. Springer, 2014. ",
1127
+ "bbox": [
1128
+ 174,
1129
+ 608,
1130
+ 826,
1131
+ 651
1132
+ ],
1133
+ "page_idx": 8
1134
+ },
1135
+ {
1136
+ "type": "text",
1137
+ "text": "Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. ",
1138
+ "bbox": [
1139
+ 174,
1140
+ 660,
1141
+ 826,
1142
+ 704
1143
+ ],
1144
+ "page_idx": 8
1145
+ },
1146
+ {
1147
+ "type": "text",
1148
+ "text": "Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 3431–3440, 2015. ",
1149
+ "bbox": [
1150
+ 171,
1151
+ 713,
1152
+ 825,
1153
+ 756
1154
+ ],
1155
+ "page_idx": 8
1156
+ },
1157
+ {
1158
+ "type": "text",
1159
+ "text": "Ping Luo, Jiamin Ren, and Zhanglin Peng. Differentiable learning-to-normalize via switchable normalization. arXiv preprint arXiv:1806.10779, 2018. ",
1160
+ "bbox": [
1161
+ 169,
1162
+ 765,
1163
+ 825,
1164
+ 795
1165
+ ],
1166
+ "page_idx": 8
1167
+ },
1168
+ {
1169
+ "type": "text",
1170
+ "text": "Hyeonseob Nam and Hyo-Eun Kim. Batch-instance normalization for adaptively style-invariant neural networks. In Advances in Neural Information Processing Systems, pp. 2563–2572, 2018. ",
1171
+ "bbox": [
1172
+ 171,
1173
+ 804,
1174
+ 823,
1175
+ 833
1176
+ ],
1177
+ "page_idx": 8
1178
+ },
1179
+ {
1180
+ "type": "text",
1181
+ "text": "Chao Peng, Tete Xiao, Zeming Li, Yuning Jiang, Xiangyu Zhang, Kai Jia, Gang Yu, and Jian Sun. Megdet: A large mini-batch object detector. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 6181–6189, 2018. ",
1182
+ "bbox": [
1183
+ 176,
1184
+ 843,
1185
+ 823,
1186
+ 886
1187
+ ],
1188
+ "page_idx": 8
1189
+ },
1190
+ {
1191
+ "type": "text",
1192
+ "text": "Siyuan Qiao, Huiyu Wang, Chenxi Liu, Wei Shen, and Alan Yuille. Weight standardization. arXiv preprint arXiv:1903.10520, 2019. ",
1193
+ "bbox": [
1194
+ 173,
1195
+ 895,
1196
+ 821,
1197
+ 924
1198
+ ],
1199
+ "page_idx": 8
1200
+ },
1201
+ {
1202
+ "type": "text",
1203
+ "text": "Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in neural information processing systems, pp. 91–99, 2015. ",
1204
+ "bbox": [
1205
+ 174,
1206
+ 103,
1207
+ 823,
1208
+ 146
1209
+ ],
1210
+ "page_idx": 9
1211
+ },
1212
+ {
1213
+ "type": "text",
1214
+ "text": "Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211–252, 2015. ",
1215
+ "bbox": [
1216
+ 176,
1217
+ 155,
1218
+ 823,
1219
+ 198
1220
+ ],
1221
+ "page_idx": 9
1222
+ },
1223
+ {
1224
+ "type": "text",
1225
+ "text": "Tim Salimans and Diederik P Kingma. Weight normalization: A simple reparameterization to accelerate training of deep neural networks. In Advances in Neural Information Processing Systems, pp. 901–909, 2016. ",
1226
+ "bbox": [
1227
+ 176,
1228
+ 207,
1229
+ 823,
1230
+ 250
1231
+ ],
1232
+ "page_idx": 9
1233
+ },
1234
+ {
1235
+ "type": "text",
1236
+ "text": "Wenqi Shao, Tianjian Meng, Jingyu Li, Ruimao Zhang, Yudian Li, Xiaogang Wang, and Ping Luo. Ssn: Learning sparse switchable normalization via sparsestmax. arXiv preprint arXiv:1903.03793, 2019. ",
1237
+ "bbox": [
1238
+ 174,
1239
+ 257,
1240
+ 826,
1241
+ 300
1242
+ ],
1243
+ "page_idx": 9
1244
+ },
1245
+ {
1246
+ "type": "text",
1247
+ "text": "J Sola and Joaquin Sevilla. Importance of input data normalization for the application of neural networks to complex industrial problems. IEEE Transactions on nuclear science, 44(3):1464–1468, 1997. ",
1248
+ "bbox": [
1249
+ 173,
1250
+ 309,
1251
+ 826,
1252
+ 352
1253
+ ],
1254
+ "page_idx": 9
1255
+ },
1256
+ {
1257
+ "type": "text",
1258
+ "text": "Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016. ",
1259
+ "bbox": [
1260
+ 173,
1261
+ 361,
1262
+ 823,
1263
+ 390
1264
+ ],
1265
+ "page_idx": 9
1266
+ },
1267
+ {
1268
+ "type": "text",
1269
+ "text": "Guangrun Wang, Ping Luo, Xinjiang Wang, Liang Lin, et al. Kalman normalization: Normalizing internal representations across network layers. In Advances in Neural Information Processing Systems, pp. 21–31, 2018a. ",
1270
+ "bbox": [
1271
+ 173,
1272
+ 398,
1273
+ 826,
1274
+ 441
1275
+ ],
1276
+ "page_idx": 9
1277
+ },
1278
+ {
1279
+ "type": "text",
1280
+ "text": "Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 7794–7803, 2018b. ",
1281
+ "bbox": [
1282
+ 173,
1283
+ 450,
1284
+ 825,
1285
+ 492
1286
+ ],
1287
+ "page_idx": 9
1288
+ },
1289
+ {
1290
+ "type": "text",
1291
+ "text": "Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European Conference on Computer Vision (ECCV), pp. 3–19, 2018. ",
1292
+ "bbox": [
1293
+ 173,
1294
+ 502,
1295
+ 823,
1296
+ 531
1297
+ ],
1298
+ "page_idx": 9
1299
+ },
1300
+ {
1301
+ "type": "text",
1302
+ "text": "A ALGORITHM OUTLINE",
1303
+ "text_level": 1,
1304
+ "bbox": [
1305
+ 176,
1306
+ 102,
1307
+ 398,
1308
+ 118
1309
+ ],
1310
+ "page_idx": 10
1311
+ },
1312
+ {
1313
+ "type": "text",
1314
+ "text": "Algorithm 1 presents an outline of our proposed Cross-Iteration Batch Normalization (CBN). ",
1315
+ "bbox": [
1316
+ 169,
1317
+ 137,
1318
+ 782,
1319
+ 152
1320
+ ],
1321
+ "page_idx": 10
1322
+ },
1323
+ {
1324
+ "type": "text",
1325
+ "text": "Algorithm 1: Cross-Iteration Batch Normalization(CBN) ",
1326
+ "text_level": 1,
1327
+ "bbox": [
1328
+ 173,
1329
+ 176,
1330
+ 553,
1331
+ 191
1332
+ ],
1333
+ "page_idx": 10
1334
+ },
1335
+ {
1336
+ "type": "text",
1337
+ "text": "Input: Feature responses of a network node of the $l$ -th layer at the $t$ -th iteration $\\{ x _ { t , i } ^ { l } ( \\theta _ { t } ) \\} _ { i = 1 } ^ { m }$ , network weights $\\{ \\theta _ { t - \\tau } ^ { l } \\} _ { \\tau = 0 } ^ { k - 1 }$ , statistics $\\{ \\mu _ { t - \\tau } ^ { l } ( \\theta _ { t - \\tau } ) \\} _ { \\tau = 1 } ^ { k - 1 }$ and $\\{ \\nu _ { t - \\tau } ^ { l } ( \\theta _ { t - \\tau } ) \\} _ { \\tau = 1 } ^ { k - 1 }$ , and gradients $\\{ \\partial \\mu _ { t - \\tau } ( \\theta _ { t - \\tau } ) / \\partial \\theta _ { t - \\tau } ^ { l } \\} _ { \\tau = 1 } ^ { k - 1 }$ =0 and $\\{ \\partial \\nu _ { t - \\tau } ( \\theta _ { t - \\tau } ) / \\partial \\theta _ { t - \\tau } ^ { l } \\} _ { \\tau = 1 } ^ { k - 1 }$ τ τfrom most recent $k - 1$ iterations Output: $\\{ y _ { t , i } ^ { l } ( \\theta _ { t } ) = \\mathbf { C B N } ( x _ { t , i } ^ { l } ( \\theta _ { t } ) ) \\}$ \n1 $\\begin{array} { r } { \\mu _ { t } ( \\theta _ { t } ) \\gets \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } x _ { t , i } ( \\theta _ { t } ) , \\nu _ { t } ( \\theta _ { t } ) \\gets \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } x _ { t , i } ^ { 2 } ( \\theta _ { t } ) } \\end{array}$ //statistics on the current iteration \n2 for $\\tau \\in \\{ 1 , \\ldots , k \\}$ do \n3 $\\begin{array} { r l } & { \\mu _ { t - \\tau } ^ { l } ( \\boldsymbol { \\theta } _ { t } ) \\dot { \\mu } _ { t - \\tau } ^ { l } ( \\boldsymbol { \\theta } _ { t - \\tau } ) + \\frac { \\partial \\mu _ { t - \\tau } ^ { l } ( \\boldsymbol { \\theta } _ { t - \\tau } ) } { \\partial \\boldsymbol { \\theta } _ { t - \\tau } ^ { l } } ( \\boldsymbol { \\theta } _ { t } ^ { l } - \\boldsymbol { \\theta } _ { t - \\tau } ^ { l } ) } \\\\ & { \\nu _ { t - \\tau } ^ { l } ( \\boldsymbol { \\theta } _ { t } ) \\nu _ { t - \\tau } ^ { l } ( \\boldsymbol { \\theta } _ { t - \\tau } ) + \\frac { \\partial \\nu _ { t - \\tau } ^ { l } ( \\boldsymbol { \\theta } _ { t - \\tau } ) } { \\partial \\boldsymbol { \\theta } _ { t - \\tau } ^ { l } } ( \\boldsymbol { \\theta } _ { t } ^ { l } - \\boldsymbol { \\theta } _ { t - \\tau } ^ { l } ) } \\end{array}$ //approximation from recent iterations \n4 //approximation from recent iterations \n5 end \n6 $\\begin{array} { r } { \\bar { \\mu } _ { t , k } ^ { l } ( \\theta _ { t } ) \\gets \\frac { 1 } { k } \\sum _ { \\tau = 0 } ^ { k - 1 } \\mu _ { t - \\tau } ^ { l } ( \\theta _ { t } ) } \\end{array}$ //averaging over recent iterations \n7 $\\begin{array} { r } { \\bar { \\nu } _ { t , k } ^ { l } ( \\theta _ { t } ) \\gets \\frac { 1 } { k } \\sum _ { \\tau = 0 } ^ { k - 1 } \\operatorname* { m a x } \\left[ \\nu _ { t - \\tau } ^ { l } ( \\theta _ { t } ) , \\mu _ { t - \\tau } ^ { l } ( \\theta _ { t } ) ^ { 2 } \\right] } \\end{array}$ //validation and averaging over recent iterations \n8 $\\bar { \\sigma } _ { t , k } ^ { l } ( \\theta _ { t } ) ^ { 2 } \\gets \\bar { \\nu } _ { t , k } ^ { l } ( \\theta _ { t } ) - \\bar { \\mu } _ { t , k } ^ { l } ( \\theta _ { t } ) ^ { 2 }$ $\\begin{array} { r } { \\hat { x } _ { t , i } ^ { l } ( \\theta _ { t } ) = \\frac { x _ { t , i } ^ { l } ( \\theta _ { t } ) - \\bar { \\mu } _ { t , k } ^ { l } ( \\theta _ { t } ) } { \\sqrt { \\bar { \\sigma } _ { t , k } ^ { l } ( \\theta _ { t } ) ^ { 2 } + \\varepsilon } } } \\end{array}$ //normalize \n10 $y _ { t , i } ^ { l } ( \\theta _ { t } ) \\gets \\gamma \\hat { x } _ { t , i } ^ { l } ( \\theta _ { t } ) + \\beta$ //scale and shift ",
1338
+ "bbox": [
1339
+ 160,
1340
+ 190,
1341
+ 826,
1342
+ 474
1343
+ ],
1344
+ "page_idx": 10
1345
+ },
1346
+ {
1347
+ "type": "text",
1348
+ "text": "B EFFICIENT IMPLEMENTATION OF $\\partial \\mu _ { t - \\tau } ^ { l } ( \\theta _ { t - \\tau } ) / \\partial \\theta _ { t - \\tau } ^ { l } \\ \\cdot$ ND $\\partial \\nu _ { t - \\tau } ^ { l } ( \\theta _ { t - \\tau } ) / \\partial \\theta _ { t - \\tau } ^ { l }$ ",
1349
+ "bbox": [
1350
+ 173,
1351
+ 516,
1352
+ 683,
1353
+ 556
1354
+ ],
1355
+ "page_idx": 10
1356
+ },
1357
+ {
1358
+ "type": "text",
1359
+ "text": "Let $C ^ { l }$ and $C ^ { l - 1 }$ denote the channel dimension of the $l$ -th layer and the $( l - 1 )$ -th layer, respectively, and $K$ denotes the kernel size of $\\theta _ { t - \\tau } ^ { l }$ . $\\mu _ { t - \\tau } ^ { l }$ and $\\nu _ { t - \\tau } ^ { l }$ are thus of $C ^ { l }$ dimensions in channels, and $\\theta _ { t - \\tau } ^ { l }$ is a $C ^ { l } \\times C ^ { l - 1 } \\times K$ dimensional tensor. A naive implementation of $\\partial \\mu _ { t - \\tau } ^ { l } ( \\theta _ { t - \\tau } ) / \\partial \\theta _ { t - \\tau } ^ { l }$ and $\\partial \\nu _ { t - \\tau } ^ { l } ( \\theta _ { t - \\tau } ) / \\partial \\theta _ { t - \\tau } ^ { l }$ involves computational overhead of $O ( C ^ { l } \\times C ^ { l } \\times C ^ { l - 1 } \\times K )$ . Here we find that the operations of $\\mu$ and $\\nu$ can be implemented efficiently in $O ( C ^ { l - 1 } \\times K )$ and $O ( C ^ { l } \\times C ^ { l - 1 } \\times K )$ , respectively, thanks to the averaging of feature responses in $\\mu$ and $\\nu$ . ",
1360
+ "bbox": [
1361
+ 173,
1362
+ 575,
1363
+ 826,
1364
+ 670
1365
+ ],
1366
+ "page_idx": 10
1367
+ },
1368
+ {
1369
+ "type": "text",
1370
+ "text": "Here we derive the efficient implementation of $\\partial \\mu _ { t - \\tau } ^ { l } ( \\theta _ { t - \\tau } ) / \\partial \\theta _ { t - \\tau } ^ { l }$ . That of $\\partial \\nu _ { t - \\tau } ^ { l } ( \\theta _ { t - \\tau } ) / \\partial \\theta _ { t - \\tau } ^ { l }$ is about the same. Let us first simplify the notations a bit. Let $\\mu ^ { l }$ and $\\theta ^ { l }$ denote $\\mu _ { t - \\tau } ^ { l } ( \\theta _ { t - \\tau } )$ and $\\theta _ { t - \\tau } ^ { l }$ respectively, by removing the irrelevant notations for iterations. The element-wise computation in the forward pass can be computed as ",
1371
+ "bbox": [
1372
+ 173,
1373
+ 674,
1374
+ 825,
1375
+ 734
1376
+ ],
1377
+ "page_idx": 10
1378
+ },
1379
+ {
1380
+ "type": "equation",
1381
+ "img_path": "images/c469471b2c0acd34a4dc9e502b62ca681092b35952e1c01828c5668ac4c00709.jpg",
1382
+ "text": "$$\n\\mu _ { j } ^ { l } = \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } x _ { i , j } ^ { l } ,\n$$",
1383
+ "text_format": "latex",
1384
+ "bbox": [
1385
+ 447,
1386
+ 741,
1387
+ 549,
1388
+ 777
1389
+ ],
1390
+ "page_idx": 10
1391
+ },
1392
+ {
1393
+ "type": "text",
1394
+ "text": "where $\\mu _ { j } ^ { l }$ denotes the $j$ -th channel in $\\mu ^ { l }$ , and $x _ { i , j } ^ { l }$ denotes the $j$ -th channel in the $i$ -th example. $x _ { i , j } ^ { l }$ is computed as ",
1395
+ "bbox": [
1396
+ 171,
1397
+ 790,
1398
+ 825,
1399
+ 821
1400
+ ],
1401
+ "page_idx": 10
1402
+ },
1403
+ {
1404
+ "type": "equation",
1405
+ "img_path": "images/d4bae55d2c574a594d40d78bade9d1d0a817eea8add0d45beed1b0b5e5a7d254.jpg",
1406
+ "text": "$$\nx _ { i , j } ^ { l } = \\sum _ { n = 1 } ^ { C ^ { l - 1 } } \\sum _ { k = 1 } ^ { K } \\theta _ { j , n , k } ^ { l } \\cdot y _ { i + \\mathrm { o f f s e t } ( k ) , n } ^ { l - 1 } ,\n$$",
1407
+ "text_format": "latex",
1408
+ "bbox": [
1409
+ 392,
1410
+ 829,
1411
+ 604,
1412
+ 871
1413
+ ],
1414
+ "page_idx": 10
1415
+ },
1416
+ {
1417
+ "type": "text",
1418
+ "text": "where $n$ and $k$ enumerate the input feature dimension and the convolution kernel index, respectively, offset $( k )$ denotes the spatial offset in applying the $k$ -th kernel, and $y ^ { l - 1 }$ is the output of the $\\bar { ( } l - 1 )$ -th layer. ",
1419
+ "bbox": [
1420
+ 174,
1421
+ 881,
1422
+ 825,
1423
+ 924
1424
+ ],
1425
+ "page_idx": 10
1426
+ },
1427
+ {
1428
+ "type": "text",
1429
+ "text": "The element-wise calculation of ${ \\partial \\mu ^ { l } } / { \\partial \\theta ^ { l } } \\in \\mathbb { R } ^ { C ^ { l } \\times C ^ { l } \\times C ^ { l - 1 } \\times K }$ is as follows, taking Eq. (13) and Eq. (14) into consideration: ",
1430
+ "bbox": [
1431
+ 173,
1432
+ 101,
1433
+ 823,
1434
+ 132
1435
+ ],
1436
+ "page_idx": 11
1437
+ },
1438
+ {
1439
+ "type": "equation",
1440
+ "img_path": "images/fb878bf44bffcb5428e4e131e077d363a2e316ff726080968857774e152c9caf.jpg",
1441
+ "text": "$$\n\\begin{array} { l } { \\displaystyle { | \\frac { \\partial \\mu ^ { l } } { \\partial \\theta ^ { l } } | _ { j , q , p , \\eta } = \\frac { \\partial \\mu _ { j } ^ { l } } { \\partial \\theta _ { q , p , \\eta } ^ { l } } } } \\\\ { \\displaystyle { \\ } = \\frac { \\partial \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } x _ { i , j } ^ { l } } { \\partial \\theta _ { i , p , \\eta } ^ { l } } } \\\\ { \\displaystyle { \\ } = \\frac { \\partial \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\sum _ { n = 1 } ^ { c ^ { l - 1 } } \\sum _ { k = 1 } ^ { K } \\theta _ { j , n , k } ^ { l } \\cdot y _ { i + \\mathrm { o f i s e t } ( k ) , n } ^ { l - 1 } } { \\partial \\theta _ { q , p , \\eta } ^ { l } } } \\\\ { \\displaystyle { \\ } = \\left\\{ \\begin{array} { l l } { \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } y _ { i + \\mathrm { o f i s e t } ( \\eta ) , p } ^ { l - 1 } } & { , j = q } \\\\ { 0 } & { , j \\neq q } \\end{array} \\right. . } \\end{array}\n$$",
1442
+ "text_format": "latex",
1443
+ "bbox": [
1444
+ 323,
1445
+ 131,
1446
+ 673,
1447
+ 295
1448
+ ],
1449
+ "page_idx": 11
1450
+ },
1451
+ {
1452
+ "type": "text",
1453
+ "text": "Thus, $\\lbrack { \\frac { \\partial \\mu ^ { l } } { \\partial \\theta ^ { l } } } \\rbrack _ { j , q , p , \\eta }$ takes non-zero values only when $j = q$ . This operation can be implemented efficiently in $O ( C ^ { l - 1 } \\times K )$ . Similarly, the calculation of $\\partial \\nu ^ { l } / \\partial \\theta ^ { l }$ can be obtained in $O ( C ^ { l } \\times C ^ { l - 1 } \\times K )$ . ",
1454
+ "bbox": [
1455
+ 176,
1456
+ 299,
1457
+ 828,
1458
+ 334
1459
+ ],
1460
+ "page_idx": 11
1461
+ },
1462
+ {
1463
+ "type": "text",
1464
+ "text": "C ADDITIONAL EXPERIMENTS ",
1465
+ "text_level": 1,
1466
+ "bbox": [
1467
+ 176,
1468
+ 353,
1469
+ 444,
1470
+ 369
1471
+ ],
1472
+ "page_idx": 11
1473
+ },
1474
+ {
1475
+ "type": "text",
1476
+ "text": "CIFAR-10 is selected for the experiments in this section. It consists of 50k training images and 10k test images from 10 classes. We train the standard ResNet-18 for 160 epochs on one GPU by SGD. The momentum and weight decay parameters are set to 0.9 and 0.0001, respectively. We experiment with batch sizes of 32, 16, 8, 4, and 2 images per GPU. The learning rate is scaled linearly to different batch sizes, following the practice in (Peng et al., 2018). The initial learning rate is $0 . 0 2 5 * N / 3 2$ for a batch size per ",
1477
+ "bbox": [
1478
+ 174,
1479
+ 383,
1480
+ 485,
1481
+ 536
1482
+ ],
1483
+ "page_idx": 11
1484
+ },
1485
+ {
1486
+ "type": "table",
1487
+ "img_path": "images/f3d38c29e7034b039ec33e84c3f1481b1587ef0a9b83b51cad84a30621dc6536.jpg",
1488
+ "table_caption": [],
1489
+ "table_footnote": [
1490
+ "Table 7: Top-1 accuracy of ResNet-18 with different trials on CIFAR-10. The batch size per GPU is 16 and 4 for BN-bs16 and the other methods, respectively. "
1491
+ ],
1492
+ "table_body": "<table><tr><td>Trials</td><td>1 2</td><td>3</td><td>4</td><td></td><td>overall</td></tr><tr><td>BN-bs16</td><td>95.3</td><td>95.0 95.4</td><td>95.3</td><td></td><td>95.395.26±0.14</td></tr><tr><td>BN-bs4</td><td>93.6</td><td>93.5 93.6</td><td>93.6</td><td></td><td>93.8 93.62±0.11</td></tr><tr><td>BRN</td><td>94.4</td><td>94.7 94.5</td><td>94.3</td><td></td><td>94.4 94.46±0.15</td></tr><tr><td>GN</td><td>94.3</td><td>94.0 94.4</td><td>94.2</td><td></td><td>94.4 94.26±0.17</td></tr><tr><td>CBN</td><td>95.095.0</td><td>94.6</td><td></td><td></td><td>95.0 95.2 94.96±0.22</td></tr></table>",
1493
+ "bbox": [
1494
+ 501,
1495
+ 401,
1496
+ 816,
1497
+ 481
1498
+ ],
1499
+ "page_idx": 11
1500
+ },
1501
+ {
1502
+ "type": "text",
1503
+ "text": "iteration of $N$ . The learning rate is divided by 10 at epochs 80 and 120. The images are of $3 2 \\times 3 2$ pixels with per-image standardization in both training and inference. Random flipping is applied in training. ",
1504
+ "bbox": [
1505
+ 174,
1506
+ 536,
1507
+ 826,
1508
+ 578
1509
+ ],
1510
+ "page_idx": 11
1511
+ },
1512
+ {
1513
+ "type": "text",
1514
+ "text": "We report the results of BN, BRN, GN, CBN with five trials on CIFAR-10, as shown in Table 7. CBN has the smallest gap with BN-bs16 compared to BN-bs4, BRN, and GN. This result is consistent with previous experiments on ImageNet and COCO. Also the std is tiny, indicating that performance on CIFAR-10 is stable enough for some empirical studies. ",
1515
+ "bbox": [
1516
+ 174,
1517
+ 584,
1518
+ 825,
1519
+ 640
1520
+ ],
1521
+ "page_idx": 11
1522
+ },
1523
+ {
1524
+ "type": "image",
1525
+ "img_path": "images/4c51de780bc1b1aafd921d0195e94e264c7efedfdbef719949f0085fb325a4b2.jpg",
1526
+ "image_caption": [
1527
+ "Figure 5: Results of different burn-in periods (in epochs) on CBN, with batch size per iteration of 4, on CIFAR-10 and COCO. "
1528
+ ],
1529
+ "image_footnote": [],
1530
+ "bbox": [
1531
+ 261,
1532
+ 651,
1533
+ 732,
1534
+ 800
1535
+ ],
1536
+ "page_idx": 11
1537
+ },
1538
+ {
1539
+ "type": "text",
1540
+ "text": "On the burn-in period length $T _ { \\mathbf { b u r n - i n } }$ . We further study the influence of varying the burn-in period length $T _ { \\mathrm { b u r n - i n } }$ , at 4 images per GPU on both CIFAR-10 image classification (ResNet-18) and COCO object detection (Faster R-CNN with FPN and ResNet-50). ",
1541
+ "bbox": [
1542
+ 173,
1543
+ 845,
1544
+ 825,
1545
+ 888
1546
+ ],
1547
+ "page_idx": 11
1548
+ },
1549
+ {
1550
+ "type": "text",
1551
+ "text": "Figure 5(a) and 5(b) present the results. When the burn-in period is too short, the accuracy suffers. This is because at the beginning of training, the network weights change rapidly, causing the compensation across iterations to be less effective. On the other hand, the accuracy is stable for a wide range of burn-in periods $T _ { \\mathrm { b u r n - i n } }$ that are not too short. ",
1552
+ "bbox": [
1553
+ 174,
1554
+ 895,
1555
+ 823,
1556
+ 924
1557
+ ],
1558
+ "page_idx": 11
1559
+ },
1560
+ {
1561
+ "type": "text",
1562
+ "text": "",
1563
+ "bbox": [
1564
+ 171,
1565
+ 103,
1566
+ 823,
1567
+ 132
1568
+ ],
1569
+ "page_idx": 12
1570
+ },
1571
+ {
1572
+ "type": "text",
1573
+ "text": "On the effect of using more than one layer. The efficient implementation is no longer applicable when more than one layer of compensation is adopted. Therefore, we only conduct a two-layer experiment of ResNet-18 on CIFAR-10 in consideration of the heavy extra memory and computational overhead. CBN using two layers for compensation achieves 95.0 on CIFAR-10 (batch size $^ { = 4 }$ , $\\mathrm { k } { = } 4$ ), which is comparable to CBN using only one layer. As using more layers does not further improve performance but consumes more FLOPs, we adopt one-layer compensation on CBN in practice. ",
1574
+ "bbox": [
1575
+ 173,
1576
+ 138,
1577
+ 825,
1578
+ 223
1579
+ ],
1580
+ "page_idx": 12
1581
+ },
1582
+ {
1583
+ "type": "image",
1584
+ "img_path": "images/90be16eff10751e7954b755905278b75e3f3bb0cef4c82697b14302e2d9bb5b5.jpg",
1585
+ "image_caption": [
1586
+ "Figure 6: Comparison of gradients of statistics w.r.t. current layer vs. that w.r.t. previous layers on CIFAR-10. "
1587
+ ],
1588
+ "image_footnote": [],
1589
+ "bbox": [
1590
+ 184,
1591
+ 247,
1592
+ 813,
1593
+ 422
1594
+ ],
1595
+ "page_idx": 12
1596
+ },
1597
+ {
1598
+ "type": "text",
1599
+ "text": "On the gradients from different layers. The key assumption in Eq. (7) and Eq. (8) is that for a node at the $l$ -th layer, the gradient of its statistics with respect to the network weights at the l-th layer is larger than that of weights from the prior layers, i.e., || ∂ µlt−τ (θt−τ )∂ θ l || $| | \\frac { \\partial \\mu _ { t - \\tau } ^ { l } ( \\theta _ { t - \\tau } ) } { \\partial \\theta _ { t - \\tau } ^ { l } } | | _ { F } \\gg | | \\frac { \\partial \\bar { \\mu _ { t - \\tau } ^ { l } } ( \\theta _ { t - \\tau } ) } { \\partial \\theta _ { t - \\tau } ^ { r } } | | _ { F }$ and $| | \\frac { \\partial \\boldsymbol { v } _ { t - \\tau } ^ { l } ( \\theta _ { t - \\tau } ) } { \\partial \\theta _ { t - \\tau } ^ { l } } | | _ { F } \\gg | | \\frac { \\partial \\boldsymbol { v } _ { t - \\tau } ^ { l } ( \\theta _ { t - \\tau } ) } { \\partial \\theta _ { t - \\tau } ^ { r } } | | _ { F }$ || ∂ νlt−τ (θt−τ )∂ θ r− ||F for r < l, where || · ||F denotes the Frobenius norm. Here we τexamine this assumption empirically for networks trained on CIFAR-10 image recognition. ",
1600
+ "bbox": [
1601
+ 173,
1602
+ 467,
1603
+ 826,
1604
+ 561
1605
+ ],
1606
+ "page_idx": 12
1607
+ },
1608
+ {
1609
+ "type": "text",
1610
+ "text": "Figure 6 presents the computed ratio of $| | \\frac { \\partial \\mu _ { t - \\tau } ^ { l } ( \\theta _ { t - \\tau } ) } { \\partial \\theta _ { t - \\tau } ^ { r } } | | _ { F } / | | \\frac { \\partial \\mu _ { t - \\tau } ^ { l } ( \\theta _ { t - \\tau } ) } { \\partial \\theta _ { t - \\tau } ^ { l } } | | _ { F }$ and $| | \\frac { \\partial \\nu _ { t - \\tau } ^ { l } ( \\theta _ { t - \\tau } ) } { \\partial \\theta _ { t - \\tau } ^ { r } } | | _ { F } / | | \\frac { \\partial \\nu _ { t - \\tau } ^ { l } ( \\theta _ { t - \\tau } ) } { \\partial \\theta _ { t - \\tau } ^ { l } } | | _ { F }$ || ∂ νlt−τ (θt−τ )l ||F for r ≤ l, at different training epochs. The results suggest that $| | \\frac { \\partial \\boldsymbol { \\mu } _ { t - \\tau } ^ { l } ( \\boldsymbol { \\theta } _ { t - \\tau } ) } { \\partial \\boldsymbol { \\theta } _ { t - \\tau } ^ { l } } | | _ { F } \\gg | | \\frac { \\partial \\boldsymbol { \\mu } _ { t - \\tau } ^ { l } ( \\boldsymbol { \\theta } _ { t - \\tau } ) } { \\partial \\boldsymbol { \\theta } _ { t - \\tau } ^ { r } } | | _ { F }$ || ∂ µlt−τ (θt−τ )r ||F and | $| | \\frac { \\partial \\boldsymbol { v } _ { t - \\tau } ^ { l } ( \\boldsymbol { \\theta } _ { t - \\tau } ) } { \\partial \\boldsymbol { \\theta } _ { t - \\tau } ^ { l } } | | _ { F } \\gg | | \\frac { \\partial \\boldsymbol { v } _ { t - \\tau } ^ { l } ( \\boldsymbol { \\theta } _ { t - \\tau } ) } { \\partial \\boldsymbol { \\theta } _ { t - \\tau } ^ { r } } | | _ { F }$ || ∂ νlt−τ (θt−τ )r ||F hold for r < l, thus validating the approximation in Eq. (7) and Eq. (8). ",
1611
+ "bbox": [
1612
+ 173,
1613
+ 565,
1614
+ 826,
1615
+ 661
1616
+ ],
1617
+ "page_idx": 12
1618
+ },
1619
+ {
1620
+ "type": "text",
1621
+ "text": "We also study the gradients of non-ResNet models. $| | \\frac { \\partial \\mu _ { t - \\tau } ^ { l } ( \\theta _ { t - \\tau } ) } { \\partial \\theta _ { t - \\tau } ^ { l - 2 } } | | _ { F } / | | \\frac { \\partial \\mu _ { t - \\tau } ^ { l } ( \\theta _ { t - \\tau } ) } { \\partial \\theta _ { t - \\tau } ^ { l } } | | _ { F }$ and $| | \\frac { \\partial \\nu _ { t - \\tau } ^ { l } ( \\theta _ { t - \\tau } ) } { \\partial \\theta _ { t - \\tau } ^ { l - 2 } } | | _ { F } / | | \\frac { \\partial \\nu _ { t - \\tau } ^ { l } ( \\theta _ { t - \\tau } ) } { \\partial \\theta _ { t - \\tau } ^ { l } } | | _ { F }$ || ∂ νlt−τ (θt−τ )l ||F on VGG-16 and InceptionV3 are (0.22 and 0.46) and (0.17 and 0.38), respectively, which is similar to ResNet-18 (0.13 and 0.40), indicating that the assumption should also hold for VGG and the Inception series. ",
1622
+ "bbox": [
1623
+ 173,
1624
+ 666,
1625
+ 826,
1626
+ 748
1627
+ ],
1628
+ "page_idx": 12
1629
+ }
1630
+ ]
parse/train/BylJUTEKvB/BylJUTEKvB_middle.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/BylJUTEKvB/BylJUTEKvB_model.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/C0GmZH2RnVR/C0GmZH2RnVR.md ADDED
@@ -0,0 +1,254 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Breaking the Dilemma of Medical Image-to-image Translation
2
+
3
+ Lingke Kong∗ Manteia Tech konglingke@manteiatech.com
4
+
5
+ Chenyu Lian∗ Xiamen University cylian@stu.xmu.edu.cn
6
+
7
+ Detian Huang Huaqiao University huangdetian@hqu.edu.cn
8
+
9
+ Zhenjiang Li Shandong University zhenjli1987@163.com
10
+
11
+ Yanle Hu† Mayo Clinic Arizona Hu.Yanle@mayo.edu
12
+
13
+ Qichao Zhou† Manteia Tech zhouqc@manteiatech.com
14
+
15
+ # Abstract
16
+
17
+ Supervised Pix2Pix and unsupervised Cycle-consistency are two modes that dominate the field of medical image-to-image translation. However, neither modes are ideal. The Pix2Pix mode has excellent performance. But it requires paired and well pixel-wise aligned images, which may not always be achievable due to respiratory motion or anatomy change between times that paired images are acquired. The Cycle-consistency mode is less stringent with training data and works well on unpaired or misaligned images. But its performance may not be optimal. In order to break the dilemma of the existing modes, we propose a new unsupervised mode called RegGAN for medical image-to-image translation. It is based on the theory of "loss-correction". In RegGAN, the misaligned target images are considered as noisy labelsaand the generator is trained with an addi- ˘ tional registration network to fit the misaligned noise distribution adaptively. The goal is to search for the common optimal solution to both image-to-image translation and registration tasks. We incorporated RegGAN into a few state-of-the-art image-to-image translation methods and demonstrated that RegGAN could be easily combined with these methods to improve their performances. Such as a simple CycleGAN in our mode surpasses latest NICEGAN even though using less network parameters. Based on our results, RegGAN outperformed both Pix2Pix on aligned data and Cycle-consistency on misaligned or unpaired data. RegGAN is insensitive to noises which makes it a better choice for a wide range of scenarios, especially for medical image-to-image translation tasks in which well pixel-wise aligned data are not available. Code and data used in this study can be found at https://github.com/Kid-Liet/Reg-GAN.
18
+
19
+ # 1 Introduction
20
+
21
+ Generative adversarial networks (GANs)[1] is a framework that simultaneously trains a generator $G$ and a discriminator $D$ through an adversarial process. The generator is used to translate the distribu
22
+
23
+ tion of source domain images $X$ to the distribution of target domain images $Y$ . The discriminator is used to determine if the target domain images are likely from the generator or from the real data.
24
+
25
+ $$
26
+ \operatorname* { m i n } _ { G } \operatorname* { m a x } _ { D } \mathcal { L } _ { A d v } \left( G , D \right) = \mathbb { E } _ { y } \left[ l o g \left( D \left( y \right) \right) \right] + \mathbb { E } _ { x } \left[ l o g \left( 1 - D \left( G \left( x \right) \right) \right) \right]
27
+ $$
28
+
29
+ Supervised Pix2Pix[2] and unsupervised Cycle-consistency[3] are the two commonly used modes in GANs. Pix2Pix updates the generator $G : X Y$ ) by minimizing pixel-level $L 1$ loss between the source image $x$ and the target image $y$ . Therefore, it requires well aligned paired images, where each pixel has a corresponding label.
30
+
31
+ $$
32
+ \operatorname* { m i n } _ { G } { \mathcal { L } } _ { L 1 } \left( G \right) = \mathbb { E } _ { x , y } \left[ \left\| y - G \left( x \right) \right\| _ { 1 } \right]
33
+ $$
34
+
35
+ Well aligned paired images, however, are not always available in real-world scenarios. To address the challenges caused by misaligned images, Cycle-consistency was developed which was based on the assumption that the generator $G$ from the source domain $X$ to the target domain $Y$ $( G : X \to Y )$ was the reverse of the generator $F$ from $Y$ to $X$ $F : Y X$ ). Compared to the Pix2Pix mode, the Cycle-consistency mode works better on misaligned or unpaired images.
36
+
37
+ $$
38
+ \underset { G } { \mathop { \operatorname* { m i n } } } \underset { F } { \operatorname* { m i n } } \mathcal { L } _ { C y c } \left( G , F \right) = \mathbb { E } _ { x } \left[ \| F \left( G \left( x \right) \right) - x \| _ { 1 } \right] + \mathbb { E } _ { y } \left[ \| G \left( F \left( y \right) \right) - y \| _ { 1 } \right]
39
+ $$
40
+
41
+ The Cycle-consistency mode, however, has its limitations. In the field of medical image-to-image translation, it requires not only the style translation between image domains, but also the translation between specific pair of images. The optimal solution should be unique. For example, the translated images should maintain the anatomical features of the original images as much as possible. It is known that the Cycle-consistency mode may produce multiple solutions[4, 5], meaning that the training process may be relatively perturbing and the results may not be accurate. The pix2pix mode is not ideal either. Even though it has a unique solution, it is difficult to satisfy the requirement asking for well aligned paired images. With misaligned images, the errors are propagated through the Pix2Pix mode which may result in unreasonable displacements on the final translated images.
42
+
43
+ As of today, there is no image-to-image translation mode that can outperform both the Pix2Pix mode on aligned data and the Cycle-consistency mode on misaligned or unpaired data. Inspired by[6–10], we consider the misaligned target images as noisy labels, which means that the existing problem is regarded as supervised learning with noisy labels. So we introduce a new image-to-image translation mode called RegGAN. Figure 1 provides a comparison of the three modes: Pix2Pix, Cycle-consistency and RegGAN. To facilitate reading, we summarize our contributions as follows.
44
+
45
+ • We demonstrate the feasibility of RegGAN from the theoretical perspective of "losscorrection". Specifically, we train the generator using an additional registration network to fit the misaligned noise distribution adaptively, with the goal to search for the common optimal solution for both image-to-image translation and registration tasks. RegGAN eliminates the requirement for well aligned paired images and searches unique solution in training process. Based on our results, RegGAN outperformed both Pix2Pix on aligned data and Cycle-consistencyaon misaligned or unpaired data. ˘ RegGAN can be integrated into other methods without changing the original network architecture. Compared to Cycle-consistency with two generators and discriminators, RegGAN can provide better performance using less network parameters.
46
+
47
+ # 2 Related Work
48
+
49
+ Image-to-image Translation: Generative adversarial networks (GANs) have shown great potential in the field of image-to-image translation[11–16]. It has been successfully implemented in medical image analysis like segmentation[17], registration[18, 19] and dose calculation[20]. The existing modes, however, have their limitations. Specifically, the Pix2Pix mode[2] requires well aligned paired images which may not always be available. The Cycle-consistency mode can achieve unsupervised image-to-image translation. With a Cycle-consistency loss, it can be used for misaligned images. Based on Cycle-consistency, many methods[3, 21–30] have been developed including CycleGAN[3] and its variants such as MUNIT[31] and UNIT[32] in which both image content and style information are used to decouple and reconstruct the image-to-image translation task; U-gat-it[33] with a self-attention mechanism added; and NICEGAN[34] proposed to reuse the discriminator for encoding. The main limitation of Cycle-consistency is that it may produce multiple solutions and therefore is sensitive to perturbation, making it difficult to meet the high accuracy requirements of medical image-to-image translation tasks.
50
+
51
+ ![](images/af1c6b6474d3dfc147a4911d8cae1c1814796ffc4a02561f528abce736db83bd.jpg)
52
+ Figure 1: Comparison among the modes of Pix2Pix, CycleGAN and RegGAN.
53
+
54
+ Learning From Noisy Labels: Neural network anti-noise training has made great progress. Current research are mainly focused on: estimating the noise transition matrix[7, 35–40], designing a robust loss function[41–44], correcting the noise label[45–50], sampling importance weighting[51– 55] and meta-learning[56–59]. Our work is in the category of estimating the noise transition matrix. Compared to conventional noise transition estimation, we mitigate the issue and simplify the task by acquiring prior knowledge of noise distribution.
55
+
56
+ Deformable Registration: Traditional image registration methods have gained widespread acceptance, such as Demons[60], B-spline[61] and elastic deformation model[62]. One of the most popular deep learning methods is Voxelmorph[63]. In this study, a CNN model was trained to predict the deformable vector filed (DVF). The time-consuming gradient derivation process was thus skipped to improve the calculation efficiency. Affine registration and a vector momentum-parameterized stationary velocity field (vSVF)[64] was implemented to get better transformation regulation. Fast Symmetric method[65] used symmetric maximum similarity. Deep flash[66] outperformed other models in terms of training and calculation time.
57
+
58
+ Closest to our work, Arar.M et al[67] introduced a multi-modal registration method for natural images based on geometry preserving. But their work focused only on registration and did not demonstrate results of image-to-image translation or discuss the relationship between registration and image-to-image translation. The key insight of our work is that we demonstrated the feasibility of using registration to significantly improve the performance of image-to-image translation because the noise could be eliminated adaptively during the joint training process. What we propose in the paper is a completely new mode for medical image-to-image translation.
59
+
60
+ # 3 Methodology
61
+
62
+ # 3.1 Theoretical Motivation
63
+
64
+ If we consideramisaligned target images as noisy labels, ˘ athe training for image-to-image translation ˘ becomes aasupervised learning process ˘ awith noisy labels. Given a training dataset ˘ $\{ ( x _ { n } , \widetilde { y } _ { n } ) \} _ { n = 1 } ^ { N }$ with $N$ noisy labels in which $x _ { n }$ , $\widetilde { y } _ { n }$ are images from two modalities and assume $y _ { n }$ is the correct label for $x _ { n }$ e, but it is unknown in real-world scenarios. Our goal is to train a generator using the dataset $\{ ( x _ { n } , \widetilde { y } _ { n } ) \} _ { n = 1 } ^ { N }$ with noisy labels and achieve the performance equivalent to trained on clean dataset not wor $\left\{ \left( x _ { n } , y _ { n } \right) \right\} _ { n = 1 } ^ { N }$ as much as possible. Direct optimization based on Equations 4 usually does to bad results because the generator cannot squeeze out the influence of noise.
65
+
66
+ $$
67
+ \hat { G } = \underset { G } { \arg \operatorname* { m i n } } \frac { 1 } { N } \sum _ { n = 1 } ^ { N } \mathcal { L } \left( G \left( x _ { n } \right) , \widetilde { y } _ { n } \right)
68
+ $$
69
+
70
+ To address the noise issue, we propose a solution based on "loss-correction"[7] shown in Equations 5. Our solution corrects the output of the generator $G ( x _ { n } )$ by modeling a noise transition $\phi$ to match the noise distribution. Previously, Patrini et al[7] proved mathematically that the model trained with the noisy labels could be equivalent to the model trained with the clean labels, if the noise transition $\phi$ matches the noise distribution.
71
+
72
+ $$
73
+ \hat { G } = \underset { G } { \arg \operatorname* { m i n } } \frac { 1 } { N } \sum _ { n = 1 } ^ { N } \mathcal { L } \left( \phi \circ G \left( x _ { n } \right) , \widetilde { y } _ { n } \right)
74
+ $$
75
+
76
+ To achieve this, Goldberger et al[36] proposed to view the correct label as a latent random variable and explicitly model the label noise as a part of the network architecture, denoted by $R$ . Then, Equations 5 can be rewritten in the form of log-likelihood, which is used as the loss function for neural network training.
77
+
78
+ $$
79
+ \begin{array} { l } { \displaystyle \mathcal { L } \left( G , R \right) = - \sum _ { n = 1 } ^ { N } l o g \left( p \left( \widetilde { y } _ { n } | y _ { n } ; R \right) p \left( y _ { n } | x _ { n } ; G \right) \right) } \\ { \displaystyle = - \sum _ { n = 1 } ^ { N } l o g \left( p \left( \widetilde { y } _ { n } | x _ { n } ; G , R \right) \right) } \end{array}
80
+ $$
81
+
82
+ # 3.2 RegGAN
83
+
84
+ Compared to existing methods that use expectation-maximum[7, 36], fully connected layers[35], anchor point estimate[37] and Drichlet-distribution[38] to solve Equations 6. In our problem, the type of noise distribution is clearer, it can be expressed as displacement error: $\widetilde { y } = y \circ T$ . Here $T$ is expressed as a random deformation field, which produces random displacement for each pixel. So we adopt a registration network $R$ after the generator $G$ as label noise model to correct the results. The Correction loss is shown Equations 7:
85
+
86
+ $$
87
+ \operatorname* { m i n } _ { G , R } { \mathcal { L } } _ { C o r r } \left( G , R \right) = \mathbb { E } _ { x , \widetilde { y } } \left[ \left. \widetilde { y } - G \left( x \right) \circ R \left( G \left( x \right) , \widetilde { y } \right) \right. _ { 1 } \right]
88
+ $$
89
+
90
+ where, $R \left( G \left( x \right) , \widetilde { y } \right)$ is the deformation field and $\circ$ represents the resamples operation. The registration network is based on U-Net[68]. A smoothness loss[63] is defined in Equations 8 to evaluate the smoothness of the deformation field and minimize the gradient of the deformation field.
91
+
92
+ $$
93
+ \operatorname* { m i n } _ { R } \mathcal { L } _ { S m o o t h } \left( R \right) = \mathbb { E } _ { x , \widetilde { y } } \left[ \| \nabla R \left( G \left( x \right) , \widetilde { y } \right) \| ^ { 2 } \right]
94
+ $$
95
+
96
+ Finally, we add the Aversarial loss between the generator and the discriminator (Equations 1), and the total loss is expressed in Equations 9.
97
+
98
+ $$
99
+ \operatorname* { m i n } _ { G , R } \operatorname* { m a x } _ { D } \mathcal { L } _ { T o t a l } \left( G , R , D \right) = \mathcal { L } _ { C o r r } + \mathcal { L } _ { S m o o t h } + \mathcal { L } _ { A d v }
100
+ $$
101
+
102
+ # 4 Experiments
103
+
104
+ Performance evaluation of RegGAN was conducted through three investigations to 1) demonstrate the feasibility and superiority of the RegGAN mode in various methods, and 2) assess RegGANs sensitivity to noise, and 3) explore the availability of the RegGAN on unpaired data.
105
+
106
+ # 4.1 Dataset
107
+
108
+ The open-access dataset (BraTS 2018[69]) was used to evaluate the proposed RegGAN mode. The training dataset and testing dataset contained 8457 and 979 pairs of T1 and T2 MR images, respectively. BraTS 2018 was selected because the original images were paired and well aligned. We created misaligned images by randomly adding different levels of rotation, translation and rescaling to the original images. And we randomly sample one image from T1 and the other one from T2 when training on unpaired images. The availability of well aligned paired images, misaligned paired images, and unpaired images allow us to evaluate the performances of all three modes (Pix2Pix, Cycle-consistency and RegGAN).
109
+
110
+ # 4.2 Performances in Different Methods
111
+
112
+ The primary motivation of introducing RegGAN was to address challenges caused by misaligned data. Therefore, in this section, misaligned data were used in model training to demonstrate the feasibility and superiority of RegGAN. We selected the most popular CycleGAN[3] and its variants MUNIT[31], UNIT[32], and NICEGAN[34] as the methods for evaluation and compared the following four modes for each method.
113
+
114
+ • C(Cycle-consistency): The most primitive mode of all methods, with Cycle-consistency loss (Equations 3) as the main constraint. Two generators and two discriminators are required in this mode. $\mathbf { C } { \ + } \mathbf { R }$ (Cycle-consistency $^ +$ Registration): The RegGAN mode is combined with the mode C. Registration network $( R )$ and Correction loss (Equations 7) are added to the constraints. NC(Non Cycle-consistency): Only Adversarial loss (Equations 1) is used for updating. Compared to the mode C, Cycle-consistency loss is removed. Only one generator and one discriminator are required in this mode.
115
+ • $\mathbf { N C } { + } \mathbf { R }$ (Non Cycle-consistency $^ +$ Registration): A registration network $( R )$ and Correction loss (Equations 7) are added to the mode NC. It is the proposed RegGAN mode.
116
+
117
+ Table 1: Comparison of CycleGAN, MUNIT, UNIT and NICEGAN using four training modes(C, $\mathrm { C } { \ + } \mathrm { R }$ , NC and $\mathrm { N C } { + } \mathrm { R }$ ).
118
+
119
+ <table><tr><td>Modes Methods</td><td rowspan="3"></td><td rowspan="3">CycleGAN</td><td rowspan="3">MUNIT</td><td rowspan="3">UNIT</td><td rowspan="3">NICEGAN</td></tr><tr><td>Index</td><td></td></tr><tr><td>C</td><td>0.089 0.11</td></tr><tr><td rowspan="2">NMAE↓</td><td>C+R NC</td><td>(-0.012)0.077 0.11</td><td>(-0.022)0.088 0.10</td><td>0.087 (-0.013)0.074 0.098</td><td>0.082 (-0.011)0.071 0.089</td></tr><tr><td>NC+R C</td><td>(-0.038)0.072 23.5</td><td>(-0.021)0.079 20.6</td><td>(-0.027)0.071 24.6</td><td>(-0.019)0.070 25.2</td></tr><tr><td>PSNR ↑</td><td>C+R NC NC+R</td><td>(+0.3)23.8 20.2 (+5.4)25.6</td><td>(+2.1)22.7 21.5 (+2.3)23.8</td><td>(+0.7)25.3 23.7 (+1.8)25.5</td><td>(+0.9)26.1 23.5 (+2.8)26.3</td></tr><tr><td>SSIM↑</td><td>C C+R NC NC+R</td><td>0.83 (+0.02)0.85 0.79 (+0.07)0.86</td><td>0.80 (+0.03) 0.83 0.81 (+0.04)0.85</td><td>0.84 (+0.02)0.86 0.83 (+0.03) 0.86</td><td>0.83 (+0.03)0.86 0.84 (+0.02)0.86</td></tr></table>
120
+
121
+ To evaluate the performance of each method on misaligned data, we randomly added $[ - 5 , + 5 ]$ degrees of angle rotation, $[ - 5 , + 5 ]$ percent of translation, and $[ - 5 , + 5 ]$ percent of rescaling to the original T1 and T2 images on the training dataset.
122
+
123
+ To ensure fair comparison, we used the same training strategy and hyperparameters for all methods and modes (see supplementary materials for details). The Normalized Mean Absolute Error (NMAE), Peak Signal to Noise Ratio (PSNR) and Structural Similarity (SSIM) were used as metrics to evaluate the performances of trained models based on the testing dataset. To avoid false high results of index, we excluded the image background from the calculation. Table 1 summarized the results for all methods and modes under the current investigation.
124
+
125
+ ![](images/eedbfe72566d6ed42ea2f3ea9e4dccafb5a309ef45d96acc8b055ab8e2b1bf94.jpg)
126
+ Figure 2: The errors of different modes in different methods.
127
+
128
+ Based on the results from the Table 1, we can reach several conclusions. First, adding the registration network $\mathbf { \tau } ( + \mathbf { R } )$ significantly improves the performances of the methods. This is true for all methods in both C and NC modes. It clearly demonstrates that $\operatorname { R e g G A N }$ can be incorporated in various methods or combined with different network architectures to improve the performances. Second, the C mode is in general better than the NC mode for most of methods. Adding the registration network $\mathbf { \tau } ( + \mathbf { R } )$ improves the performance of the NC mode more than that of the C mode. In fact, our results show that the $\mathbf { N C } { + } \mathbf { R }$ mode is even better than the $\mathbf { C } { \ + } \mathbf { R }$ mode, implying that "Cycle-consistency loss" may play a negative role when it is combined with RegGAN. Compared with the commonly used C mode with two generators and two discriminators, RegGAN has fewer parameters but provides better performance. The simple CycleGAN method in the $\mathbf { N C } { + } \mathbf { R }$ mode outperforms the current state-of-the-art method NICEGAN in the C mode by 0.01, 0.4, 0.03 for NMAE, PSNR and SSIM, respectively. The $\mathbf { N C } { + } \mathbf { R }$ mode can also be used to improve the performance of NICEGAN. In fact, the performance of NICEGAN in the $\mathbf { N C } { + } \mathbf { R }$ mode is the best among all combinations of the 4 methods and 4 modes.
129
+
130
+ Figure 2 shows representative results from various combinations of the 4 methods (CycleGAN, MUNIT, UNIT and NICEGAN) and 4 modes (C, $\mathbf { C } { \ + } \mathbf { R }$ , NC and $\mathbf { N C + R }$ ). For all aspects of the image (from the tumor areas and the details), the combinations that use the registration network $\mathbf { \tau } ( + \mathbf { R } )$ always provide more realistic and accurate results than those that do not use the registration network $\mathbf { \tau } ( + \mathbf { R } )$ .
131
+
132
+ ![](images/1d3e8120380b91f9184bcfce0c98cc767b03f1f62b9eb2e29c1da17e47ccd98c.jpg)
133
+ Figure 3: Example images at seven different levels of introduced noise.
134
+
135
+ Table 2: Comparison of the NMAE, PSNR and SSIM for CycleGAN(C), Pix2Pix and RegGAN under 7 levels of noise.
136
+
137
+ <table><tr><td colspan="2"></td><td>Noise.0</td><td>Noise.1</td><td>Noise.2</td><td>Noise.3</td><td>Noise.4</td><td></td><td>Noise.5 Noise.NA</td></tr><tr><td rowspan="3">Setting</td><td>Rotate</td><td>0°</td><td>±1°</td><td>±2°</td><td>±3°</td><td>±4°</td><td>±5°</td><td>X</td></tr><tr><td>Translation</td><td>0%</td><td>±2%</td><td>±4%</td><td>±6%</td><td>±8%</td><td>±10%</td><td>X</td></tr><tr><td>Rescaling</td><td>0%</td><td>±2%</td><td>±4%</td><td>±6%</td><td>±8%</td><td>±10%</td><td>X</td></tr><tr><td rowspan="3">CycleGAN(C)</td><td>NMAE↓</td><td>0.084</td><td>0.095</td><td>0.087</td><td>0.094</td><td>0.087</td><td>0.110</td><td>0.091</td></tr><tr><td>PSNR ↑</td><td>23.9</td><td>22.5</td><td>23.7</td><td>23.3</td><td>23.9</td><td>23.7</td><td>23.5</td></tr><tr><td>SSIM↑</td><td>0.83</td><td>0.83</td><td>0.82</td><td>0.81</td><td>0.82</td><td>0.79</td><td>0.83</td></tr><tr><td rowspan="3">Pix2Pix</td><td>NMAE ↓</td><td>0.075</td><td>0.103</td><td>0.139</td><td>0.161</td><td>0.175</td><td>0.181</td><td>0.086</td></tr><tr><td>PSNR ↑</td><td>25.6</td><td>22.3</td><td>18.9</td><td>16.2</td><td>15.3</td><td>15.0</td><td></td></tr><tr><td>SSIM ↑</td><td>0.85</td><td>0.82</td><td>0.78</td><td>0.76</td><td>0.74</td><td>0.74</td><td>21.1</td></tr><tr><td rowspan="3">RegGAN</td><td>NMAE ↓</td><td>0.071</td><td>0.073</td><td>0.071</td><td>0.072</td><td>0.072</td><td>0.072</td><td>0.82 0.071</td></tr><tr><td>PSNR ↑</td><td>26</td><td>25.6</td><td>25.9</td><td>25.7</td><td>25.4</td><td>25.2</td><td></td></tr><tr><td>SSIM ↑</td><td>0.86</td><td>0.86</td><td>0.86</td><td>0.86</td><td>0.86</td><td>0.85</td><td>25.9 0.86</td></tr></table>
138
+
139
+ # 4.3 Performances in Different Noise Levels
140
+
141
+ To evaluate the sensitivity of $\operatorname { R e g G A N }$ to noise, we selected a simple network architecture.(CycleGAN) with the intention to minimize interference from other factors. The same network architecture was used for all three modes: CycleGAN(C), Pix2Pix and RegGAN. Seven levels of noise were used in the evaluation. Table 2 lists the specific noise setting and range for each noise level. Noise.0 means the original dataset with no added noise. Noise.5 is the highest level of noise. At Noise.5, the data are likely from different patients. In addition, we also made non-affine noise(Noise.NA) settings. The non-affine noise is generated by spatially transforming T1 and T2 using elastic transformations on control points followed by Gaussian smoothing. Figure 3 shows example images at different levels of introduced noise.
142
+
143
+ Table 2 lists the quantitative evaluation metrics from 3 modes at 7 levels of noise. It is clear that RegGAN outperforms CycleGAN(C) under all noise levels. Figure 4(a) shows the test results from each epoch during the training process for both RegGAN and CycleGAN(C). Curves of different colors corresponds to different levels of noise. We notice that CycleGAN(C) is not very stable during the training process. The test results fluctuate significantly and cannot converge well. This may be caused by the fact that the solution of CycleGAN (C) is not unique. As a comparison, RegGAN is quite stable. Although the results from different levels of noise may vary at the beginning of training, all curves converge to a similar result after multiple epoches of training, indicating that RegGAN is more robust to noise compared to CycleGAN (C).
144
+
145
+ Based on Table 2, we notice that the performance of Pix2Pix deteriorates rapidly as the noise increases. This is as expected because Pix2Pix requires well aligned paired images. Surprisingly, the performances of RegGAN at all noise levels exceed those of Pix2Pix with no noise. Figure 4(b) shows the test results at each epoch of RegGAN and Pix2Pix under Noise.0 (i.e., no noise). Theoretically, the performances of RegGAN and Pix2Pix should be similar on perfectly aligned paired datasets because the registration network of RegGAN does not help and RegGAN is equivalent to
146
+
147
+ ![](images/f662bd1bbeb7045dac456003b31281c1beb1a993af4d26cdff8fba4ce3a7d075.jpg)
148
+ Figure 4: Quantitative evaluation metrics at different epochs in the training process. (a) Comparison of CycleGAN and $\operatorname { R e g G A N }$ at different levels of noise. (b) Comparison of $\mathrm { P i x 2 P i x }$ and $\mathsf { R e g G A N }$ at Noise.0 (i.e., no noise). (c) RegGAN’s Smoothness loss under different levels of noise.
149
+
150
+ ![](images/326a09b807acfdbf4500bf0e29287859625c86c57910fc62e0d5bca1bce29663.jpg)
151
+ Figure 5: The misalignment of orginal image pairs and corresponding deformation fields.
152
+
153
+ Pix2Pix. A possible explanation to our results is that in the medical field, the perfectly pixel-wise aligned dataset may not practically exist. Even for BraTS 2018[63] which is recognized as well aligned, it is still possible that there exists slight misalignment. As a result, adding the registration network is always likely to improve the performances in real-world scenarios. To verify our explanation, we plotted the Smoothness loss of RegGAN under different noise levels as shown in Figure 4(c). Large Smoothness loss corresponds to large deformation field displacement. First, we notice that the Smoothness loss under Noise.0 never completely goes to 0, indicating the existence of misalignment and potential usefulness of the registration network. Second, the noise level and Smoothness loss show a step-like positive correlation, which means that RegGAN can adaptively handle the noise distribution, i.e., the registration network can determine the range of deformation according to the noise level. In addition, we can see that even under the setting of non-affine noise, the above conclusion still holds. Because what the registration network corrects is deformation noise.
154
+
155
+ we also show some original image pairs and visualize the corresponding deformation fields output by registration network in Figure 5. Obviously, there is some misalignment between the original T1 and T2 images, and such misalignment is represented by the deformation fields (highlighted by red circle).
156
+
157
+ ![](images/27d4890be00f9554f08f14f90822eb89bc97096dee32bda414c7eac84631e0c8.jpg)
158
+ Figure 6: Performance comparison of the three modes (CycleGAN(C), Pix2Pix and RegGAN) on unpaired dataset.
159
+
160
+ ![](images/048a4fd457b36a04d9cc483fe88373cc9c0bedef0e626f7bb79deca8ef1b4a45.jpg)
161
+ Figure 7: Display of RegGAN’s output on unpaired data. T1 and T2 are unpaired images. The Translated represents the translation result of T1 to T2. Registered represents the registration result of the translated images. D.F represents deformation fields.
162
+
163
+ # 4.4 Performances on Unpaired Dataset
164
+
165
+ So far, our investigations are based on paired datasets. We also want to explore how RegGAN performs using unpaired datasets. In practice, this is not recommended because even different patients may have similarities in their body tissues of adjacent layers. For unpaired datasets, we can conduct rigid registration first in 3D space and then use RegGAN for training. Unpaired data can be treated as having larger scale noise. If the correction capability is strong enough, RegGAN can still work effectively. The comparison of the performances of three modes on the unpaired dataset is shown in Figure 6.
166
+
167
+ With unpaired datasets, Pix2Pix no longer considers the characteristics of the input T1 images and thus has the worst performance. Due to the challenges in fitting the noise, the performance improvement from replacing CycleGAN(C) with RegGAN using unpaired datasets may not be as dramatic as that demonstrated using paired datasets, but RegGAN still has the best performance under unpaired conditions. In Figure 7, we show some examples of how RegGAN corrects noise on unpaired dataset. It can be seen that RegGAN will try its best to eliminate the influnce of noise through registration.
168
+
169
+ Based on our results, it is reasonable to reach the conclusions below. In all circumstances, RegGAN demonstrates better performance compared to Pix2Pix and CycleGAN(C).
170
+
171
+ • For paired and aligned conditions, RegGA $\mathsf { N } \geq \mathsf { P i x 2 P i x } > \mathsf { C y c l e }$ GAN(C).
172
+
173
+ • For paired but misaligned conditions, RegGAN $>$ CycleGAN(C) $>$ Pix2Pix.
174
+
175
+ • For unpaired conditions, RegGAN $>$ CycleGAN(C) $>$ Pix2Pix.
176
+
177
+ # Conclusion
178
+
179
+ In this study, we introduced a new image-to-image translation mode RegGAN to the medical community that can break the dilemma of image-to-image translation task. Using a public BraTS 2018 dataset, we demonstrated the feasibility of RegGAN and its superior performance compared to Pix2Pix and Cycle-consistency. We validated that RegGAN can be incorporated into various existing methods to improve their performances. We also evaluated the sensitivity of RegGAN to noise. Our results confirmed that RegGAN could adapt well to various scenarios from no noise to large-scale noise. The superior performance of RegGAN makes it a better choice over Pix2Pix and Cycle-consistency whether datasets are aligned or not. However, this mode may not work well on natural images. The noise may cannot be considered simply as deformation errors due to the differences in natural images are much greater than those in medical images.
180
+
181
+ # Broader Impact
182
+
183
+ Image-to-image translation has been one of the main focuses in medical image analysis, as it aids in diagnosis and treatment. Previously, physicians had to use different medical imaging equipments if they wanted to get different image sequences of a patient, which was time-consuming and expensive. Pix2Pix mode is expected to solve this problem by its outstanding performance in image-to-image translation. In most of clinical scenarios, however, it is not practical to create such a large well aligned dataset for Pix2Pix mode. Cycle-consistence mode does not need well aligned dataset but can not meet the high-precision requirements of medical image analysis. Our work aims to provide a general image-to-image translation mode, which not only has no strict requirements on the dataset, but also can meet the clinical requirements in terms of image quality. In the future, we will attempt to obtain multi-modal dataset(eg MR-CT) for clinical verification. We foresee positive impacts if the mode is applied to diagnosis in radiology, treatment planning and research.
184
+
185
+ References
186
+ [1] Goodfellow I, Pouget-Abadie J, Mirza M, et al. Generative adversarial nets[C]//Ghahramani Z, Welling M, Cortes C, et al. Advances in Neural Information Processing Systems. Curran Associates, Inc., 2014.
187
+ [2] Isola P, Zhu J Y, Zhou T, et al. Image-to-image translation with conditional adversarial networks[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 1125-1134.
188
+ [3] Zhu J Y, Park T, Isola P, et al. Unpaired image-to-image translation using cycle-consistent adversarial networks[C]//Proceedings of the IEEE international conference on computer vision. 2017: 2223-2232.
189
+ [4] Sim B, Oh G, Lim S, et al. Optimal transport, cyclegan, and penalized ls for unsupervised learning in inverse problems[J]. 2019.
190
+ [5] Moriakov N, Adler J, Teuwen J. Kernel of cyclegan as a principal homogeneous space[C]// International Conference on Learning Representations. 2020.
191
+ [6] Huang L, Zhang C, Zhang H. Self-adaptive training: beyond empirical risk minimization[J]. Advances in Neural Information Processing Systems, 2020, 33.
192
+ [7] Patrini G, Rozza A, Krishna Menon A, et al. Making deep neural networks robust to label noise: A loss correction approach[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2017: 1944-1952.
193
+ [8] Dgani Y, Greenspan H, Goldberger J. Training a neural network based on unreliable human annotation of medical images[C]//2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018). IEEE, 2018: 39-42.
194
+ [9] Yao Y, Liu T, Han B, et al. Dual t: Reducing estimation error for transition matrix in labelnoise learning[C/OL]//Larochelle H, Ranzato M, Hadsell R, et al. Advances in Neural Information Processing Systems: volume 33. Curran Associates, Inc., 2020: 7260-7271. https: //proceedings.neurips.cc/paper/2020/file/512c5cad6c37edb98ae91c8a76c3a291-Paper.pdf.
195
+ [10] Zhang L, Tanno R, Xu M C, et al. Disentangling human error from ground truth in segmentation of medical images[C/OL]//Larochelle H, Ranzato M, Hadsell R, et al. Advances in Neural Information Processing Systems: volume 33. Curran Associates, Inc., 2020: 15750-15762. https://proceedings.neurips.cc/paper/2020/file/ b5d17ed2b502da15aa727af0d51508d6-Paper.pdf.
196
+ [11] wang y, Yu L, van de Weijer J. Deepi2i: Enabling deep hierarchical image-toimage translation by transferring from gans[C/OL]//Larochelle H, Ranzato M, Hadsell R, et al. Advances in Neural Information Processing Systems: volume 33. Curran Associates, Inc., 2020: 11803-11815. https://proceedings.neurips.cc/paper/2020/file/ 88855547570f7ff053fff7c54e5148cc-Paper.pdf.
197
+ [12] Zhang P, Zhang B, Chen D, et al. Cross-domain correspondence learning for exemplar-based image translation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020: 5143-5153.
198
+ [13] Kim T, Cha M, Kim H, et al. Learning to discover cross-domain relations with generative adversarial networks[C]//International Conference on Machine Learning. PMLR, 2017: 1857- 1865.
199
+ [14] Cao Y, Wan X. Divgan: Towards diverse paraphrase generation via diversified generative adversarial network[C]//Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings. 2020: 2411-2421.
200
+ [15] Choi Y, Choi M, Kim M, et al. Stargan: Unified generative adversarial networks for multidomain image-to-image translation[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 8789-8797.
201
+ [16] Choi Y, Uh Y, Yoo J, et al. Stargan v2: Diverse image synthesis for multiple domains[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020: 8188-8197.
202
+ [17] Chen C, Dou Q, Chen H, et al. Synergistic image and feature adaptation: Towards crossmodality domain adaptation for medical image segmentation[C]//Proceedings of the AAAI Conference on Artificial Intelligence: volume 33. 2019: 865-872.
203
+ [18] Fan J, Cao X, Wang Q, et al. Adversarial learning for mono-or multi-modal registration[J]. Medical image analysis, 2019, 58:101545.
204
+ [19] Qin C, Shi B, Liao R, et al. Unsupervised deformable registration for multi-modal images via disentangled representations[C]//International Conference on Information Processing in Medical Imaging. Springer, 2019: 249-261.
205
+ [20] Sahiner B, Pezeshk A, Hadjiiski L M, et al. Deep learning in medical imaging and radiation therapy[J]. Medical physics, 2019, 46(1):e1-e36.
206
+ [21] Hoffman J, Tzeng E, Park T, et al. Cycada: Cycle-consistent adversarial domain adaptation [C]//International conference on machine learning. PMLR, 2018: 1989-1998.
207
+ [22] Li M, Huang H, Ma L, et al. Unsupervised image-to-image translation with stacked cycleconsistent adversarial networks[C]//Proceedings of the European conference on computer vision (ECCV). 2018: 184-199.
208
+ [23] Deng W, Zheng L, Ye Q, et al. Image-image domain adaptation with preserved self-similarity and domain-dissimilarity for person re-identification[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 994-1003.
209
+ [24] Li J, Chen E, Ding Z, et al. Cycle-consistent conditional adversarial transfer networks[C]// Proceedings of the 27th ACM International Conference on Multimedia. 2019: 747-755.
210
+ [25] Schmidt V, Luccioni A, Mukkavilli S K, et al. Visualizing the consequences of climate change using cycle-consistent adversarial networks[J]. arXiv preprint arXiv:1905.03709, 2019.
211
+ [26] Chen Z, Li J, Luo Y, et al. Canzsl: Cycle-consistent adversarial networks for zero-shot learning from natural language[C]//Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2020: 874-883.
212
+ [27] Ren C X, Ziemann A, Theiler J, et al. Cycle-consistent adversarial networks for realistic pervasive change generation in remote sensing imagery[C]//2020 IEEE Southwest Symposium on Image Analysis and Interpretation (SSIAI). IEEE, 2020: 42-45.
213
+ [28] Liu J, Ding Y, Xiong J, et al. Multi-cycle-consistent adversarial networks for ct image denoising[C]//2020 IEEE 17th International Symposium on Biomedical Imaging (ISBI). IEEE, 2020: 614-618.
214
+ [29] Zheng C, Pan L, Wu P. Camu: Cycle-consistent adversarial mapping model for user alignment across social networks[J]. IEEE Transactions on Cybernetics, 2021.
215
+ [30] Shah M, Chen X, Rohrbach M, et al. Cycle-consistency for robust visual question answering [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019: 6649-6658.
216
+ [31] Huang X, Liu M Y, Belongie S, et al. Multimodal unsupervised image-to-image translation [C]//Proceedings of the European conference on computer vision (ECCV). 2018: 172-189.
217
+ [32] Liu M Y, Breuel T, Kautz J. Unsupervised image-to-image translation networks[C/OL]// Guyon I, Luxburg U V, Bengio S, et al. Advances in Neural Information Processing Systems: volume 30. Curran Associates, Inc., 2017. https://proceedings.neurips.cc/paper/2017/ file/dc6a6489640ca02b0d42dabeb8e46bb7-Paper.pdf.
218
+ [33] Kim J, Kim M, Kang H, et al. U-gat-it: Unsupervised generative attentional networks with adaptive layer-instance normalization for image-to-image translation[C/OL]// International Conference on Learning Representations. 2020. https://openreview.net/forum? id=BJlZ5ySKPH.
219
+ [34] Chen R, Huang W, Huang B, et al. Reusing discriminators for encoding: Towards unsupervised image-to-image translation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020: 8168-8177.
220
+ [35] S. Sukhbaatar M P L B R F, J. Bruna. Training convolutional networks with noisy labels[C]// International Conference on Learning Representations. 2015.
221
+ [36] Goldberger J, Ben-Reuven E. Training deep neural-networks using a noise adaptation layer[J]. 2016.
222
+ [37] Xia X, Liu T, Wang N, et al. Are anchor points really indispensable in labelnoise learning?[C/OL]//NeurIPS. 2019: 6835-6846. http://papers.nips.cc/paper/ 8908-are-anchor-points-really-indispensable-in-label-noise-learning.
223
+ [38] Yao J, Wu H, Zhang Y, et al. Safeguarded dynamic label regression for noisy supervision[C]// Proceedings of the AAAI Conference on Artificial Intelligence: volume 33. 2019: 9103-9110.
224
+ [39] Xiao T, Xia T, Yang Y, et al. Learning from massive noisy labeled data for image classification [C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2015: 2691-2699.
225
+ [40] Misra I, Lawrence Zitnick C, Mitchell M, et al. Seeing through the human reporting bias: Visual classifiers from noisy human-centric labels[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2016: 2930-2939.
226
+ [41] Tanno R, Saeedi A, Sankaranarayanan S, et al. Learning from noisy labels by regularized estimation of annotator confusion[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019: 11244-11253.
227
+ [42] Rodrigues F, Pereira F. Deep learning from crowds[C]//Proceedings of the AAAI Conference on Artificial Intelligence: volume 32. 2018.
228
+ [43] Branson S, Van Horn G, Perona P. Lean crowdsourcing: Combining humans and machines in an online system[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2017: 7474-7483.
229
+ [44] Izadinia H, Russell B C, Farhadi A, et al. Deep classifiers from image tags in the wild[C]// Proceedings of the 2015 Workshop on Community-Organized Multimodal Mining: Opportunities for Novel Solutions. 2015: 13-18.
230
+ [45] Jaehwan L, Donggeun Y, Hyo-Eun K. Photometric transformer networks and label adjustment for breast density prediction[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. 2019: 0-0.
231
+ [46] Veit A, Alldrin N, Chechik G, et al. Learning from noisy large-scale datasets with minimal supervision[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 839-847.
232
+ [47] Tanaka D, Ikami D, Yamasaki T, et al. Joint optimization framework for learning with noisy labels[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018: 5552-5560.
233
+ [48] Zheng S, Wu P, Goswami A, et al. Error-bounded correction of noisy labels[C]//International Conference on Machine Learning. PMLR, 2020: 11447-11457.
234
+ [49] Han J, Luo P, Wang X. Deep self-learning from noisy labels[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. 2019: 5138-5147.
235
+ [50] Durand T, Mehrasa N, Mori G. Learning a deep convnet for multi-label classification with partial labels[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019: 647-657.
236
+ [51] Wu X, He R, Sun Z, et al. A light cnn for deep face representation with noisy labels[J]. IEEE Transactions on Information Forensics and Security, 2018, 13(11):2884-2896.
237
+ [52] Huang J, Qu L, Jia R, et al. O2u-net: A simple noisy label detection approach for deep neural networks[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. 2019: 3326-3334.
238
+ [53] Li J, Socher R, Hoi S C. Dividemix: Learning with noisy labels as semi-supervised learning [C/OL]//International Conference on Learning Representations. 2020. https://openreview.net/ forum?id=HJgExaVtwr.
239
+ [54] Yan Y, Xu Z, Tsang I, et al. Robust semi-supervised learning through label aggregation[C]// Proceedings of the AAAI Conference on Artificial Intelligence: volume 30. 2016.
240
+ [55] Jiang J, Ma J, Wang Z, et al. Hyperspectral image classification in the presence of noisy labels [J]. IEEE Transactions on Geoscience and Remote Sensing, 2018, 57(2):851-865.
241
+ [56] Li J, Wong Y, Zhao Q, et al. Learning to learn from noisy labeled data[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019: 5051-5059.
242
+ [57] Li Y, Yang J, Song Y, et al. Learning from noisy labels with distillation[C]//Proceedings of the IEEE International Conference on Computer Vision. 2017: 1910-1918.
243
+ [58] Algan G, Ulusoy I. Meta soft label generation for noisy labels[C]//2020 25th International Conference on Pattern Recognition (ICPR). IEEE, 2021: 7142-7148.
244
+ [59] Garcia L P, de Carvalho A C, Lorena A C. Noise detection in the meta-learning level[J]. Neurocomputing, 2016, 176:14-25.
245
+ [60] Thirion J P. Image matching as a diffusion process: an analogy with maxwell’s demons[J]. Medical image analysis, 1998, 2:243260.
246
+ [61] Rueckert D, Sonoda L I, Hayes C, et al. Nonrigid registration using free-form deformations: application to breast mr images[J]. IEEE transactions on medical imaging, 1999, 18(8):712- 721.
247
+ [62] Shen D, Davatzikos C. Hammer: hierarchical attribute matching mechanism for elastic registration[J]. IEEE transactions on medical imaging, 2002, 21(11):1421-1439.
248
+ [63] Balakrishnan G, Zhao A, Sabuncu M R, et al. Voxelmorph: a learning framework for deformable medical image registration[J]. IEEE transactions on medical imaging, 2019, 38(8): 1788-1800.
249
+ [64] Shen Z, Han X, Xu Z, et al. Networks for joint affine and non-parametric image registration [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2019: 4224-4233.
250
+ [65] Mok T C, Chung A. Fast symmetric diffeomorphic image registration with convolutional neural networks[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020: 4644-4653.
251
+ [66] Wang J, Zhang M. Deepflash: An efficient network for learning-based medical image registration[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020: 4444-4452.
252
+ [67] Arar M, Ginger Y, Danon D, et al. Unsupervised multi-modal image registration via geometry preserving image-to-image translation[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020: 13410-13419.
253
+ [68] Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation[C]//International Conference on Medical image computing and computer-assisted intervention. Springer, 2015: 234-241.
254
+ [69] Menze B H, Jakab A, Bauer S, et al. The multimodal brain tumor image segmentation benchmark (brats)[J]. IEEE transactions on medical imaging, 2014, 34(10):1993-2024.
parse/train/C0GmZH2RnVR/C0GmZH2RnVR_content_list.json ADDED
@@ -0,0 +1,1007 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "type": "text",
4
+ "text": "Breaking the Dilemma of Medical Image-to-image Translation ",
5
+ "text_level": 1,
6
+ "bbox": [
7
+ 196,
8
+ 122,
9
+ 803,
10
+ 171
11
+ ],
12
+ "page_idx": 0
13
+ },
14
+ {
15
+ "type": "text",
16
+ "text": "Lingke Kong∗ Manteia Tech konglingke@manteiatech.com ",
17
+ "bbox": [
18
+ 251,
19
+ 222,
20
+ 478,
21
+ 263
22
+ ],
23
+ "page_idx": 0
24
+ },
25
+ {
26
+ "type": "text",
27
+ "text": "Chenyu Lian∗ Xiamen University cylian@stu.xmu.edu.cn ",
28
+ "bbox": [
29
+ 562,
30
+ 222,
31
+ 745,
32
+ 263
33
+ ],
34
+ "page_idx": 0
35
+ },
36
+ {
37
+ "type": "text",
38
+ "text": "Detian Huang Huaqiao University huangdetian@hqu.edu.cn ",
39
+ "bbox": [
40
+ 205,
41
+ 285,
42
+ 398,
43
+ 327
44
+ ],
45
+ "page_idx": 0
46
+ },
47
+ {
48
+ "type": "text",
49
+ "text": "Zhenjiang Li Shandong University zhenjli1987@163.com ",
50
+ "bbox": [
51
+ 436,
52
+ 285,
53
+ 602,
54
+ 327
55
+ ],
56
+ "page_idx": 0
57
+ },
58
+ {
59
+ "type": "text",
60
+ "text": "Yanle Hu† Mayo Clinic Arizona Hu.Yanle@mayo.edu ",
61
+ "bbox": [
62
+ 643,
63
+ 285,
64
+ 789,
65
+ 327
66
+ ],
67
+ "page_idx": 0
68
+ },
69
+ {
70
+ "type": "text",
71
+ "text": "Qichao Zhou† Manteia Tech zhouqc@manteiatech.com ",
72
+ "bbox": [
73
+ 403,
74
+ 348,
75
+ 594,
76
+ 390
77
+ ],
78
+ "page_idx": 0
79
+ },
80
+ {
81
+ "type": "text",
82
+ "text": "Abstract ",
83
+ "text_level": 1,
84
+ "bbox": [
85
+ 462,
86
+ 426,
87
+ 535,
88
+ 443
89
+ ],
90
+ "page_idx": 0
91
+ },
92
+ {
93
+ "type": "text",
94
+ "text": "Supervised Pix2Pix and unsupervised Cycle-consistency are two modes that dominate the field of medical image-to-image translation. However, neither modes are ideal. The Pix2Pix mode has excellent performance. But it requires paired and well pixel-wise aligned images, which may not always be achievable due to respiratory motion or anatomy change between times that paired images are acquired. The Cycle-consistency mode is less stringent with training data and works well on unpaired or misaligned images. But its performance may not be optimal. In order to break the dilemma of the existing modes, we propose a new unsupervised mode called RegGAN for medical image-to-image translation. It is based on the theory of \"loss-correction\". In RegGAN, the misaligned target images are considered as noisy labelsaand the generator is trained with an addi- ˘ tional registration network to fit the misaligned noise distribution adaptively. The goal is to search for the common optimal solution to both image-to-image translation and registration tasks. We incorporated RegGAN into a few state-of-the-art image-to-image translation methods and demonstrated that RegGAN could be easily combined with these methods to improve their performances. Such as a simple CycleGAN in our mode surpasses latest NICEGAN even though using less network parameters. Based on our results, RegGAN outperformed both Pix2Pix on aligned data and Cycle-consistency on misaligned or unpaired data. RegGAN is insensitive to noises which makes it a better choice for a wide range of scenarios, especially for medical image-to-image translation tasks in which well pixel-wise aligned data are not available. Code and data used in this study can be found at https://github.com/Kid-Liet/Reg-GAN. ",
95
+ "bbox": [
96
+ 233,
97
+ 458,
98
+ 766,
99
+ 776
100
+ ],
101
+ "page_idx": 0
102
+ },
103
+ {
104
+ "type": "text",
105
+ "text": "1 Introduction ",
106
+ "text_level": 1,
107
+ "bbox": [
108
+ 174,
109
+ 801,
110
+ 310,
111
+ 819
112
+ ],
113
+ "page_idx": 0
114
+ },
115
+ {
116
+ "type": "text",
117
+ "text": "Generative adversarial networks (GANs)[1] is a framework that simultaneously trains a generator $G$ and a discriminator $D$ through an adversarial process. The generator is used to translate the distribu",
118
+ "bbox": [
119
+ 174,
120
+ 833,
121
+ 821,
122
+ 862
123
+ ],
124
+ "page_idx": 0
125
+ },
126
+ {
127
+ "type": "text",
128
+ "text": "tion of source domain images $X$ to the distribution of target domain images $Y$ . The discriminator is used to determine if the target domain images are likely from the generator or from the real data. ",
129
+ "bbox": [
130
+ 171,
131
+ 90,
132
+ 825,
133
+ 119
134
+ ],
135
+ "page_idx": 1
136
+ },
137
+ {
138
+ "type": "equation",
139
+ "img_path": "images/dade2c019b0560448bdd52e608999246358f4532037b006375c9be4593f99f56.jpg",
140
+ "text": "$$\n\\operatorname* { m i n } _ { G } \\operatorname* { m a x } _ { D } \\mathcal { L } _ { A d v } \\left( G , D \\right) = \\mathbb { E } _ { y } \\left[ l o g \\left( D \\left( y \\right) \\right) \\right] + \\mathbb { E } _ { x } \\left[ l o g \\left( 1 - D \\left( G \\left( x \\right) \\right) \\right) \\right]\n$$",
141
+ "text_format": "latex",
142
+ "bbox": [
143
+ 267,
144
+ 126,
145
+ 728,
146
+ 148
147
+ ],
148
+ "page_idx": 1
149
+ },
150
+ {
151
+ "type": "text",
152
+ "text": "Supervised Pix2Pix[2] and unsupervised Cycle-consistency[3] are the two commonly used modes in GANs. Pix2Pix updates the generator $G : X Y$ ) by minimizing pixel-level $L 1$ loss between the source image $x$ and the target image $y$ . Therefore, it requires well aligned paired images, where each pixel has a corresponding label. ",
153
+ "bbox": [
154
+ 173,
155
+ 154,
156
+ 825,
157
+ 210
158
+ ],
159
+ "page_idx": 1
160
+ },
161
+ {
162
+ "type": "equation",
163
+ "img_path": "images/6edc84ad30cbfa49c43979599b8e8906becc2076cf8dbbce66e89f54edef8645.jpg",
164
+ "text": "$$\n\\operatorname* { m i n } _ { G } { \\mathcal { L } } _ { L 1 } \\left( G \\right) = \\mathbb { E } _ { x , y } \\left[ \\left\\| y - G \\left( x \\right) \\right\\| _ { 1 } \\right]\n$$",
165
+ "text_format": "latex",
166
+ "bbox": [
167
+ 375,
168
+ 217,
169
+ 622,
170
+ 239
171
+ ],
172
+ "page_idx": 1
173
+ },
174
+ {
175
+ "type": "text",
176
+ "text": "Well aligned paired images, however, are not always available in real-world scenarios. To address the challenges caused by misaligned images, Cycle-consistency was developed which was based on the assumption that the generator $G$ from the source domain $X$ to the target domain $Y$ $( G : X \\to Y )$ was the reverse of the generator $F$ from $Y$ to $X$ $F : Y X$ ). Compared to the Pix2Pix mode, the Cycle-consistency mode works better on misaligned or unpaired images. ",
177
+ "bbox": [
178
+ 174,
179
+ 244,
180
+ 825,
181
+ 315
182
+ ],
183
+ "page_idx": 1
184
+ },
185
+ {
186
+ "type": "equation",
187
+ "img_path": "images/49f6230c305fd3cc7db635af55af4ede259c3e4e150c429c648894a411f3cf87.jpg",
188
+ "text": "$$\n\\underset { G } { \\mathop { \\operatorname* { m i n } } } \\underset { F } { \\operatorname* { m i n } } \\mathcal { L } _ { C y c } \\left( G , F \\right) = \\mathbb { E } _ { x } \\left[ \\| F \\left( G \\left( x \\right) \\right) - x \\| _ { 1 } \\right] + \\mathbb { E } _ { y } \\left[ \\| G \\left( F \\left( y \\right) \\right) - y \\| _ { 1 } \\right]\n$$",
189
+ "text_format": "latex",
190
+ "bbox": [
191
+ 254,
192
+ 321,
193
+ 743,
194
+ 344
195
+ ],
196
+ "page_idx": 1
197
+ },
198
+ {
199
+ "type": "text",
200
+ "text": "The Cycle-consistency mode, however, has its limitations. In the field of medical image-to-image translation, it requires not only the style translation between image domains, but also the translation between specific pair of images. The optimal solution should be unique. For example, the translated images should maintain the anatomical features of the original images as much as possible. It is known that the Cycle-consistency mode may produce multiple solutions[4, 5], meaning that the training process may be relatively perturbing and the results may not be accurate. The pix2pix mode is not ideal either. Even though it has a unique solution, it is difficult to satisfy the requirement asking for well aligned paired images. With misaligned images, the errors are propagated through the Pix2Pix mode which may result in unreasonable displacements on the final translated images. ",
201
+ "bbox": [
202
+ 173,
203
+ 357,
204
+ 825,
205
+ 483
206
+ ],
207
+ "page_idx": 1
208
+ },
209
+ {
210
+ "type": "text",
211
+ "text": "As of today, there is no image-to-image translation mode that can outperform both the Pix2Pix mode on aligned data and the Cycle-consistency mode on misaligned or unpaired data. Inspired by[6–10], we consider the misaligned target images as noisy labels, which means that the existing problem is regarded as supervised learning with noisy labels. So we introduce a new image-to-image translation mode called RegGAN. Figure 1 provides a comparison of the three modes: Pix2Pix, Cycle-consistency and RegGAN. To facilitate reading, we summarize our contributions as follows. ",
212
+ "bbox": [
213
+ 174,
214
+ 488,
215
+ 825,
216
+ 571
217
+ ],
218
+ "page_idx": 1
219
+ },
220
+ {
221
+ "type": "text",
222
+ "text": "• We demonstrate the feasibility of RegGAN from the theoretical perspective of \"losscorrection\". Specifically, we train the generator using an additional registration network to fit the misaligned noise distribution adaptively, with the goal to search for the common optimal solution for both image-to-image translation and registration tasks. RegGAN eliminates the requirement for well aligned paired images and searches unique solution in training process. Based on our results, RegGAN outperformed both Pix2Pix on aligned data and Cycle-consistencyaon misaligned or unpaired data. ˘ RegGAN can be integrated into other methods without changing the original network architecture. Compared to Cycle-consistency with two generators and discriminators, RegGAN can provide better performance using less network parameters. ",
223
+ "bbox": [
224
+ 215,
225
+ 583,
226
+ 825,
227
+ 732
228
+ ],
229
+ "page_idx": 1
230
+ },
231
+ {
232
+ "type": "text",
233
+ "text": "2 Related Work ",
234
+ "text_level": 1,
235
+ "bbox": [
236
+ 174,
237
+ 755,
238
+ 321,
239
+ 772
240
+ ],
241
+ "page_idx": 1
242
+ },
243
+ {
244
+ "type": "text",
245
+ "text": "Image-to-image Translation: Generative adversarial networks (GANs) have shown great potential in the field of image-to-image translation[11–16]. It has been successfully implemented in medical image analysis like segmentation[17], registration[18, 19] and dose calculation[20]. The existing modes, however, have their limitations. Specifically, the Pix2Pix mode[2] requires well aligned paired images which may not always be available. The Cycle-consistency mode can achieve unsupervised image-to-image translation. With a Cycle-consistency loss, it can be used for misaligned images. Based on Cycle-consistency, many methods[3, 21–30] have been developed including CycleGAN[3] and its variants such as MUNIT[31] and UNIT[32] in which both image content and style information are used to decouple and reconstruct the image-to-image translation task; U-gat-it[33] with a self-attention mechanism added; and NICEGAN[34] proposed to reuse the discriminator for encoding. The main limitation of Cycle-consistency is that it may produce multiple solutions and therefore is sensitive to perturbation, making it difficult to meet the high accuracy requirements of medical image-to-image translation tasks. ",
246
+ "bbox": [
247
+ 174,
248
+ 786,
249
+ 826,
250
+ 911
251
+ ],
252
+ "page_idx": 1
253
+ },
254
+ {
255
+ "type": "image",
256
+ "img_path": "images/af1c6b6474d3dfc147a4911d8cae1c1814796ffc4a02561f528abce736db83bd.jpg",
257
+ "image_caption": [
258
+ "Figure 1: Comparison among the modes of Pix2Pix, CycleGAN and RegGAN. "
259
+ ],
260
+ "image_footnote": [],
261
+ "bbox": [
262
+ 179,
263
+ 93,
264
+ 816,
265
+ 327
266
+ ],
267
+ "page_idx": 2
268
+ },
269
+ {
270
+ "type": "text",
271
+ "text": "",
272
+ "bbox": [
273
+ 174,
274
+ 387,
275
+ 825,
276
+ 443
277
+ ],
278
+ "page_idx": 2
279
+ },
280
+ {
281
+ "type": "text",
282
+ "text": "Learning From Noisy Labels: Neural network anti-noise training has made great progress. Current research are mainly focused on: estimating the noise transition matrix[7, 35–40], designing a robust loss function[41–44], correcting the noise label[45–50], sampling importance weighting[51– 55] and meta-learning[56–59]. Our work is in the category of estimating the noise transition matrix. Compared to conventional noise transition estimation, we mitigate the issue and simplify the task by acquiring prior knowledge of noise distribution. ",
283
+ "bbox": [
284
+ 173,
285
+ 449,
286
+ 826,
287
+ 532
288
+ ],
289
+ "page_idx": 2
290
+ },
291
+ {
292
+ "type": "text",
293
+ "text": "Deformable Registration: Traditional image registration methods have gained widespread acceptance, such as Demons[60], B-spline[61] and elastic deformation model[62]. One of the most popular deep learning methods is Voxelmorph[63]. In this study, a CNN model was trained to predict the deformable vector filed (DVF). The time-consuming gradient derivation process was thus skipped to improve the calculation efficiency. Affine registration and a vector momentum-parameterized stationary velocity field (vSVF)[64] was implemented to get better transformation regulation. Fast Symmetric method[65] used symmetric maximum similarity. Deep flash[66] outperformed other models in terms of training and calculation time. ",
294
+ "bbox": [
295
+ 174,
296
+ 537,
297
+ 825,
298
+ 650
299
+ ],
300
+ "page_idx": 2
301
+ },
302
+ {
303
+ "type": "text",
304
+ "text": "Closest to our work, Arar.M et al[67] introduced a multi-modal registration method for natural images based on geometry preserving. But their work focused only on registration and did not demonstrate results of image-to-image translation or discuss the relationship between registration and image-to-image translation. The key insight of our work is that we demonstrated the feasibility of using registration to significantly improve the performance of image-to-image translation because the noise could be eliminated adaptively during the joint training process. What we propose in the paper is a completely new mode for medical image-to-image translation. ",
305
+ "bbox": [
306
+ 174,
307
+ 655,
308
+ 825,
309
+ 752
310
+ ],
311
+ "page_idx": 2
312
+ },
313
+ {
314
+ "type": "text",
315
+ "text": "3 Methodology ",
316
+ "text_level": 1,
317
+ "bbox": [
318
+ 174,
319
+ 776,
320
+ 313,
321
+ 792
322
+ ],
323
+ "page_idx": 2
324
+ },
325
+ {
326
+ "type": "text",
327
+ "text": "3.1 Theoretical Motivation ",
328
+ "text_level": 1,
329
+ "bbox": [
330
+ 176,
331
+ 809,
332
+ 372,
333
+ 825
334
+ ],
335
+ "page_idx": 2
336
+ },
337
+ {
338
+ "type": "text",
339
+ "text": "If we consideramisaligned target images as noisy labels, ˘ athe training for image-to-image translation ˘ becomes aasupervised learning process ˘ awith noisy labels. Given a training dataset ˘ $\\{ ( x _ { n } , \\widetilde { y } _ { n } ) \\} _ { n = 1 } ^ { N }$ with $N$ noisy labels in which $x _ { n }$ , $\\widetilde { y } _ { n }$ are images from two modalities and assume $y _ { n }$ is the correct label for $x _ { n }$ e, but it is unknown in real-world scenarios. Our goal is to train a generator using the dataset $\\{ ( x _ { n } , \\widetilde { y } _ { n } ) \\} _ { n = 1 } ^ { N }$ with noisy labels and achieve the performance equivalent to trained on clean dataset not wor $\\left\\{ \\left( x _ { n } , y _ { n } \\right) \\right\\} _ { n = 1 } ^ { N }$ as much as possible. Direct optimization based on Equations 4 usually does to bad results because the generator cannot squeeze out the influence of noise. ",
340
+ "bbox": [
341
+ 174,
342
+ 837,
343
+ 825,
344
+ 912
345
+ ],
346
+ "page_idx": 2
347
+ },
348
+ {
349
+ "type": "text",
350
+ "text": "",
351
+ "bbox": [
352
+ 173,
353
+ 88,
354
+ 826,
355
+ 121
356
+ ],
357
+ "page_idx": 3
358
+ },
359
+ {
360
+ "type": "equation",
361
+ "img_path": "images/0643f279dbd3eda0a5ea287241962ed5d5b869b67bae060f19991656d4d41e38.jpg",
362
+ "text": "$$\n\\hat { G } = \\underset { G } { \\arg \\operatorname* { m i n } } \\frac { 1 } { N } \\sum _ { n = 1 } ^ { N } \\mathcal { L } \\left( G \\left( x _ { n } \\right) , \\widetilde { y } _ { n } \\right)\n$$",
363
+ "text_format": "latex",
364
+ "bbox": [
365
+ 377,
366
+ 128,
367
+ 620,
368
+ 171
369
+ ],
370
+ "page_idx": 3
371
+ },
372
+ {
373
+ "type": "text",
374
+ "text": "To address the noise issue, we propose a solution based on \"loss-correction\"[7] shown in Equations 5. Our solution corrects the output of the generator $G ( x _ { n } )$ by modeling a noise transition $\\phi$ to match the noise distribution. Previously, Patrini et al[7] proved mathematically that the model trained with the noisy labels could be equivalent to the model trained with the clean labels, if the noise transition $\\phi$ matches the noise distribution. ",
375
+ "bbox": [
376
+ 173,
377
+ 172,
378
+ 825,
379
+ 242
380
+ ],
381
+ "page_idx": 3
382
+ },
383
+ {
384
+ "type": "equation",
385
+ "img_path": "images/ce568a9f17c471973cedcd570ba78b10facb5d8e8c4023f857a5c776e231e8fb.jpg",
386
+ "text": "$$\n\\hat { G } = \\underset { G } { \\arg \\operatorname* { m i n } } \\frac { 1 } { N } \\sum _ { n = 1 } ^ { N } \\mathcal { L } \\left( \\phi \\circ G \\left( x _ { n } \\right) , \\widetilde { y } _ { n } \\right)\n$$",
387
+ "text_format": "latex",
388
+ "bbox": [
389
+ 366,
390
+ 247,
391
+ 633,
392
+ 290
393
+ ],
394
+ "page_idx": 3
395
+ },
396
+ {
397
+ "type": "text",
398
+ "text": "To achieve this, Goldberger et al[36] proposed to view the correct label as a latent random variable and explicitly model the label noise as a part of the network architecture, denoted by $R$ . Then, Equations 5 can be rewritten in the form of log-likelihood, which is used as the loss function for neural network training. ",
399
+ "bbox": [
400
+ 176,
401
+ 300,
402
+ 823,
403
+ 357
404
+ ],
405
+ "page_idx": 3
406
+ },
407
+ {
408
+ "type": "equation",
409
+ "img_path": "images/eebb34dd03b6d35943b583d5de6201cca50c0e9548a7b6cc0e40b3d837a3627a.jpg",
410
+ "text": "$$\n\\begin{array} { l } { \\displaystyle \\mathcal { L } \\left( G , R \\right) = - \\sum _ { n = 1 } ^ { N } l o g \\left( p \\left( \\widetilde { y } _ { n } | y _ { n } ; R \\right) p \\left( y _ { n } | x _ { n } ; G \\right) \\right) } \\\\ { \\displaystyle = - \\sum _ { n = 1 } ^ { N } l o g \\left( p \\left( \\widetilde { y } _ { n } | x _ { n } ; G , R \\right) \\right) } \\end{array}\n$$",
411
+ "text_format": "latex",
412
+ "bbox": [
413
+ 331,
414
+ 362,
415
+ 665,
416
+ 450
417
+ ],
418
+ "page_idx": 3
419
+ },
420
+ {
421
+ "type": "text",
422
+ "text": "3.2 RegGAN ",
423
+ "text_level": 1,
424
+ "bbox": [
425
+ 173,
426
+ 462,
427
+ 276,
428
+ 478
429
+ ],
430
+ "page_idx": 3
431
+ },
432
+ {
433
+ "type": "text",
434
+ "text": "Compared to existing methods that use expectation-maximum[7, 36], fully connected layers[35], anchor point estimate[37] and Drichlet-distribution[38] to solve Equations 6. In our problem, the type of noise distribution is clearer, it can be expressed as displacement error: $\\widetilde { y } = y \\circ T$ . Here $T$ is expressed as a random deformation field, which produces random displacement for each pixel. So we adopt a registration network $R$ after the generator $G$ as label noise model to correct the results. The Correction loss is shown Equations 7: ",
435
+ "bbox": [
436
+ 173,
437
+ 487,
438
+ 826,
439
+ 571
440
+ ],
441
+ "page_idx": 3
442
+ },
443
+ {
444
+ "type": "equation",
445
+ "img_path": "images/d6224ac185e4206b98c8a5e1c1b6bcfdc6bc1bc85dcec898d92cabd2fb18b148.jpg",
446
+ "text": "$$\n\\operatorname* { m i n } _ { G , R } { \\mathcal { L } } _ { C o r r } \\left( G , R \\right) = \\mathbb { E } _ { x , \\widetilde { y } } \\left[ \\left. \\widetilde { y } - G \\left( x \\right) \\circ R \\left( G \\left( x \\right) , \\widetilde { y } \\right) \\right. _ { 1 } \\right]\n$$",
447
+ "text_format": "latex",
448
+ "bbox": [
449
+ 308,
450
+ 575,
451
+ 689,
452
+ 601
453
+ ],
454
+ "page_idx": 3
455
+ },
456
+ {
457
+ "type": "text",
458
+ "text": "where, $R \\left( G \\left( x \\right) , \\widetilde { y } \\right)$ is the deformation field and $\\circ$ represents the resamples operation. The registration network is based on U-Net[68]. A smoothness loss[63] is defined in Equations 8 to evaluate the smoothness of the deformation field and minimize the gradient of the deformation field. ",
459
+ "bbox": [
460
+ 173,
461
+ 606,
462
+ 826,
463
+ 648
464
+ ],
465
+ "page_idx": 3
466
+ },
467
+ {
468
+ "type": "equation",
469
+ "img_path": "images/83187509d08ed0c4e0021f309bea195d5f71177add3272c248bfcee3e8c47705.jpg",
470
+ "text": "$$\n\\operatorname* { m i n } _ { R } \\mathcal { L } _ { S m o o t h } \\left( R \\right) = \\mathbb { E } _ { x , \\widetilde { y } } \\left[ \\| \\nabla R \\left( G \\left( x \\right) , \\widetilde { y } \\right) \\| ^ { 2 } \\right]\n$$",
471
+ "text_format": "latex",
472
+ "bbox": [
473
+ 341,
474
+ 654,
475
+ 653,
476
+ 678
477
+ ],
478
+ "page_idx": 3
479
+ },
480
+ {
481
+ "type": "text",
482
+ "text": "Finally, we add the Aversarial loss between the generator and the discriminator (Equations 1), and the total loss is expressed in Equations 9. ",
483
+ "bbox": [
484
+ 171,
485
+ 681,
486
+ 823,
487
+ 710
488
+ ],
489
+ "page_idx": 3
490
+ },
491
+ {
492
+ "type": "equation",
493
+ "img_path": "images/04c231c002f5af59868fcd8a3ef1bf7135dc8c40a88688f72f0467da4bf32a3f.jpg",
494
+ "text": "$$\n\\operatorname* { m i n } _ { G , R } \\operatorname* { m a x } _ { D } \\mathcal { L } _ { T o t a l } \\left( G , R , D \\right) = \\mathcal { L } _ { C o r r } + \\mathcal { L } _ { S m o o t h } + \\mathcal { L } _ { A d v }\n$$",
495
+ "text_format": "latex",
496
+ "bbox": [
497
+ 310,
498
+ 714,
499
+ 687,
500
+ 739
501
+ ],
502
+ "page_idx": 3
503
+ },
504
+ {
505
+ "type": "text",
506
+ "text": "4 Experiments ",
507
+ "text_level": 1,
508
+ "bbox": [
509
+ 174,
510
+ 755,
511
+ 312,
512
+ 772
513
+ ],
514
+ "page_idx": 3
515
+ },
516
+ {
517
+ "type": "text",
518
+ "text": "Performance evaluation of RegGAN was conducted through three investigations to 1) demonstrate the feasibility and superiority of the RegGAN mode in various methods, and 2) assess RegGANs sensitivity to noise, and 3) explore the availability of the RegGAN on unpaired data. ",
519
+ "bbox": [
520
+ 173,
521
+ 785,
522
+ 825,
523
+ 829
524
+ ],
525
+ "page_idx": 3
526
+ },
527
+ {
528
+ "type": "text",
529
+ "text": "4.1 Dataset ",
530
+ "text_level": 1,
531
+ "bbox": [
532
+ 173,
533
+ 843,
534
+ 266,
535
+ 858
536
+ ],
537
+ "page_idx": 3
538
+ },
539
+ {
540
+ "type": "text",
541
+ "text": "The open-access dataset (BraTS 2018[69]) was used to evaluate the proposed RegGAN mode. The training dataset and testing dataset contained 8457 and 979 pairs of T1 and T2 MR images, respectively. BraTS 2018 was selected because the original images were paired and well aligned. We created misaligned images by randomly adding different levels of rotation, translation and rescaling to the original images. And we randomly sample one image from T1 and the other one from T2 when training on unpaired images. The availability of well aligned paired images, misaligned paired images, and unpaired images allow us to evaluate the performances of all three modes (Pix2Pix, Cycle-consistency and RegGAN). ",
542
+ "bbox": [
543
+ 174,
544
+ 868,
545
+ 825,
546
+ 911
547
+ ],
548
+ "page_idx": 3
549
+ },
550
+ {
551
+ "type": "text",
552
+ "text": "",
553
+ "bbox": [
554
+ 174,
555
+ 90,
556
+ 825,
557
+ 161
558
+ ],
559
+ "page_idx": 4
560
+ },
561
+ {
562
+ "type": "text",
563
+ "text": "4.2 Performances in Different Methods ",
564
+ "text_level": 1,
565
+ "bbox": [
566
+ 176,
567
+ 176,
568
+ 459,
569
+ 193
570
+ ],
571
+ "page_idx": 4
572
+ },
573
+ {
574
+ "type": "text",
575
+ "text": "The primary motivation of introducing RegGAN was to address challenges caused by misaligned data. Therefore, in this section, misaligned data were used in model training to demonstrate the feasibility and superiority of RegGAN. We selected the most popular CycleGAN[3] and its variants MUNIT[31], UNIT[32], and NICEGAN[34] as the methods for evaluation and compared the following four modes for each method. ",
576
+ "bbox": [
577
+ 174,
578
+ 202,
579
+ 826,
580
+ 272
581
+ ],
582
+ "page_idx": 4
583
+ },
584
+ {
585
+ "type": "text",
586
+ "text": "• C(Cycle-consistency): The most primitive mode of all methods, with Cycle-consistency loss (Equations 3) as the main constraint. Two generators and two discriminators are required in this mode. $\\mathbf { C } { \\ + } \\mathbf { R }$ (Cycle-consistency $^ +$ Registration): The RegGAN mode is combined with the mode C. Registration network $( R )$ and Correction loss (Equations 7) are added to the constraints. NC(Non Cycle-consistency): Only Adversarial loss (Equations 1) is used for updating. Compared to the mode C, Cycle-consistency loss is removed. Only one generator and one discriminator are required in this mode. \n• $\\mathbf { N C } { + } \\mathbf { R }$ (Non Cycle-consistency $^ +$ Registration): A registration network $( R )$ and Correction loss (Equations 7) are added to the mode NC. It is the proposed RegGAN mode. ",
587
+ "bbox": [
588
+ 215,
589
+ 285,
590
+ 825,
591
+ 439
592
+ ],
593
+ "page_idx": 4
594
+ },
595
+ {
596
+ "type": "table",
597
+ "img_path": "images/12d3e7a40a78c4217140ad57003c9f3887681b085f2676b02b2cf200340f9bf1.jpg",
598
+ "table_caption": [
599
+ "Table 1: Comparison of CycleGAN, MUNIT, UNIT and NICEGAN using four training modes(C, $\\mathrm { C } { \\ + } \\mathrm { R }$ , NC and $\\mathrm { N C } { + } \\mathrm { R }$ ). "
600
+ ],
601
+ "table_footnote": [],
602
+ "table_body": "<table><tr><td>Modes Methods</td><td rowspan=\"3\"></td><td rowspan=\"3\">CycleGAN</td><td rowspan=\"3\">MUNIT</td><td rowspan=\"3\">UNIT</td><td rowspan=\"3\">NICEGAN</td></tr><tr><td>Index</td><td></td></tr><tr><td>C</td><td>0.089 0.11</td></tr><tr><td rowspan=\"2\">NMAE↓</td><td>C+R NC</td><td>(-0.012)0.077 0.11</td><td>(-0.022)0.088 0.10</td><td>0.087 (-0.013)0.074 0.098</td><td>0.082 (-0.011)0.071 0.089</td></tr><tr><td>NC+R C</td><td>(-0.038)0.072 23.5</td><td>(-0.021)0.079 20.6</td><td>(-0.027)0.071 24.6</td><td>(-0.019)0.070 25.2</td></tr><tr><td>PSNR ↑</td><td>C+R NC NC+R</td><td>(+0.3)23.8 20.2 (+5.4)25.6</td><td>(+2.1)22.7 21.5 (+2.3)23.8</td><td>(+0.7)25.3 23.7 (+1.8)25.5</td><td>(+0.9)26.1 23.5 (+2.8)26.3</td></tr><tr><td>SSIM↑</td><td>C C+R NC NC+R</td><td>0.83 (+0.02)0.85 0.79 (+0.07)0.86</td><td>0.80 (+0.03) 0.83 0.81 (+0.04)0.85</td><td>0.84 (+0.02)0.86 0.83 (+0.03) 0.86</td><td>0.83 (+0.03)0.86 0.84 (+0.02)0.86</td></tr></table>",
603
+ "bbox": [
604
+ 173,
605
+ 505,
606
+ 839,
607
+ 770
608
+ ],
609
+ "page_idx": 4
610
+ },
611
+ {
612
+ "type": "text",
613
+ "text": "To evaluate the performance of each method on misaligned data, we randomly added $[ - 5 , + 5 ]$ degrees of angle rotation, $[ - 5 , + 5 ]$ percent of translation, and $[ - 5 , + 5 ]$ percent of rescaling to the original T1 and T2 images on the training dataset. ",
614
+ "bbox": [
615
+ 174,
616
+ 779,
617
+ 825,
618
+ 821
619
+ ],
620
+ "page_idx": 4
621
+ },
622
+ {
623
+ "type": "text",
624
+ "text": "To ensure fair comparison, we used the same training strategy and hyperparameters for all methods and modes (see supplementary materials for details). The Normalized Mean Absolute Error (NMAE), Peak Signal to Noise Ratio (PSNR) and Structural Similarity (SSIM) were used as metrics to evaluate the performances of trained models based on the testing dataset. To avoid false high results of index, we excluded the image background from the calculation. Table 1 summarized the results for all methods and modes under the current investigation. ",
625
+ "bbox": [
626
+ 174,
627
+ 828,
628
+ 825,
629
+ 911
630
+ ],
631
+ "page_idx": 4
632
+ },
633
+ {
634
+ "type": "image",
635
+ "img_path": "images/eedbfe72566d6ed42ea2f3ea9e4dccafb5a309ef45d96acc8b055ab8e2b1bf94.jpg",
636
+ "image_caption": [
637
+ "Figure 2: The errors of different modes in different methods. "
638
+ ],
639
+ "image_footnote": [],
640
+ "bbox": [
641
+ 173,
642
+ 92,
643
+ 813,
644
+ 520
645
+ ],
646
+ "page_idx": 5
647
+ },
648
+ {
649
+ "type": "text",
650
+ "text": "Based on the results from the Table 1, we can reach several conclusions. First, adding the registration network $\\mathbf { \\tau } ( + \\mathbf { R } )$ significantly improves the performances of the methods. This is true for all methods in both C and NC modes. It clearly demonstrates that $\\operatorname { R e g G A N }$ can be incorporated in various methods or combined with different network architectures to improve the performances. Second, the C mode is in general better than the NC mode for most of methods. Adding the registration network $\\mathbf { \\tau } ( + \\mathbf { R } )$ improves the performance of the NC mode more than that of the C mode. In fact, our results show that the $\\mathbf { N C } { + } \\mathbf { R }$ mode is even better than the $\\mathbf { C } { \\ + } \\mathbf { R }$ mode, implying that \"Cycle-consistency loss\" may play a negative role when it is combined with RegGAN. Compared with the commonly used C mode with two generators and two discriminators, RegGAN has fewer parameters but provides better performance. The simple CycleGAN method in the $\\mathbf { N C } { + } \\mathbf { R }$ mode outperforms the current state-of-the-art method NICEGAN in the C mode by 0.01, 0.4, 0.03 for NMAE, PSNR and SSIM, respectively. The $\\mathbf { N C } { + } \\mathbf { R }$ mode can also be used to improve the performance of NICEGAN. In fact, the performance of NICEGAN in the $\\mathbf { N C } { + } \\mathbf { R }$ mode is the best among all combinations of the 4 methods and 4 modes. ",
651
+ "bbox": [
652
+ 174,
653
+ 641,
654
+ 826,
655
+ 835
656
+ ],
657
+ "page_idx": 5
658
+ },
659
+ {
660
+ "type": "text",
661
+ "text": "Figure 2 shows representative results from various combinations of the 4 methods (CycleGAN, MUNIT, UNIT and NICEGAN) and 4 modes (C, $\\mathbf { C } { \\ + } \\mathbf { R }$ , NC and $\\mathbf { N C + R }$ ). For all aspects of the image (from the tumor areas and the details), the combinations that use the registration network $\\mathbf { \\tau } ( + \\mathbf { R } )$ always provide more realistic and accurate results than those that do not use the registration network $\\mathbf { \\tau } ( + \\mathbf { R } )$ . ",
662
+ "bbox": [
663
+ 174,
664
+ 842,
665
+ 825,
666
+ 911
667
+ ],
668
+ "page_idx": 5
669
+ },
670
+ {
671
+ "type": "image",
672
+ "img_path": "images/1d3e8120380b91f9184bcfce0c98cc767b03f1f62b9eb2e29c1da17e47ccd98c.jpg",
673
+ "image_caption": [
674
+ "Figure 3: Example images at seven different levels of introduced noise. "
675
+ ],
676
+ "image_footnote": [],
677
+ "bbox": [
678
+ 179,
679
+ 93,
680
+ 818,
681
+ 143
682
+ ],
683
+ "page_idx": 6
684
+ },
685
+ {
686
+ "type": "table",
687
+ "img_path": "images/e4c536347294c88b33ab00346675067bab9876661946a2008ae547c9ea839dad.jpg",
688
+ "table_caption": [
689
+ "Table 2: Comparison of the NMAE, PSNR and SSIM for CycleGAN(C), Pix2Pix and RegGAN under 7 levels of noise. "
690
+ ],
691
+ "table_footnote": [],
692
+ "table_body": "<table><tr><td colspan=\"2\"></td><td>Noise.0</td><td>Noise.1</td><td>Noise.2</td><td>Noise.3</td><td>Noise.4</td><td></td><td>Noise.5 Noise.NA</td></tr><tr><td rowspan=\"3\">Setting</td><td>Rotate</td><td>0°</td><td>±1°</td><td>±2°</td><td>±3°</td><td>±4°</td><td>±5°</td><td>X</td></tr><tr><td>Translation</td><td>0%</td><td>±2%</td><td>±4%</td><td>±6%</td><td>±8%</td><td>±10%</td><td>X</td></tr><tr><td>Rescaling</td><td>0%</td><td>±2%</td><td>±4%</td><td>±6%</td><td>±8%</td><td>±10%</td><td>X</td></tr><tr><td rowspan=\"3\">CycleGAN(C)</td><td>NMAE↓</td><td>0.084</td><td>0.095</td><td>0.087</td><td>0.094</td><td>0.087</td><td>0.110</td><td>0.091</td></tr><tr><td>PSNR ↑</td><td>23.9</td><td>22.5</td><td>23.7</td><td>23.3</td><td>23.9</td><td>23.7</td><td>23.5</td></tr><tr><td>SSIM↑</td><td>0.83</td><td>0.83</td><td>0.82</td><td>0.81</td><td>0.82</td><td>0.79</td><td>0.83</td></tr><tr><td rowspan=\"3\">Pix2Pix</td><td>NMAE ↓</td><td>0.075</td><td>0.103</td><td>0.139</td><td>0.161</td><td>0.175</td><td>0.181</td><td>0.086</td></tr><tr><td>PSNR ↑</td><td>25.6</td><td>22.3</td><td>18.9</td><td>16.2</td><td>15.3</td><td>15.0</td><td></td></tr><tr><td>SSIM ↑</td><td>0.85</td><td>0.82</td><td>0.78</td><td>0.76</td><td>0.74</td><td>0.74</td><td>21.1</td></tr><tr><td rowspan=\"3\">RegGAN</td><td>NMAE ↓</td><td>0.071</td><td>0.073</td><td>0.071</td><td>0.072</td><td>0.072</td><td>0.072</td><td>0.82 0.071</td></tr><tr><td>PSNR ↑</td><td>26</td><td>25.6</td><td>25.9</td><td>25.7</td><td>25.4</td><td>25.2</td><td></td></tr><tr><td>SSIM ↑</td><td>0.86</td><td>0.86</td><td>0.86</td><td>0.86</td><td>0.86</td><td>0.85</td><td>25.9 0.86</td></tr></table>",
693
+ "bbox": [
694
+ 173,
695
+ 237,
696
+ 831,
697
+ 518
698
+ ],
699
+ "page_idx": 6
700
+ },
701
+ {
702
+ "type": "text",
703
+ "text": "4.3 Performances in Different Noise Levels ",
704
+ "text_level": 1,
705
+ "bbox": [
706
+ 174,
707
+ 540,
708
+ 485,
709
+ 555
710
+ ],
711
+ "page_idx": 6
712
+ },
713
+ {
714
+ "type": "text",
715
+ "text": "To evaluate the sensitivity of $\\operatorname { R e g G A N }$ to noise, we selected a simple network architecture.(CycleGAN) with the intention to minimize interference from other factors. The same network architecture was used for all three modes: CycleGAN(C), Pix2Pix and RegGAN. Seven levels of noise were used in the evaluation. Table 2 lists the specific noise setting and range for each noise level. Noise.0 means the original dataset with no added noise. Noise.5 is the highest level of noise. At Noise.5, the data are likely from different patients. In addition, we also made non-affine noise(Noise.NA) settings. The non-affine noise is generated by spatially transforming T1 and T2 using elastic transformations on control points followed by Gaussian smoothing. Figure 3 shows example images at different levels of introduced noise. ",
716
+ "bbox": [
717
+ 173,
718
+ 565,
719
+ 826,
720
+ 690
721
+ ],
722
+ "page_idx": 6
723
+ },
724
+ {
725
+ "type": "text",
726
+ "text": "Table 2 lists the quantitative evaluation metrics from 3 modes at 7 levels of noise. It is clear that RegGAN outperforms CycleGAN(C) under all noise levels. Figure 4(a) shows the test results from each epoch during the training process for both RegGAN and CycleGAN(C). Curves of different colors corresponds to different levels of noise. We notice that CycleGAN(C) is not very stable during the training process. The test results fluctuate significantly and cannot converge well. This may be caused by the fact that the solution of CycleGAN (C) is not unique. As a comparison, RegGAN is quite stable. Although the results from different levels of noise may vary at the beginning of training, all curves converge to a similar result after multiple epoches of training, indicating that RegGAN is more robust to noise compared to CycleGAN (C). ",
727
+ "bbox": [
728
+ 173,
729
+ 696,
730
+ 825,
731
+ 821
732
+ ],
733
+ "page_idx": 6
734
+ },
735
+ {
736
+ "type": "text",
737
+ "text": "Based on Table 2, we notice that the performance of Pix2Pix deteriorates rapidly as the noise increases. This is as expected because Pix2Pix requires well aligned paired images. Surprisingly, the performances of RegGAN at all noise levels exceed those of Pix2Pix with no noise. Figure 4(b) shows the test results at each epoch of RegGAN and Pix2Pix under Noise.0 (i.e., no noise). Theoretically, the performances of RegGAN and Pix2Pix should be similar on perfectly aligned paired datasets because the registration network of RegGAN does not help and RegGAN is equivalent to ",
738
+ "bbox": [
739
+ 174,
740
+ 828,
741
+ 825,
742
+ 911
743
+ ],
744
+ "page_idx": 6
745
+ },
746
+ {
747
+ "type": "image",
748
+ "img_path": "images/f662bd1bbeb7045dac456003b31281c1beb1a993af4d26cdff8fba4ce3a7d075.jpg",
749
+ "image_caption": [
750
+ "Figure 4: Quantitative evaluation metrics at different epochs in the training process. (a) Comparison of CycleGAN and $\\operatorname { R e g G A N }$ at different levels of noise. (b) Comparison of $\\mathrm { P i x 2 P i x }$ and $\\mathsf { R e g G A N }$ at Noise.0 (i.e., no noise). (c) RegGAN’s Smoothness loss under different levels of noise. "
751
+ ],
752
+ "image_footnote": [],
753
+ "bbox": [
754
+ 178,
755
+ 90,
756
+ 818,
757
+ 236
758
+ ],
759
+ "page_idx": 7
760
+ },
761
+ {
762
+ "type": "image",
763
+ "img_path": "images/326a09b807acfdbf4500bf0e29287859625c86c57910fc62e0d5bca1bce29663.jpg",
764
+ "image_caption": [
765
+ "Figure 5: The misalignment of orginal image pairs and corresponding deformation fields. "
766
+ ],
767
+ "image_footnote": [],
768
+ "bbox": [
769
+ 176,
770
+ 333,
771
+ 820,
772
+ 587
773
+ ],
774
+ "page_idx": 7
775
+ },
776
+ {
777
+ "type": "text",
778
+ "text": "Pix2Pix. A possible explanation to our results is that in the medical field, the perfectly pixel-wise aligned dataset may not practically exist. Even for BraTS 2018[63] which is recognized as well aligned, it is still possible that there exists slight misalignment. As a result, adding the registration network is always likely to improve the performances in real-world scenarios. To verify our explanation, we plotted the Smoothness loss of RegGAN under different noise levels as shown in Figure 4(c). Large Smoothness loss corresponds to large deformation field displacement. First, we notice that the Smoothness loss under Noise.0 never completely goes to 0, indicating the existence of misalignment and potential usefulness of the registration network. Second, the noise level and Smoothness loss show a step-like positive correlation, which means that RegGAN can adaptively handle the noise distribution, i.e., the registration network can determine the range of deformation according to the noise level. In addition, we can see that even under the setting of non-affine noise, the above conclusion still holds. Because what the registration network corrects is deformation noise. ",
779
+ "bbox": [
780
+ 173,
781
+ 670,
782
+ 825,
783
+ 849
784
+ ],
785
+ "page_idx": 7
786
+ },
787
+ {
788
+ "type": "text",
789
+ "text": "we also show some original image pairs and visualize the corresponding deformation fields output by registration network in Figure 5. Obviously, there is some misalignment between the original T1 and T2 images, and such misalignment is represented by the deformation fields (highlighted by red circle). ",
790
+ "bbox": [
791
+ 174,
792
+ 856,
793
+ 823,
794
+ 911
795
+ ],
796
+ "page_idx": 7
797
+ },
798
+ {
799
+ "type": "image",
800
+ "img_path": "images/27d4890be00f9554f08f14f90822eb89bc97096dee32bda414c7eac84631e0c8.jpg",
801
+ "image_caption": [
802
+ "Figure 6: Performance comparison of the three modes (CycleGAN(C), Pix2Pix and RegGAN) on unpaired dataset. "
803
+ ],
804
+ "image_footnote": [],
805
+ "bbox": [
806
+ 176,
807
+ 92,
808
+ 834,
809
+ 250
810
+ ],
811
+ "page_idx": 8
812
+ },
813
+ {
814
+ "type": "image",
815
+ "img_path": "images/048a4fd457b36a04d9cc483fe88373cc9c0bedef0e626f7bb79deca8ef1b4a45.jpg",
816
+ "image_caption": [
817
+ "Figure 7: Display of RegGAN’s output on unpaired data. T1 and T2 are unpaired images. The Translated represents the translation result of T1 to T2. Registered represents the registration result of the translated images. D.F represents deformation fields. "
818
+ ],
819
+ "image_footnote": [],
820
+ "bbox": [
821
+ 174,
822
+ 313,
823
+ 823,
824
+ 467
825
+ ],
826
+ "page_idx": 8
827
+ },
828
+ {
829
+ "type": "text",
830
+ "text": "4.4 Performances on Unpaired Dataset ",
831
+ "text_level": 1,
832
+ "bbox": [
833
+ 176,
834
+ 556,
835
+ 457,
836
+ 571
837
+ ],
838
+ "page_idx": 8
839
+ },
840
+ {
841
+ "type": "text",
842
+ "text": "So far, our investigations are based on paired datasets. We also want to explore how RegGAN performs using unpaired datasets. In practice, this is not recommended because even different patients may have similarities in their body tissues of adjacent layers. For unpaired datasets, we can conduct rigid registration first in 3D space and then use RegGAN for training. Unpaired data can be treated as having larger scale noise. If the correction capability is strong enough, RegGAN can still work effectively. The comparison of the performances of three modes on the unpaired dataset is shown in Figure 6. ",
843
+ "bbox": [
844
+ 174,
845
+ 587,
846
+ 825,
847
+ 684
848
+ ],
849
+ "page_idx": 8
850
+ },
851
+ {
852
+ "type": "text",
853
+ "text": "With unpaired datasets, Pix2Pix no longer considers the characteristics of the input T1 images and thus has the worst performance. Due to the challenges in fitting the noise, the performance improvement from replacing CycleGAN(C) with RegGAN using unpaired datasets may not be as dramatic as that demonstrated using paired datasets, but RegGAN still has the best performance under unpaired conditions. In Figure 7, we show some examples of how RegGAN corrects noise on unpaired dataset. It can be seen that RegGAN will try its best to eliminate the influnce of noise through registration. ",
854
+ "bbox": [
855
+ 174,
856
+ 690,
857
+ 825,
858
+ 787
859
+ ],
860
+ "page_idx": 8
861
+ },
862
+ {
863
+ "type": "text",
864
+ "text": "Based on our results, it is reasonable to reach the conclusions below. In all circumstances, RegGAN demonstrates better performance compared to Pix2Pix and CycleGAN(C). ",
865
+ "bbox": [
866
+ 174,
867
+ 794,
868
+ 823,
869
+ 821
870
+ ],
871
+ "page_idx": 8
872
+ },
873
+ {
874
+ "type": "text",
875
+ "text": "• For paired and aligned conditions, RegGA $\\mathsf { N } \\geq \\mathsf { P i x 2 P i x } > \\mathsf { C y c l e }$ GAN(C). ",
876
+ "bbox": [
877
+ 215,
878
+ 838,
879
+ 715,
880
+ 853
881
+ ],
882
+ "page_idx": 8
883
+ },
884
+ {
885
+ "type": "text",
886
+ "text": "• For paired but misaligned conditions, RegGAN $>$ CycleGAN(C) $>$ Pix2Pix. ",
887
+ "bbox": [
888
+ 214,
889
+ 867,
890
+ 732,
891
+ 883
892
+ ],
893
+ "page_idx": 8
894
+ },
895
+ {
896
+ "type": "text",
897
+ "text": "• For unpaired conditions, RegGAN $>$ CycleGAN(C) $>$ Pix2Pix. ",
898
+ "bbox": [
899
+ 215,
900
+ 896,
901
+ 650,
902
+ 911
903
+ ],
904
+ "page_idx": 8
905
+ },
906
+ {
907
+ "type": "text",
908
+ "text": "Conclusion ",
909
+ "text_level": 1,
910
+ "bbox": [
911
+ 174,
912
+ 89,
913
+ 269,
914
+ 106
915
+ ],
916
+ "page_idx": 9
917
+ },
918
+ {
919
+ "type": "text",
920
+ "text": "In this study, we introduced a new image-to-image translation mode RegGAN to the medical community that can break the dilemma of image-to-image translation task. Using a public BraTS 2018 dataset, we demonstrated the feasibility of RegGAN and its superior performance compared to Pix2Pix and Cycle-consistency. We validated that RegGAN can be incorporated into various existing methods to improve their performances. We also evaluated the sensitivity of RegGAN to noise. Our results confirmed that RegGAN could adapt well to various scenarios from no noise to large-scale noise. The superior performance of RegGAN makes it a better choice over Pix2Pix and Cycle-consistency whether datasets are aligned or not. However, this mode may not work well on natural images. The noise may cannot be considered simply as deformation errors due to the differences in natural images are much greater than those in medical images. ",
921
+ "bbox": [
922
+ 174,
923
+ 121,
924
+ 825,
925
+ 258
926
+ ],
927
+ "page_idx": 9
928
+ },
929
+ {
930
+ "type": "text",
931
+ "text": "Broader Impact ",
932
+ "text_level": 1,
933
+ "bbox": [
934
+ 174,
935
+ 277,
936
+ 310,
937
+ 295
938
+ ],
939
+ "page_idx": 9
940
+ },
941
+ {
942
+ "type": "text",
943
+ "text": "Image-to-image translation has been one of the main focuses in medical image analysis, as it aids in diagnosis and treatment. Previously, physicians had to use different medical imaging equipments if they wanted to get different image sequences of a patient, which was time-consuming and expensive. Pix2Pix mode is expected to solve this problem by its outstanding performance in image-to-image translation. In most of clinical scenarios, however, it is not practical to create such a large well aligned dataset for Pix2Pix mode. Cycle-consistence mode does not need well aligned dataset but can not meet the high-precision requirements of medical image analysis. Our work aims to provide a general image-to-image translation mode, which not only has no strict requirements on the dataset, but also can meet the clinical requirements in terms of image quality. In the future, we will attempt to obtain multi-modal dataset(eg MR-CT) for clinical verification. We foresee positive impacts if the mode is applied to diagnosis in radiology, treatment planning and research. ",
944
+ "bbox": [
945
+ 174,
946
+ 309,
947
+ 825,
948
+ 462
949
+ ],
950
+ "page_idx": 9
951
+ },
952
+ {
953
+ "type": "text",
954
+ "text": "References \n[1] Goodfellow I, Pouget-Abadie J, Mirza M, et al. Generative adversarial nets[C]//Ghahramani Z, Welling M, Cortes C, et al. Advances in Neural Information Processing Systems. Curran Associates, Inc., 2014. \n[2] Isola P, Zhu J Y, Zhou T, et al. Image-to-image translation with conditional adversarial networks[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 1125-1134. \n[3] Zhu J Y, Park T, Isola P, et al. Unpaired image-to-image translation using cycle-consistent adversarial networks[C]//Proceedings of the IEEE international conference on computer vision. 2017: 2223-2232. \n[4] Sim B, Oh G, Lim S, et al. Optimal transport, cyclegan, and penalized ls for unsupervised learning in inverse problems[J]. 2019. \n[5] Moriakov N, Adler J, Teuwen J. Kernel of cyclegan as a principal homogeneous space[C]// International Conference on Learning Representations. 2020. \n[6] Huang L, Zhang C, Zhang H. Self-adaptive training: beyond empirical risk minimization[J]. Advances in Neural Information Processing Systems, 2020, 33. \n[7] Patrini G, Rozza A, Krishna Menon A, et al. Making deep neural networks robust to label noise: A loss correction approach[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2017: 1944-1952. \n[8] Dgani Y, Greenspan H, Goldberger J. Training a neural network based on unreliable human annotation of medical images[C]//2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018). IEEE, 2018: 39-42. \n[9] Yao Y, Liu T, Han B, et al. Dual t: Reducing estimation error for transition matrix in labelnoise learning[C/OL]//Larochelle H, Ranzato M, Hadsell R, et al. Advances in Neural Information Processing Systems: volume 33. Curran Associates, Inc., 2020: 7260-7271. https: //proceedings.neurips.cc/paper/2020/file/512c5cad6c37edb98ae91c8a76c3a291-Paper.pdf. \n[10] Zhang L, Tanno R, Xu M C, et al. Disentangling human error from ground truth in segmentation of medical images[C/OL]//Larochelle H, Ranzato M, Hadsell R, et al. Advances in Neural Information Processing Systems: volume 33. Curran Associates, Inc., 2020: 15750-15762. https://proceedings.neurips.cc/paper/2020/file/ b5d17ed2b502da15aa727af0d51508d6-Paper.pdf. \n[11] wang y, Yu L, van de Weijer J. Deepi2i: Enabling deep hierarchical image-toimage translation by transferring from gans[C/OL]//Larochelle H, Ranzato M, Hadsell R, et al. Advances in Neural Information Processing Systems: volume 33. Curran Associates, Inc., 2020: 11803-11815. https://proceedings.neurips.cc/paper/2020/file/ 88855547570f7ff053fff7c54e5148cc-Paper.pdf. \n[12] Zhang P, Zhang B, Chen D, et al. Cross-domain correspondence learning for exemplar-based image translation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020: 5143-5153. \n[13] Kim T, Cha M, Kim H, et al. Learning to discover cross-domain relations with generative adversarial networks[C]//International Conference on Machine Learning. PMLR, 2017: 1857- 1865. \n[14] Cao Y, Wan X. Divgan: Towards diverse paraphrase generation via diversified generative adversarial network[C]//Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings. 2020: 2411-2421. \n[15] Choi Y, Choi M, Kim M, et al. Stargan: Unified generative adversarial networks for multidomain image-to-image translation[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 8789-8797. \n[16] Choi Y, Uh Y, Yoo J, et al. Stargan v2: Diverse image synthesis for multiple domains[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020: 8188-8197. \n[17] Chen C, Dou Q, Chen H, et al. Synergistic image and feature adaptation: Towards crossmodality domain adaptation for medical image segmentation[C]//Proceedings of the AAAI Conference on Artificial Intelligence: volume 33. 2019: 865-872. \n[18] Fan J, Cao X, Wang Q, et al. Adversarial learning for mono-or multi-modal registration[J]. Medical image analysis, 2019, 58:101545. \n[19] Qin C, Shi B, Liao R, et al. Unsupervised deformable registration for multi-modal images via disentangled representations[C]//International Conference on Information Processing in Medical Imaging. Springer, 2019: 249-261. \n[20] Sahiner B, Pezeshk A, Hadjiiski L M, et al. Deep learning in medical imaging and radiation therapy[J]. Medical physics, 2019, 46(1):e1-e36. \n[21] Hoffman J, Tzeng E, Park T, et al. Cycada: Cycle-consistent adversarial domain adaptation [C]//International conference on machine learning. PMLR, 2018: 1989-1998. \n[22] Li M, Huang H, Ma L, et al. Unsupervised image-to-image translation with stacked cycleconsistent adversarial networks[C]//Proceedings of the European conference on computer vision (ECCV). 2018: 184-199. \n[23] Deng W, Zheng L, Ye Q, et al. Image-image domain adaptation with preserved self-similarity and domain-dissimilarity for person re-identification[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 994-1003. \n[24] Li J, Chen E, Ding Z, et al. Cycle-consistent conditional adversarial transfer networks[C]// Proceedings of the 27th ACM International Conference on Multimedia. 2019: 747-755. \n[25] Schmidt V, Luccioni A, Mukkavilli S K, et al. Visualizing the consequences of climate change using cycle-consistent adversarial networks[J]. arXiv preprint arXiv:1905.03709, 2019. \n[26] Chen Z, Li J, Luo Y, et al. Canzsl: Cycle-consistent adversarial networks for zero-shot learning from natural language[C]//Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2020: 874-883. \n[27] Ren C X, Ziemann A, Theiler J, et al. Cycle-consistent adversarial networks for realistic pervasive change generation in remote sensing imagery[C]//2020 IEEE Southwest Symposium on Image Analysis and Interpretation (SSIAI). IEEE, 2020: 42-45. \n[28] Liu J, Ding Y, Xiong J, et al. Multi-cycle-consistent adversarial networks for ct image denoising[C]//2020 IEEE 17th International Symposium on Biomedical Imaging (ISBI). IEEE, 2020: 614-618. \n[29] Zheng C, Pan L, Wu P. Camu: Cycle-consistent adversarial mapping model for user alignment across social networks[J]. IEEE Transactions on Cybernetics, 2021. \n[30] Shah M, Chen X, Rohrbach M, et al. Cycle-consistency for robust visual question answering [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019: 6649-6658. \n[31] Huang X, Liu M Y, Belongie S, et al. Multimodal unsupervised image-to-image translation [C]//Proceedings of the European conference on computer vision (ECCV). 2018: 172-189. \n[32] Liu M Y, Breuel T, Kautz J. Unsupervised image-to-image translation networks[C/OL]// Guyon I, Luxburg U V, Bengio S, et al. Advances in Neural Information Processing Systems: volume 30. Curran Associates, Inc., 2017. https://proceedings.neurips.cc/paper/2017/ file/dc6a6489640ca02b0d42dabeb8e46bb7-Paper.pdf. \n[33] Kim J, Kim M, Kang H, et al. U-gat-it: Unsupervised generative attentional networks with adaptive layer-instance normalization for image-to-image translation[C/OL]// International Conference on Learning Representations. 2020. https://openreview.net/forum? id=BJlZ5ySKPH. \n[34] Chen R, Huang W, Huang B, et al. Reusing discriminators for encoding: Towards unsupervised image-to-image translation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020: 8168-8177. \n[35] S. Sukhbaatar M P L B R F, J. Bruna. Training convolutional networks with noisy labels[C]// International Conference on Learning Representations. 2015. \n[36] Goldberger J, Ben-Reuven E. Training deep neural-networks using a noise adaptation layer[J]. 2016. \n[37] Xia X, Liu T, Wang N, et al. Are anchor points really indispensable in labelnoise learning?[C/OL]//NeurIPS. 2019: 6835-6846. http://papers.nips.cc/paper/ 8908-are-anchor-points-really-indispensable-in-label-noise-learning. \n[38] Yao J, Wu H, Zhang Y, et al. Safeguarded dynamic label regression for noisy supervision[C]// Proceedings of the AAAI Conference on Artificial Intelligence: volume 33. 2019: 9103-9110. \n[39] Xiao T, Xia T, Yang Y, et al. Learning from massive noisy labeled data for image classification [C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2015: 2691-2699. \n[40] Misra I, Lawrence Zitnick C, Mitchell M, et al. Seeing through the human reporting bias: Visual classifiers from noisy human-centric labels[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2016: 2930-2939. \n[41] Tanno R, Saeedi A, Sankaranarayanan S, et al. Learning from noisy labels by regularized estimation of annotator confusion[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019: 11244-11253. \n[42] Rodrigues F, Pereira F. Deep learning from crowds[C]//Proceedings of the AAAI Conference on Artificial Intelligence: volume 32. 2018. \n[43] Branson S, Van Horn G, Perona P. Lean crowdsourcing: Combining humans and machines in an online system[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2017: 7474-7483. \n[44] Izadinia H, Russell B C, Farhadi A, et al. Deep classifiers from image tags in the wild[C]// Proceedings of the 2015 Workshop on Community-Organized Multimodal Mining: Opportunities for Novel Solutions. 2015: 13-18. \n[45] Jaehwan L, Donggeun Y, Hyo-Eun K. Photometric transformer networks and label adjustment for breast density prediction[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. 2019: 0-0. \n[46] Veit A, Alldrin N, Chechik G, et al. Learning from noisy large-scale datasets with minimal supervision[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 839-847. \n[47] Tanaka D, Ikami D, Yamasaki T, et al. Joint optimization framework for learning with noisy labels[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018: 5552-5560. \n[48] Zheng S, Wu P, Goswami A, et al. Error-bounded correction of noisy labels[C]//International Conference on Machine Learning. PMLR, 2020: 11447-11457. \n[49] Han J, Luo P, Wang X. Deep self-learning from noisy labels[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. 2019: 5138-5147. \n[50] Durand T, Mehrasa N, Mori G. Learning a deep convnet for multi-label classification with partial labels[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019: 647-657. \n[51] Wu X, He R, Sun Z, et al. A light cnn for deep face representation with noisy labels[J]. IEEE Transactions on Information Forensics and Security, 2018, 13(11):2884-2896. \n[52] Huang J, Qu L, Jia R, et al. O2u-net: A simple noisy label detection approach for deep neural networks[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. 2019: 3326-3334. \n[53] Li J, Socher R, Hoi S C. Dividemix: Learning with noisy labels as semi-supervised learning [C/OL]//International Conference on Learning Representations. 2020. https://openreview.net/ forum?id=HJgExaVtwr. \n[54] Yan Y, Xu Z, Tsang I, et al. Robust semi-supervised learning through label aggregation[C]// Proceedings of the AAAI Conference on Artificial Intelligence: volume 30. 2016. \n[55] Jiang J, Ma J, Wang Z, et al. Hyperspectral image classification in the presence of noisy labels [J]. IEEE Transactions on Geoscience and Remote Sensing, 2018, 57(2):851-865. \n[56] Li J, Wong Y, Zhao Q, et al. Learning to learn from noisy labeled data[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019: 5051-5059. \n[57] Li Y, Yang J, Song Y, et al. Learning from noisy labels with distillation[C]//Proceedings of the IEEE International Conference on Computer Vision. 2017: 1910-1918. \n[58] Algan G, Ulusoy I. Meta soft label generation for noisy labels[C]//2020 25th International Conference on Pattern Recognition (ICPR). IEEE, 2021: 7142-7148. \n[59] Garcia L P, de Carvalho A C, Lorena A C. Noise detection in the meta-learning level[J]. Neurocomputing, 2016, 176:14-25. \n[60] Thirion J P. Image matching as a diffusion process: an analogy with maxwell’s demons[J]. Medical image analysis, 1998, 2:243260. \n[61] Rueckert D, Sonoda L I, Hayes C, et al. Nonrigid registration using free-form deformations: application to breast mr images[J]. IEEE transactions on medical imaging, 1999, 18(8):712- 721. \n[62] Shen D, Davatzikos C. Hammer: hierarchical attribute matching mechanism for elastic registration[J]. IEEE transactions on medical imaging, 2002, 21(11):1421-1439. \n[63] Balakrishnan G, Zhao A, Sabuncu M R, et al. Voxelmorph: a learning framework for deformable medical image registration[J]. IEEE transactions on medical imaging, 2019, 38(8): 1788-1800. \n[64] Shen Z, Han X, Xu Z, et al. Networks for joint affine and non-parametric image registration [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2019: 4224-4233. \n[65] Mok T C, Chung A. Fast symmetric diffeomorphic image registration with convolutional neural networks[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020: 4644-4653. \n[66] Wang J, Zhang M. Deepflash: An efficient network for learning-based medical image registration[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020: 4444-4452. \n[67] Arar M, Ginger Y, Danon D, et al. Unsupervised multi-modal image registration via geometry preserving image-to-image translation[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020: 13410-13419. \n[68] Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation[C]//International Conference on Medical image computing and computer-assisted intervention. Springer, 2015: 234-241. \n[69] Menze B H, Jakab A, Bauer S, et al. The multimodal brain tumor image segmentation benchmark (brats)[J]. IEEE transactions on medical imaging, 2014, 34(10):1993-2024. ",
955
+ "bbox": [
956
+ 171,
957
+ 55,
958
+ 826,
959
+ 917
960
+ ],
961
+ "page_idx": 10
962
+ },
963
+ {
964
+ "type": "text",
965
+ "text": "",
966
+ "bbox": [
967
+ 169,
968
+ 78,
969
+ 828,
970
+ 915
971
+ ],
972
+ "page_idx": 11
973
+ },
974
+ {
975
+ "type": "text",
976
+ "text": "",
977
+ "bbox": [
978
+ 171,
979
+ 98,
980
+ 828,
981
+ 916
982
+ ],
983
+ "page_idx": 12
984
+ },
985
+ {
986
+ "type": "text",
987
+ "text": "",
988
+ "bbox": [
989
+ 169,
990
+ 71,
991
+ 828,
992
+ 916
993
+ ],
994
+ "page_idx": 13
995
+ },
996
+ {
997
+ "type": "text",
998
+ "text": "",
999
+ "bbox": [
1000
+ 171,
1001
+ 90,
1002
+ 828,
1003
+ 171
1004
+ ],
1005
+ "page_idx": 14
1006
+ }
1007
+ ]
parse/train/C0GmZH2RnVR/C0GmZH2RnVR_middle.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/C0GmZH2RnVR/C0GmZH2RnVR_model.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/DfGu8WwT0d/DfGu8WwT0d.md ADDED
@@ -0,0 +1,368 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Large Scale Learning on Non-Homophilous Graphs: New Benchmarks and Strong Simple Methods
2
+
3
+ Derek Lim⇤ Cornell University dl772@cornell.edu
4
+
5
+ Felix Hohne⇤ Cornell University fmh42@cornell.edu
6
+
7
+ Xiuyu Li⇤ Cornell University xl289@cornell.edu
8
+
9
+ Sijia Linda Huang Cornell University sh837@cornell.edu
10
+
11
+ Vaishnavi Gupta Cornell University vg222@cornell.edu
12
+
13
+ Omkar Bhalerao Cornell University opb7@cornell.edu
14
+
15
+ Ser-Nam Lim Facebook AI sernam@gmail.com
16
+
17
+ # Abstract
18
+
19
+ Many widely used datasets for graph machine learning tasks have generally been homophilous, where nodes with similar labels connect to each other. Recently, new Graph Neural Networks (GNNs) have been developed that move beyond the homophily regime; however, their evaluation has often been conducted on small graphs with limited application domains. We collect and introduce diverse nonhomophilous datasets from a variety of application areas that have up to $3 8 4 \mathrm { x }$ more nodes and 1398x more edges than prior datasets. We further show that existing scalable graph learning and graph minibatching techniques lead to performance degradation on these non-homophilous datasets, thus highlighting the need for further work on scalable non-homophilous methods. To address these concerns, we introduce LINKX — a strong simple method that admits straightforward minibatch training and inference. Extensive experimental results with representative simple methods and GNNs across our proposed datasets show that LINKX achieves state-ofthe-art performance for learning on non-homophilous graphs. Our codes and data are available at https://github.com/CUAI/Non-Homophily-Large-Scale.
20
+
21
+ # 1 Introduction
22
+
23
+ Graph learning methods generate predictions by leveraging complex inductive biases captured in the topology of the graph [7]. A large volume of work in this area, including graph neural networks (GNNs), exploits homophily as a strong inductive bias, where connected nodes tend to be similar to each other in terms of labels [46, 3]. Such assumptions of homophily, however, do not always hold true. For example, malicious node detection, a key application of graph machine learning, is known to be non-homophilous in many settings [55, 13, 25, 11].
24
+
25
+ Further, while new GNNs that work better in these non-homophilous settings have been developed [82, 44, 81, 17, 15, 73, 36, 35, 9, 54], their evaluation is limited to a few graph datasets used by Pei et al. [58] (collected by [61, 66, 48]) that have certain undesirable properties such as small size, narrow range of application areas, and high variance between different train/test splits $[ \textcircled { 8 2 } ]$ Consequently, method scalability has not been thoroughly studied in non-homophilous graph learning. In fact, many non-homophilous techniques frequently require more parameters and computational resources [82, 1, 17], which is neither evident nor detrimental when they are evaluated on very small datasets. Even though scalable graph learning techniques do exist, these methods generally cannot be directly applied to the non-homophilous setting, as they oftentimes assume homophily in their construction $\boxed { \boxed { 7 1 } \boxed { 3 2 } \boxed { 2 0 } \boxed { 1 0 } }$ .
26
+
27
+ Non-homophily in graphs also degrades proven graph learning techniques that have been instrumental to strong performance in scalable graph learning. For instance, label propagation, personalized PageRank, and low-pass graph filtering have been used for scalable graph representation learning models, but these methods all assume homophily [71, 32, 20, 10]. Moreover, we give empirical evidence that existing minibatching techniques in graph learning [16, $\textcircled { 7 7 }$ significantly degrade performance in non-homophilous settings. In response, we develop a novel model, LINKX, that addresses these concerns; LINKX outperforms existing graph learning methods on large-scale nonhomophilous datasets and admits a simple minibatching procedure that maintains strong performance.
28
+
29
+ To summarize, we demonstrate three key areas of deficiency as mentioned above, namely: (1) that there is a lack of large, high-quality datasets covering different non-homophilous applications, (2) that current graph minibatching techniques and scalable methods do not work well in non-homophilous settings, and (3) that prior non-homophilous methods are not scalable. To these ends, this paper makes the following contributions:
30
+
31
+ Dataset Collection and Benchmarking. We collect a diverse series of large, non-homophilous graph datasets and define new node features and tasks for classification. These datasets are substantially larger than previous non-homophilous datasets, span wider application areas, and capture different types of complex label-topology relationships. With these proposed datasets, we conduct extensive experiments with 14 graph learning methods and 3 graph minibatching techniques that are broadly representative of the graph machine learning model space.
32
+
33
+ Analyzing Scalable Methods and Minibatching. We analyze current graph minibatching techniques like GraphSAINT $\mathbb { [ [ \overline { { \mathsf { Z } } } ] ] }$ in non-homophilous settings, showing that they substantially degrade performance in experiments. Also, we show empirically that scalable methods for graph learning like SGC and C&S [71, 32] do not perform well in non-homophilous settings — even though they achieve state-of-the-art results on many homophilous graph benchmarks. Finally, we demonstrate that existing non-homophilous methods often suffer from issues with scalability and performance in large non-homophilous graphs, in large part due to a lack of study of large-scale non-homophilous graph learning.
34
+
35
+ LINKX: a strong, simple method. We propose a simple method LINKX that achieves excellent results for non-homophilous graphs while overcoming the above-mentioned minibatching issues. LINKX works by separately embedding the adjacency A and node features $\mathbf { X }$ , then combining them with multilayer perceptrons and simple transformations, as illustrated in Figure $^ { 1 . }$ It generalizes node feature MLP and LINK regression $\textcircled { 1 7 9 }$ , two baselines that often work well on non-homophilous graphs. This method is simple to train and evaluate in a minibatched fashion, and does not face the performance degradation that other methods do in the minibatch setting. We develop the model and give more details in Section 4.
36
+
37
+ ![](images/c3d7544b0a9ef11f26479911d9394e4a333f361b16edd53fa90275c21a294b00.jpg)
38
+ Figure 1: Our model LINKX separately embeds node features and adjacency information with MLPs, combines the embeddings together by concatenation, then uses a final MLP to generate predictions.
39
+
40
+ # 2 Prior Work
41
+
42
+ Graph Representation Learning. Graph neural networks $\textcircled { 1 2 8 } , \textcircled { 3 8 } , \textcircled { 6 9 }$ have demonstrated their utility on a variety of graph machine learning tasks. Most GNNs are constructed by stacking layers that propagate transformed node features, which are then aggregated via different mechanisms. The neighborhood aggregation used in many existing GNNs implicitly leverage homophily, so they often fail to generalize on non-homophilous graphs $\boxed { 8 2 } \boxed { 6 }$ . Indeed, a wide range of GNNs operate as low-pass graph filters $\pm \pm \pm \pm \pm \pm \pm \pm \pm \pm \pm \pm$ that smooth features over the graph topology, which produces similar representations and thus similar predictions for neighboring nodes.
43
+
44
+ Scalable methods. A variety of scalable graph learning methods have been developed for efficient computation in larger datasets [77, 16, 75, 28, 71, 32, 20, 10]. Many of these methods explicitly make use of an assumption of homophily in the data [71, 32, 20, 10]. By leveraging this assumption, several simple, inexpensive models are able to achieve state-of-the-art performance on homophilic datasets [71, 32]. However, these methods are unable to achieve comparable performance in non-homophilous settings, as we show empirically in Section 5.
45
+
46
+ Graph sampling. As node representations depend on other nodes in the graph, there are no simple minibatching techniques in graph learning as there are for i.i.d. data. To scale to large graphs, one line of work samples nodes that are used in each layer of a graph neural network $\overline { { \| 2 8 \| } } \widetilde { \bigtriangledown } \widetilde { \| 4 \| }$ . Another family of methods samples subgraphs of an input graph, then passes each subgraph through a GNN to make a prediction for each node of the subgraph [16, 76, 77]. While these methods are useful for scalable graph learning, we show that they substantially degrade performance in our non-homphilous experiments (see Section 5)
47
+
48
+ Non-Homophilous methods. Various GNNs have been proposed to achieve higher performance in low-homophily settings [82, 44, 81, 17, 15, 73, 36, 35]. Geom-GCN $\left[ \left[ 5 8 \right] \right]$ introduces a geometric aggregation scheme, MixHop [1] proposes a graph convolutional layer that mixes powers of the adjacency matrix, GPR-GNN [17] features learnable weights that can be positive and negative in feature propagation, GCNII [15] allows deep graph convolutional networks with relieved oversmoothing, which empirically performs better in non-homophilous settings, and $\mathrm { H } _ { 2 } \mathrm { G C N }$ $\pmb { \Vert 8 2 \Vert }$ shows that separation of ego and neighbor embeddings, aggregation in higher-order neighborhoods, and the combination of intermediate representations improves GNN performance in low-homophily.
49
+
50
+ There are several recurring design decisions across these methods that appear to strengthen performance in non-homophilous settings: using higher-order neighborhoods, decoupling neighbor information from ego information, and combining graph information at different scales $[ \bar { 8 2 } ] $ . Many of these design choices require additional overhead (see Section $4 . 3 )$ , thus reducing their scalability.
51
+
52
+ Datasets. The widely used citation networks Cora, Citeseer, and Pubmed $[ [ 6 2 , 7 4 ]$ are highly homophilous (see Appendix $\mathbf { A } )$ $\pmb { \Vert 8 2 \Vert }$ . Recently, the Open Graph Benchmark [31] has provided a series of datasets and leaderboards that improve the quality of evaluation in graph representation learning; however, most of the node classification datasets tend to be homophilous, as noted in past work $\lVert \tilde { \otimes } 2 \rVert$ and expanded upon in Appendix $\boxed { \mathbf { A } . 2 }$ A comparable set of high-quality benchmarks to evaluate non-homophilous methods does not currently exist.
53
+
54
+ # 3 Datasets for Non-Homophilous Graph Learning
55
+
56
+ # 3.1 Currently Used Datasets
57
+
58
+ The most widely used datasets to evaluate non-homophilous graph representation learning methods were used by Pei et al. $\left[ \left[ 5 8 \right] \right]$ (and collected by [61, 66, 48]); see our Table $^ 1$ for statistics. However, these datasets have fundamental issues. First, they are very small — the Cornell, Texas, and Wisconsin datasets have between 180-250 nodes, and the largest dataset Actor has 7,600 nodes. In analogy to certain pitfalls of graph neural network evaluation on small (homophilic) datasets discussed in $\overline { { \mathbb { B } 3 } } \|$ , evaluation on the datasets of Pei et al. $ { \mathbb { B } } { \mathbb { B } }$ is plagued by high variance across different train/test splits (see results in $\mathbb { \lVert 8 2 \rVert }$ ). The small size of these datasets may tend to create models that are more prone to overfitting $\pmb { \mathbb { D } } \mathbf { 1 } \mathbf { h }$ , which prevents the scaling up of GNNs designed for non-homophilous settings.
59
+
60
+ Peel $ { \mathbb { B } } 5 7 { \mathbb { I } }$ also studies node classification on network datasets with various types of relationships between edges and labels. However, they only study methods that act on graph topology, and thus their datasets do not necessarily have node features. We take inspiration from their work, by testing on Pokec and Facebook networks with node features that we define, and by introducing other year-prediction tasks on citation networks that have node features.
61
+
62
+ # 3.2 An Improved Homophily Measure
63
+
64
+ Various metrics have been proposed to measure the homophily of a graph. However, these metrics are sensitive to the number of classes and the number of nodes in each class. Let $G = ( V , E )$ be a graph with $n$ nodes, none of which are isolated. Further let each node $u \in V$ have a class label $k _ { u } \in \{ 0 , 1 , \ldots , C - 1 \}$ for some number of classes $C$ , and denote by $C _ { k }$ the set of nodes in class $k$ . The edge homophily $\lVert 8 2 \rVert$ is the proportion of edges that connect two nodes of the same class:
65
+
66
+ $$
67
+ h = { \frac { | \{ ( u , v ) \in E : k _ { u } = k _ { v } \} | } { | E | } } .
68
+ $$
69
+
70
+ Another related measure is what we call the node homophily $\begin{array} { r l } { \mathrm { \bf \delta } } & { { } \| \boldsymbol { \bar { 5 } } \mathrm { \bf { 8 } } \| } \end{array}$ , defined as $\begin{array} { r } { \frac { 1 } { \vert V \vert } \sum _ { u \in V } \frac { d _ { u } ^ { ( k _ { u } ) } } { d _ { u } } } \end{array}$ , in which $d _ { u }$ is the number of neighbors of node $u$ , and $d _ { u } ^ { ( k _ { u } ) }$ is the number of neighbors of $u$ that have the same class label. We focus on the edge homophily $\mathbb { \underline { { ( 1 ) } } }$ in this work, but find that node homophily tends to have similar qualitative behavior in experiments.
71
+
72
+ The sensitivity of edge homophily to the number of classes and size of each class limits its utility. We consider a null model for graphs in which the graph topology is independent of the labels; suppose that nodes with corresponding labels are fixed, and include edges uniformly at random in the graph that are independent of node labels. Under this null model, a node $u \in V$ would be expected to have $d _ { u } ^ { ( k _ { u } ) } / d _ { u } \approx | C _ { k _ { u } } | / n$ as the proportion of nodes of the same class that they connect to $\mathbb { \left[ 3 \right] }$ . For a dataset with $C$ balanced classes, we would thus expect the edge homophily to be around $\textstyle { \frac { 1 } { C } }$ , so the interpretation of the measure depends on the number of classes. Also, if classes are imbalanced, then the edge homophily may be misleadingly large. For instance, if $9 9 \%$ of nodes were of one class, then most edges would likely be within that same class, so the edge homophily would be high, even when the graph is generated from the null model where labels are independent of graph topology. Thus, the edge homophily does not capture deviation of the label distribution from the null model.
73
+
74
+ We introduce a metric that better captures the presence or absence of homophily. Unlike the edge homophily, our metric measures excess homophily that is not expected from the above null model where edges are randomly wired. Our metric does not distinguish between different non-homophilous settings (such as heterophily or independent edges); we believe that there are too many degrees of freedom in non-homophilous settings for a single scalar quantity to be able to distinguish them all. Our measure is given as:
75
+
76
+ $$
77
+ \hat { h } = \frac { 1 } { C - 1 } \sum _ { k = 0 } ^ { C - 1 } \left[ h _ { k } - \frac { | C _ { k } | } { n } \right] _ { + } ,
78
+ $$
79
+
80
+ where $[ a ] _ { + } = \operatorname* { m a x } ( a , 0 )$ , and $h _ { k }$ is the class-wise homophily metric
81
+
82
+ $$
83
+ h _ { k } = \frac { \sum _ { u \in C _ { k } } d _ { u } ^ { ( k _ { u } ) } } { \sum _ { u \in C _ { k } } d _ { u } } .
84
+ $$
85
+
86
+ Note that $\hat { h } \in [ 0 , 1 ]$ , with a fully homophilous graph (in which every node is only connected to nodes of the same class) having $\hat { h } = 1$ . Since each class-wise homophily metric $h _ { k }$ only contributes positive deviations from the null expected proportion $| C _ { k } | / n$ , the class-imbalance problem is substantially mitigated. Also, graphs in which edges are independent of node labels are expected to have $\hat { h } \approx 0$ , for any number of classes. Our measure $\hat { h }$ measures presence of homophily, but does not distinguish between the many types of possibly non-homophilous relationships. This is reasonable given the diversity of non-homophilous relationships. For example, non-homophily can imply independence of edges and classes, extreme heterophily, connections only among subsets of classes, or certain chemically / biologically determined relationships. Indeed, these relationships are very different, and are better captured by more than one scalar quantity, such as the compatibility matrices presented in the appendix. Further discussion is given in Appendix A.
87
+
88
+ Table 1: Statistics for previously used datasets from Pei et al. $\pmb { \Vert 5 8 \Vert }$ (collected by [61, 66, 48]). #C is the number of node classes. The highest number of nodes or edges overall are bolded.
89
+
90
+ <table><tr><td>Dataset</td><td># Nodes</td><td>#Edges</td><td>#Feat.</td><td>#C</td><td>Context</td><td>Edge hom.</td><td>h (ours)</td></tr><tr><td>Chameleon</td><td>2,277</td><td>36,101</td><td>2,325</td><td>5</td><td>Wiki pages</td><td>.23</td><td>.062</td></tr><tr><td>Cornell</td><td>183</td><td>295</td><td>1,703</td><td>5</td><td>Web pages</td><td>.30</td><td>.047</td></tr><tr><td>Actor</td><td>7,600</td><td>29,926</td><td>931</td><td>5</td><td>Actors in movies</td><td>.22</td><td>.011</td></tr><tr><td>Squirrel</td><td>5,201</td><td>216,933</td><td>2.089</td><td>5</td><td>Wiki pages</td><td>.22</td><td>.025</td></tr><tr><td>Texas</td><td>183</td><td>309</td><td>1,703</td><td>5</td><td>Web pages</td><td>.11</td><td>.001</td></tr><tr><td>Wisconsin</td><td>251</td><td>499</td><td>1,703</td><td>5</td><td>Web pages</td><td>.21</td><td>.094</td></tr></table>
91
+
92
+ # 3.3 Proposed Datasets
93
+
94
+ Here, we detail the non-homophilous datasets that we propose for graph machine learning evaluation. Our datasets and tasks span diverse application areas. Penn94 [67], Pokec [41], genius [43], and twitch-gamers $[ [ 6 0 ] ]$ are online social networks, where the task is to predict reported gender, certain account labels, or use of explicit content on user accounts. For the citation networks arXiv-year [31] and snap-patents [42, 41] the goal is to predict year of paper publication or the year that a patent is granted. The dataset wiki consists of Wikipedia articles, where the goal is to predict total page views of each article. Detailed descriptions about the graph structure, node features, node labels, and licenses of each dataset are given in Appendix D.2.
95
+
96
+ Most of these datasets have been used for evaluation of graph machine learning models in past work; we make adjustments such as modifying node labels and adding node features that allow for evaluation of GNNs in non-homophilous settings. We define node features for Pokec, genius, and snap-patents, and we also define node labels for arXiv-year, snap-patents, and genius. Additionally, we crawl and clean the large-scale wiki dataset — a new Wikipedia dataset where the task is to predict page views, which is non-homophilous with respect to the graph of articles connected by links between articles (see Appendix $\mathbf { D } . { \dot { 3 } } )$ . This wiki dataset has 1,925,342 nodes and 303,434,860 edges, so training and inference require scalable algorithms.
97
+
98
+ Basic dataset statistics are given in Table 2. Note the substantial difference between the size of our datasets and those of Pei et al. [58] in Table 1; our datasets have up to $3 8 4 \mathrm { x }$ more nodes and $1 3 9 8 \mathrm { x }$ more edges. The homophily measures along with the lower empirical performance of homophilyassuming models (Section $\dot { 5 } _ { , }$ and examination of compatibility matrices (Appendix $\boxed { \mathrm { A } }$ show that our datasets are indeed non-homophilous. As there is little study in large-scale non-homophilous graph learning, our proposed large datasets strongly motivate the need for developing a new, scalable approach that can accurately learn on non-homophilous graphs.
99
+
100
+ Table 2: Statistics of our proposed non-homophilous graph datasets. # C is the number of distinct node classes. Note that our datasets come from more diverse applications areas and are much larger than those shown in Table 1, with up to 384x more nodes and $1 3 9 8 \mathrm { x }$ more edges.
101
+
102
+ <table><tr><td>Dataset</td><td>#Nodes</td><td>#Edges</td><td>#Feat.</td><td>#C</td><td>Class types</td><td>Edge hom.</td><td>h (ours)</td></tr><tr><td>Penn94</td><td>41,554</td><td>1,362,229</td><td>5</td><td>2</td><td>gender</td><td>.470</td><td>.046</td></tr><tr><td>pokec</td><td>1,632,803</td><td>30,622,564</td><td>65</td><td>2</td><td>gender</td><td>.445</td><td>.000</td></tr><tr><td>arXiv-year</td><td>169,343</td><td>1,166,243</td><td>128</td><td>5</td><td> pub year</td><td>.222</td><td>.272</td></tr><tr><td>snap-patents</td><td>2,923,922</td><td>13,975,788</td><td>269</td><td>5</td><td>time granted</td><td>.073</td><td>.100</td></tr><tr><td>genius</td><td>421,961</td><td>984,979</td><td>12</td><td>2</td><td>marked act.</td><td>.618</td><td>.080</td></tr><tr><td>twitch-gamers</td><td>168,114</td><td>6,797,557</td><td>7</td><td>2</td><td>mature content</td><td>.545</td><td>.090</td></tr><tr><td>wiki</td><td>1,925,342</td><td>303,434,860</td><td>600</td><td>5</td><td>views</td><td>.389</td><td>.107</td></tr></table>
103
+
104
+ # 4 LINKX: A New Scalable Model
105
+
106
+ In this section, we introduce our novel model, LINKX, for scalable node classification in nonhomophilous settings. LINKX is built out of multilayer perceptrons (MLPs) and linear transformations, thus making it simple and scalable. It also admits simple row-wise minibatching procedures that allow it to perform well on large non-homophilous graphs. As a result, LINKX is able to circumvent aforementioned issues of graph minibatching and non-homophilous GNNs in large-scale non-homophilous settings.
107
+
108
+ # 4.1 Motivation from two simple baselines
109
+
110
+ Here, we detail two simple baselines for node classification that we build on to develop LINKX.
111
+
112
+ MLP on node features. A naïve method for node classification is to ignore the graph topology and simply train an MLP on node features. For the same reason that the graph topology has more complicated relationships with label distributions in non-homophilous graphs, many GNNs are not able to effectively leverage the graph topology in these settings. Thus, MLPs can actually perform comparatively well on non-homophilous graphs — achieving higher or approximately equal performance to various GNNs $\pmb { \Vert 8 2 \Vert }$ .
113
+
114
+ LINK regression on graph topology. On the other extreme, there is LINK $\left[ \left[ 7 9 \right] \right]$ — a simple baseline that only utilizes graph topology. In particular, we consider LINK regression, which trains a logistic regression model in which each node’s features are taken from a column of the adjacency matrix. Letting $\mathbf { A } \in \{ 0 , 1 \} ^ { n \times n }$ be the binary adjacency matrix of the graph, and $\mathbf { W } \in \mathbb { R } ^ { c \hat { \times } n }$ be a learned weight matrix, LINK computes class probabilities as
115
+
116
+ $$
117
+ \mathbf { Y } = \operatorname { s o f t m a x } ( \mathbf { W } \mathbf { A } ) .
118
+ $$
119
+
120
+ Let $u \in \{ 1 , \ldots , n \}$ be a specific node, and let $k \in \{ 1 , \ldots , c \}$ be a specific class. Then, expanding the matrix multiplication, the log-odds of node $u$ belonging to class $k$ is given by
121
+
122
+ $$
123
+ ( \mathbf { W } \mathbf { A } ) _ { k u } = \sum _ { v \in \mathcal { N } ( u ) } \mathbf { W } _ { k v } ,
124
+ $$
125
+
126
+ where $\mathcal { N } ( u )$ contains the 1-hop neighbors of $u$ . In other words, the logit is given by the sum of weights $\mathbf { W } _ { k v }$ across the 1-hop neighbors of $u$ . If a specific node $v$ has many neighbors of class $k$ , then $\mathbf { W } _ { k v }$ is probably large, as we would expect with a high probability that any neighbor of $v$ is of class $k$ . In this sense, LINK is like a 2-hop method: for a given node $u$ , the probability of being in a given class is related to the class memberships of $u$ ’s 2-hop neighbors in $\bar { \mathcal { N } } ( v )$ for each neighbor $v \in \mathcal { N } ( u )$ . Related interpretations of LINK as a method acting on 2-hop paths between nodes are given by Altenburger and Ugander $\pmb { \mathbb { B } } \|$ .
127
+
128
+ Though it is simple and has been overlooked in the recent non-homophilous GNN literature, LINK has been found to perform well in certain node classification tasks like gender prediction in social networks $\mathbb { B } \mathbb { H }$ . A major reason why LINK does well in many settings is exactly because it acts as a 2-hop method. For example, while 1-hop neighbors are often not so informative for gender prediction in social networks due to lack of homophily, 2-hop neighbors are very informative due to so-called “monophily,” whereby many nodes have extreme preferences for connecting to a certain class $\pmb { \mathbb { B } } \|$ . Beyond just gender prediction, we show in Section $\boxed { 5 }$ that LINK empirically outperforms many models across the various application areas of the non-homophilous datasets we propose.
129
+
130
+ # 4.2 LINKX
131
+
132
+ We combine these two simple baselines through simple linear transformations and component-wise nonlinearities. Let $\mathbf { X } \in \mathbb { R } ^ { \boldsymbol { \hat { D } \times n } }$ denote the matrix of node features with input dimension $D$ , and let $\left[ \mathbf { h } _ { 1 } ; \mathbf { h } _ { 2 } \right]$ denote concatenation of vectors $\mathbf { h } _ { 1 }$ and $\mathbf { h } _ { 2 }$ . Then our model outputs predictions $\mathbf { Y }$ through the following mapping:
133
+
134
+ $$
135
+ \begin{array} { r l } & { \mathbf { h } _ { \mathbf { A } } = \mathrm { M L P } _ { \mathbf { A } } ( \mathbf { A } ) \in \mathbb { R } ^ { d \times n } } \\ & { \mathbf { h } _ { \mathbf { X } } = \mathrm { M L P } _ { \mathbf { X } } ( \mathbf { X } ) \in \mathbb { R } ^ { d \times n } } \\ & { \mathbf { Y } = \mathrm { M L P } _ { f } \left( \sigma \left( \mathbf { W } [ \mathbf { h } _ { \mathbf { A } } ; \mathbf { h } _ { \mathbf { X } } ] + \mathbf { h } _ { \mathbf { A } } + \mathbf { h } _ { \mathbf { X } } \right) \right) , } \end{array}
136
+ $$
137
+
138
+ in which $d$ is the hidden dimension, $\mathbf { W } \in \mathbb { R } ^ { d \times 2 d }$ is a weight matrix, and $\sigma$ is a component-wise nonlinearity (which we take to be ReLU). We call our model LINKX, as it extends LINK with node feature information from the matrix $\mathbf { X }$ . A diagram of LINKX is given in Figure 1.
139
+
140
+ First, LINKX computes hidden representations $\mathbf { h _ { A } }$ of the adjacency (extending LINK) and $\mathbf { h } _ { \mathbf { X } }$ of the feature matrix (as in node-feature MLPs). Then it combines these hidden representations through a linear transform W of their concatenation, with skip connections that add back in $\mathbf { h _ { A } }$ and $\mathbf { h } _ { \mathbf { X } }$ to better preserve pure adjacency or node feature information. Finally, it puts this combined representation through a non-linearity and another MLP to make a prediction.
141
+
142
+ Separating then mixing adjacency and feature information. LINKX separately embeds the adjacency A to $\mathbf { h _ { A } }$ and the features $\mathbf { X }$ into $\mathbf { h } _ { \mathbf { X } }$ before mixing them for a few reasons. First, we note that this design is reminiscent of fusion architectures in multimodal networks, where data from different modalities are processed and combined in a neural network $\underline { { [ 2 4 , 7 8 ] } }$ . In our setting, we can view adjacency information and node feature information as separate modalities. Since node feature MLPs and LINK do well independently on different datasets, this allows us to preserve their individual performance if needed. Ignoring $\mathbf { h } _ { \mathbf { X } }$ information is similar to just using LINK, and ignoring $\mathbf { h _ { A } }$ information is just using an node feature MLP. Still, to preserve the ability to just learn a similar mapping to LINK or to a node feature MLP, we find that having the additive skip connections helps to get performance at least as good as either baseline. Our initial empirical results showed that simply concatenating adjacency and node features as input to a network does worse overall empirically (see Appendix $\checkmark$
143
+
144
+ There are also computational benefits to our design choices. Embedding A is beneficial for depth as adding more layers to the MLPs only gives an $\mathcal { O } ( \bar { d } ^ { 2 } )$ cost — depending only on the hidden dimension $d$ — and thus does not scale in the number of edges $| E |$ as when adding layers to message-passing GNNs. This is because the graph information in $\mathbf { A }$ is already compressed to hidden feature vectors after the first linear mapping of $\mathrm { M L P } _ { \mathbf { A } }$ , and we do not need to propagate along the graph in later steps. Moreover, this enables a sparse-dense matrix product to compute the first linear mapping of $\mathrm { M L P } _ { \mathbf { A } }$ on A, which greatly increases efficiency as $\mathbf { A }$ is typically very sparse for real-world graphs. Separate embeddings are key here, as this would not be possible if we for instance concatenated A and $\mathbf { X }$ when $\mathbf { X }$ is large and dense.
145
+
146
+ Simple minibatching. Message-passing GNNs must take graph topology into account when minibatching with techniques such as neighbor sampling, subgraph sampling, or graph partitioning. However, LINKX does not require this, as it utilizes graph information solely through defining adjacency matrix columns as features. Thus, we can train LINKX with standard stochastic gradient descent variants by taking i.i.d. samples of nodes along with the corresponding columns of the adjacency and feature matrix as features. This is much simpler than the graph minibatching procedures for message-passing GNNs, which require specific hyperparameter choices, have to avoid exponential blowup of number of neighbors per layer, and are generally more complex to implement $\mathbb { \ m }$ . In Section $\left\lceil 5 . 3 \right\rceil$ we use the simple LINKX minibatching procedure for large-scale experiments that show that LINKX with this minibatching style outperforms GNNs with graph minibatching methods. This is especially important on the scale of the wiki dataset, where none of our tested methods — other than MLP — is capable of running on a Titan RTX GPU with 24 GB GPU RAM (see Section $\textcircled { 5 }$ .
147
+
148
+ # 4.3 Complexity Analysis
149
+
150
+ Using the above notation, a forward pass of LINKX has a time complexity of $\mathcal { O } \left( d \vert E \vert + n d ^ { 2 } L \right)$ , in which $d$ is the hidden dimension (which we assume to be on the same order as the input feature dimension $D$ ), $L$ is the number of layers, $n$ is the number of nodes, and $| E |$ is the number of edges. We require a $\mathcal { O } ( d \vert E \vert )$ cost for the first linear mapping of $\mathbf { A }$ and a $\mathcal { O } ( d ^ { 2 } )$ cost per layer for MLP operations on hidden features, for $L$ total layers and each of $n$ nodes.
151
+
152
+ As mentioned above, message passing GNNs have to propagate using the adjacency in each layer, so they have an $L | E |$ term in the complexity. For instance, an $L$ -layer GCN $\pmb { \Vert 3 8 \Vert }$ with $d$ hidden dimensions has $\dot { \mathcal { O } } ( d \dot { L } \vert E \vert + n d ^ { 2 } L )$ complexity, as it costs $\mathcal { O } ( d \vert E \vert )$ to propagate features in each layer, and $\mathcal { O } ( n d ^ { 2 } )$ to multiply by the weight matrix in each layer.
153
+
154
+ Non-homophilous methods often make modifications to standard architectures that increase computational cost, such as using higher-order neighborhoods or using additional hidden embeddings $\mathbf { \widehat { \| 8 2 \| } }$ . For instance, the complexity of MixHop $\mathbb { I I }$ is $\mathcal { O } ( K ( d L \vert E \vert + n d ^ { 2 } L ) )$ , which has an extra factor $K$ that is the number of adjacency powers to propagate with. The complexity of GCNII $\mathbb { \left. \overline { { 1 5 } } \right. }$ i s asymptotically the same as that of GCN, but in practice it requires more computations per layer due to residual connections and linear combinations, and it also often achieves best performance with a large number of layers $L$ . $\mathrm { H } _ { 2 } \mathrm { G C N }$ $\pmb { \Vert 8 2 \Vert }$ is significantly more expensive due to its usage of strict two-hop neighborhoods, which requires it to form the squared adjacency ${ \bf A } ^ { 2 }$ . This makes the memory requirements intractable even for medium sized graphs (see Section $\textcircled { 5 }$
155
+
156
+ # 5 Experiments
157
+
158
+ We conduct two sets of experiments for node classification on our proposed non-homophilous datasets. One set of experiments does full batch gradient descent training for all applicable methods. This of course limits the size of each model, as the large datasets require substantial GPU memory to train on. Our other set of experiments uses minibatching methods. As all graph-based methods run out of memory on the wiki dataset, even on 24 GB GPUs, we only include wiki results in the minibatching section. In all settings, our LINKX model matches or outperforms other methods.
159
+
160
+ Table 3: Experimental results. Test accuracy is displayed for most datasets, while genius displays test ROC AUC. Standard deviations are over 5 train/val/test splits. The three best results per dataset are highlighted. (M) denotes some (or all) hyperparameter settings run out of memory.
161
+
162
+ <table><tr><td></td><td>Penn94</td><td>pokec</td><td>arXiv-year</td><td>snap-patents</td><td>genius</td><td>twitch-gamers</td></tr><tr><td>MLP</td><td>73.61 ±0.40</td><td>62.37 ±0.02</td><td>36.70±0.21</td><td>31.34 ± 0.05</td><td>86.68± 0.09</td><td>60.92 ±0.07</td></tr><tr><td>L Prop 1-hop</td><td>63.21王0.39</td><td>53.09±0.05</td><td>43.42±0.17</td><td>30.28 ±0.09</td><td>66.02±0.16</td><td>62.77±0.24</td></tr><tr><td>L Prop 2-hop</td><td>74.13±0.46</td><td>76.76±0.03</td><td>46.07 ± 0.15</td><td>38.61 ±0.07</td><td>67.04 ± 0.20</td><td>63.88±0.24</td></tr><tr><td>LINK</td><td>80.79±0.49</td><td>80.54 ±0.03</td><td>53.97 ±0.18</td><td>60.39±0.07</td><td>73.56±0.14</td><td>64.85 ± 0.21</td></tr><tr><td>SGC 1-hop</td><td>66.790.27</td><td>53.61±0.17</td><td>32.83±0.13</td><td>30.31±0.06</td><td>82.36±0.37</td><td>58.97±0.19</td></tr><tr><td>SGC 2-hop</td><td>76.09 ±0.45</td><td>62.81 ± 1.42</td><td>32.27 ±0.06</td><td>29.09 ±0.09</td><td>82.10 ±0.14</td><td>59.94 ± 0.21</td></tr><tr><td>C&amp;S1-hop</td><td>74.28 ± 1.19</td><td>62.35 ±0.06</td><td>44.51 ±0.16</td><td>35.55 ± 0.05</td><td>82.93 ±0.15</td><td>64.86±0.27</td></tr><tr><td>C&amp;S2-hop</td><td>78.40±3.12</td><td>81.69 ±0.09</td><td>49.78±0.26</td><td>49.08±0.04</td><td>84.94 ± 0.49</td><td>65.02 ±0.16</td></tr><tr><td>GCN</td><td>82.470.27</td><td>75.45±0.17</td><td>46.02±0.26</td><td>45.65±0.04</td><td>87.42±0.37</td><td>62.18±0.26</td></tr><tr><td>GAT</td><td>81.53 ±0.55</td><td>71.77±6.18(M)</td><td>46.05 ±0.51</td><td>45.37± 0.44 (M)</td><td>55.80 ±0.87</td><td>59.89 ± 4.12</td></tr><tr><td>GCNJK</td><td>81.63 ±0.54</td><td>77.00±0.14</td><td>46.28±0.29</td><td>46.88 ±0.13</td><td>89.30 ±0.19</td><td>63.45±0.22</td></tr><tr><td>GATJK</td><td>80.69 ± 0.36</td><td>71.19 ±6.96 (M)</td><td>45.80±0.72</td><td>44.78±0.50</td><td>56.70±2.07</td><td>59.98 ±2.87</td></tr><tr><td>APPNP</td><td>74.33 ±0.38</td><td>62.58±0.08</td><td>38.15±0.26</td><td>32.19 ±0.07</td><td>85.36 ±0.62</td><td>60.97 ±0.10</td></tr><tr><td>HGCN</td><td>(M)</td><td>(M)</td><td>49.09±0.10</td><td>(M)</td><td>(M)</td><td>(M)</td></tr><tr><td>MixHop</td><td>83.47 ± 0.71</td><td>81.07 ± 0.16</td><td>51.81 ± 0.17</td><td>52.16 ± 0.09 (M)</td><td>90.58 ± 0.16</td><td>65.64 ± 0.27</td></tr><tr><td>GPR-GNN GCNII</td><td>81.38±0.16 82.92 ± 0.59</td><td>78.83± 0.05</td><td>45.07 ± 0.21</td><td>40.19±0.03</td><td>90.05 ± 0.31</td><td>61.89 ± 0.29</td></tr><tr><td></td><td></td><td>78.94± 0.11 (M)</td><td>47.21 ± 0.28</td><td>37.88 ±0.69 (M)</td><td>90.24 ± 0.09</td><td>63.39 ±0.61</td></tr><tr><td>LINKX</td><td>84.71 ± 0.52</td><td>82.04 ± 0.07</td><td>56.00 ±1.34</td><td>61.95 ± 0.12</td><td>90.77±0.27</td><td>66.06 ± 0.19</td></tr></table>
163
+
164
+ # 5.1 Experimental Setup
165
+
166
+ Methods. We include both methods that are graph-agnostic and node-feature-agnostic as simple baselines. The node-feature-agnostic models of two-hop label propagation $\mathbb { \lVert \Xi \rVert }$ and LINK (logistic regression on the adjacency matrix) $\pmb { \mathbb { Z } } 9 \|$ have been found to perform well in various non-homophilous settings, but they have often been overlooked by recent graph representation learning work. Also, we include SGC $\dot { [ ] } \dot { [ ] \begin{array} { r l } \end{array} }$ and C&S $\mathbb { \lVert 3 2 \rVert }$ as simple, scalable methods that perform well on homophilic datasets. We include a two-hop propagation variant of C&S in analogy with two-step label propagation. In addition to representative general GNNs, we also include GNNs recently proposed for non-homophilous settings. The full list of methods is: Only node features: MLP $\dot { \left. 2 6 \right. }$ . Only graph topology: label propagation (standard and two-hop) [80, 57], LINK $\mathbb { \underline { { \Pi 9 } } }$ . Simple methods: SGC [71], C&S [32] and their two-hop variants. General GNNs: GCN $\pmb { \| } \pmb { \mathrm { 3 8 } } \|$ , GAT $\overline { { \| 6 9 \| } }$ , jumping knowledge networks (GCNJK, GATJK) $\pmb { \mathbb { Z } 2 }$ , and APPNP $\pmb { \mathbb { B } } \pmb { \mathrm { 9 } } \|$ . Non-homophilous methods: $\mathrm { H } _ { 2 } \mathrm { G C N }$ $[ \textcircled { 8 2 } ]$ , MixHop [1], GPR-GNN [17], GCNII [15], and LINKX (ours).
167
+
168
+ Minibatching methods. We also evaluate GNNs with various minibatching methods. We take GCNJK $\pmb { \mathbb { Z } 2 } \|$ and MixHop [1] as our base models for evaluation, as they are representative of many GNN design choices and MixHop performs very well in full batch training. As other minibatching methods are trickier to make work with these models, we use the Cluster-GCN $\boxed { 1 6 }$ and GraphSAINT $\mathbb { \ m }$ minibatching methods, which sample subgraphs. We include both the node based sampling and random walk based sampling variants of GraphSAINT. We compare these GNNs with MLP, LINK, and our LINKX, which use simple i.i.d. node minibatching.
169
+
170
+ Training and evaluation. Following other works in non-homophilous graph learning evaluation, we take a high proportion of training nodes [82, 58, 73]; we run each method on the same five random 50/25/25 train/val/test splits for each dataset. All methods requiring gradient-based optimization are run for 500 epochs, with test performance reported for the learned parameters of highest validation performance. We use ROC-AUC as the metric for the class-imbalanced genius dataset (about $80 \%$ of nodes are in the majority class), as it is less sensitive to class-imbalance than accuracy. For other datasets, we use classification accuracy as the metric. Further experimental details are in Appendix B.
171
+
172
+ # 5.2 Full-Batch Results
173
+
174
+ Table $\triangledown$ lists the results of each method across the datasets that we propose. Our datasets reveal several important properties of non-homophilous node classification. Firstly, the stability of performance across runs is better for our datasets than those of Pei et al. $\left[ \left[ 5 8 \right] \right]$ (see $\pmb { \Vert 8 2 \Vert }$ results). Secondly, as suggested by prior theory and experiments [82, 1, 17], the non-homophilous GNNs usually do well — though not necessarily on every dataset.
175
+
176
+ The core assumption of homophily in SGC and C&S that enables them to be simple and efficient does not hold on these non-homophilous datasets, and thus the performance of these methods is typically relatively low. Still, as expected, two-hop variants generally improve upon their one-hop counter-parts in these low-homophily settings.
177
+
178
+ One consequence of using larger datasets for benchmarks is that the tradeoff between scalability and learning performance of non-homophilous methods has become starker, with some methods facing memory issues. This tradeoff is especially important to consider in light of the fact that many scalable graph learning methods rely on implicit or explicit homophily assumptions [71, 32, 20, 10], and thus face issues when used in non-homophilous settings.
179
+
180
+ Finally, LINKX achieves superior performance on all datasets, taking advantage of LINK’s power, while also being able to utilize node features where they provide additional information.
181
+
182
+ # 5.3 Minibatching Results
183
+
184
+ Table 4: Minibatching results on our proposed datasets. $\dagger$ denotes that 10 random partitions of the graphs are used for testing GraphSAINT sampling. (T) denotes that five runs takes $\geq 4 8$ hours for a single hyperparameter setting. Best results up to a standard deviation are highlighted.
185
+
186
+ <table><tr><td></td><td>Penn94</td><td>pokec †</td><td>arXiv-year</td><td>snap-patents †</td><td>genius</td><td>twitch-gamers †</td><td>wiki †</td></tr><tr><td>MLP Minibatch</td><td>74.24±0.55</td><td>62.14±0.05</td><td>36.89±0.11</td><td>22.96±0.81</td><td>82.35±0.38</td><td>61.01±0.06</td><td>37.38±0.21</td></tr><tr><td>LINK Minibatch</td><td>81.61±0.34</td><td>81.15±0.25</td><td>53.76±0.28</td><td>45.65±8.25</td><td>80.95±0.07</td><td>64.38±0.26</td><td>57.11±0.26</td></tr><tr><td>GCNJK-Cluster</td><td>69.99±0.85</td><td>72.67±0.05</td><td>44.05±0.11</td><td>37.62±0.31</td><td>83.04±0.56</td><td>61.15±0.16</td><td>(T)</td></tr><tr><td>GCNJK-SAINT-Node</td><td>72.80±0.43</td><td>63.68±0.06</td><td>44.30±0.22</td><td>26.97±0.10</td><td>80.96±0.09</td><td>59.50±0.35</td><td>44.86±0.19</td></tr><tr><td>GCNJK-SAINT-RW</td><td>72.29±0.49</td><td>65.00±0.11</td><td>47.40±0.17</td><td>33.05±0.06</td><td>81.04±0.14</td><td>59.82±0.27</td><td>47.39±0.19</td></tr><tr><td>MixHop-Cluster</td><td>75.79±0.44</td><td>76.67±0.07</td><td>48.41±0.31</td><td>46.82±0.11</td><td>81.12±0.10</td><td>62.95±0.08</td><td>(T)</td></tr><tr><td>MixHop-SAINT-Node</td><td>75.61±0.55</td><td>66.42±0.06</td><td>44.84±0.18</td><td>27.45±0.11</td><td>81.06±0.08</td><td>59.58±0.27</td><td>47.39±0.18</td></tr><tr><td>MixHop-SAINT-RW</td><td>76.38±0.50</td><td>67.92±0.06</td><td>50.55±0.20</td><td>34.21±0.07</td><td>82.25±0.78</td><td>60.39±0.16</td><td>49.15±0.26</td></tr><tr><td>LINKXMinibatch</td><td>84.50±0.65</td><td>81.27±0.38</td><td>53.74±0.27</td><td>60.27±0.29</td><td>85.81±0.10</td><td>65.84±0.19</td><td>59.80±0.41</td></tr></table>
187
+
188
+ Our experimental results for minibatched methods on our proposed datasets are in Table 4. Since GraphSAINT does not partition the nodes of the graph into subgraphs that cover all nodes, we test on the full input graph for the smaller datasets and uniformly random partitions of the graph into 10 induced subgraphs for the larger datasets.
189
+
190
+ First, we note that both Cluster-GCN and GraphSAINT sampling lead to performance degradation for these methods on our proposed non-homophilous datasets. When compared to the full-batch training results of the previous section, classification accuracy is typically substantially lower. Further experiments in Appendix C.2 give evidence that the performance degradation is often more substantial in non-homophilous settings, and provides possible explanations for why this may be the case.
191
+
192
+ On the other hand, LINKX does not suffer much performance degradation with the simple i.i.d. node minibatching technique. In fact, it matches or outperforms all methods in this setting, often by a wide margin. Though LINK performs on par with LINKX in arXiv-year and pokec, our LINKX model significantly outperforms it on other datasets, again due to LINKX’s ability to integrate node feature information. We again stress that the LINKX minibatching is very simple to implement, yet it still substantially outperforms other methods. Consequently, LINKX is generally well-suited for scalable node classification across a broad range of non-homophilous settings, surpassing even specially designed non-homophilous GNNs with current graph minibatching techniques.
193
+
194
+ # 6 Discussion and Conclusion
195
+
196
+ In this paper, we propose new, high-quality non-homophilous graph learning datasets, and we benchmark simple baselines and representative graph representation learning methods across these datasets. Further, we develop LINKX: a strong, simple, and scalable method for non-homophilous classification. Our experiments show that LINKX significantly outperforms other methods on our proposed datasets, thus providing one powerful method in the underexplored area of scalable learning on non-homophilous graphs. We hope that our contributions will provide researchers with new avenues of research in learning on non-homophilous graphs, along with better tools to test models and evaluate utility of new techniques.
197
+
198
+ While we do find utility in our proposed datasets and LINKX model, this work is somewhat limited by only focusing on transductive node classification. This setting is the most natural for studying performance in the absence of homophily, since here we define homophily in terms of the node labels, and previous non-homophilous GNN work using the Pei et al. [58] data also studies this setting exclusively $\scriptstyle [ 8 2 , \boxed { 1 7 } ]$ . Using other Facebook 100 datasets besides Penn94 $[ [ 6 7 ] ]$ would allow for inductive node classification, but LINKX does not directly generalize to this setting. Our proposed datasets and model LINKX could be used for link prediction, but this is left for future work.
199
+
200
+ Broader Impact. Fundamental research in graph learning on non-homophilous graphs has the potential for positive societal benefit. As a major application, it enables malicious node detection techniques in social networks and transaction networks that are not fooled by fraudsters’ connections to legitimate users and customers. This is a widely studied task, and past works have noted that non-homophilous structures are present in many such networks [11, 25, 55]. We hope that this paper provides insight on the homophily limitations of existing scalable graph learning models and help researchers design scalable models that continue to work well in the non-homophilous regime, thus improving the quality of node classification on graphs more broadly. As our proposed datasets have diverse structures and our model performs well across all of these datasets, the potential for future application of our work to important non-homophilous tasks is high.
201
+
202
+ Nevertheless, our work could also have potential for different types of negative social consequences. Nefarious behavior by key actors could be one source of such consequences. Nonetheless, we expect that the actors that can make use of large-scale social networks for gender prediction as studied in our work are limited in number. Actors with both the capability and incentive to perform such operations probably mostly consist of entities with access to large social network data such as social media companies or government actors with auxiliary networks $\mathbb { \boldsymbol { 0 } }$ . Smaller actors can perform certain attacks, but this may be made more difficult by resource requirements such as the need for certain external information $\pmb { \Vert 5 0 \Vert }$ or the ability to add nodes and edges before an anonymized version of a social network is released $\pmb { \Vert 5 \Vert }$ . Furthermore, additional actors could make use of deanonymization attacks $\textcircled { 1 3 0 } , \textcircled { 4 9 } , \textcircled { 5 0 } $ to reveal user identities in supposedly anonymized datasets.
203
+
204
+ Also, accidental consequences and implicit biases are a potential issue, even if the applications of the learning algorithms are benign and intended to benefit society [47]. Performance of algorithms may vary substantially between intersectional subgroups of subjects — as in the case of vision-based gender predictors $\dot { \mathbb { E } } \dot { 2 } \mathbb { I }$ (and some have questioned the propriety of vision-based gender classifiers altogether). Thus, there may be disparate effects on different populations, so care should be taken to understand the impact of those differences across subgroups. Moreover, large datasets require computing resources, so projects can only be pursued by large entities at the possible expense of the individual and smaller research groups $\boxed { 1 8 }$ . This is alleviated by the fact that our experiments are each run on one GPU, and hence have significantly less GPU computing requirements than much current deep learning research. Thus, smaller research groups and independent researchers should find our work beneficial, and should be able to build on it.
205
+
206
+ Finally, the nature of collection of online user information also comes with notable ethical concerns. Common notice-and-consent policies are often ineffective in actually protecting user privacy [52]. Indeed, users may not actually have much choice in using certain platforms or sharing data due to social or economic reasons. Also, users are generally unable to fully read and understand all of the different privacy policies that they come across, and may not understand the implications of having their data available for long periods of time to entities with powerful inference algorithms. Furthermore, people may rely on obscurity for privacy $\mathbb { \left[ \left[ 2 9 \right] \right] }$ , but this assumption may be ignored in courts of law, and it may be directly broken when data leaks or is released in aggregated form without sufficient privacy protections. Overall, while we believe that our work will benefit machine learning research and enable positive applications, we must still be aware of possible negative consequences.
207
+
208
+ # Acknowledgements
209
+
210
+ We thank Abhay Singh, Austin Benson, and Horace He for insightful discussions. We also thank the rest of Cornell University Artificial Intelligence for their support and discussion. We thank Facebook AI for funding equipment that made this work possible.
211
+
212
+ # References
213
+
214
+ [1] Sami Abu-El-Haija, Bryan Perozzi, Amol Kapoor, Nazanin Alipourfard, Kristina Lerman, Hrayr Harutyunyan, Greg Ver Steeg, and Aram Galstyan. Mixhop: Higher-order graph convolutional architectures via sparsified neighborhood mixing. In International Conference on Machine Learning, pages 21–29, 2019.
215
+ [2] Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. Optuna: A next-generation hyperparameter optimization framework, 2019.
216
+ [3] Kristen M Altenburger and Johan Ugander. Monophily in social networks introduces similarity among friends-of-friends. Nature human behaviour, 2(4):284–290, 2018.
217
+ [4] Kristen M Altenburger and Johan Ugander. Node attribute prediction: An evaluation of withinversus across-network tasks. NeurIPS Workshop on Relational Representation Learning, 2018.
218
+ [5] Lars Backstrom, Cynthia Dwork, and Jon M. Kleinberg. Wherefore art thou $\mathrm { r } 3 5 7 9 \mathrm { x } ?$ : anonymized social networks, hidden patterns, and structural steganography. In WWW $^ { , } 0 7$ , 2007.
219
+ [6] Muhammet Balcilar, Guillaume Renton, Pierre Héroux, Benoit Gaüzère, Sébastien Adam, and Paul Honeine. Analyzing the expressive power of graph neural networks in a spectral perspective. In International Conference on Learning Representations, 2021.
220
+ [7] Peter W Battaglia, Jessica B Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, et al. Relational inductive biases, deep learning, and graph networks. arXiv preprint arXiv:1806.01261, 2018.
221
+ [8] Abeba Birhane, Pratyusha Kalluri, Dallas Card, William Agnew, Ravit Dotan, and Michelle Bao. The values encoded in machine learning research, 2021.
222
+ [9] Deyu Bo, Xiao Wang, Chuan Shi, and Huawei Shen. Beyond low-frequency information in graph convolutional networks. arXiv preprint arXiv:2101.00797, 2021.
223
+ [10] Aleksandar Bojchevski, Johannes Klicpera, Bryan Perozzi, Amol Kapoor, Martin Blais, Benedek Rózemberczki, Michal Lukasik, and Stephan Günnemann. Scaling graph neural networks with approximate pagerank. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 2464–2473, 2020.
224
+ [11] Adam Breuer, Roee Eilat, and Udi Weinsberg. Friend or faux: Graph-based early detection of fake accounts on social networks, 2020.
225
+ [12] Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. In Sorelle A. Friedler and Christo Wilson, editors, Proceedings of the 1st Conference on Fairness, Accountability and Transparency, volume 81 of Proceedings of Machine Learning Research, pages 77–91. PMLR, 23–24 Feb 2018. URL https:// proceedings.mlr.press/v81/buolamwini18a.html.
226
+ [13] Duen Horng Chau, Shashank Pandit, and Christos Faloutsos. Detecting fraudulent personalities in networks of online auctioneers. In European conference on principles of data mining and knowledge discovery, pages 103–114. Springer, 2006.
227
+ [14] Jie Chen, Tengfei Ma, and Cao Xiao. Fastgcn: Fast learning with graph convolutional networks via importance sampling. In International Conference on Learning Representations, 2018.
228
+
229
+ [15] Ming Chen, Zhewei Wei, Zengfeng Huang, Bolin Ding, and Yaliang Li. Simple and deep graph convolutional networks. In International Conference on Machine Learning, pages 1725–1735. PMLR, 2020.
230
+
231
+ [16] Wei-Lin Chiang, Xuanqing Liu, Si Si, Yang Li, Samy Bengio, and Cho-Jui Hsieh. Cluster-gcn: An efficient algorithm for training deep and large graph convolutional networks. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 257–266, 2019.
232
+
233
+ [17] Eli Chien, Jianhao Peng, Pan Li, and Olgica Milenkovic. Adaptive universal generalized pagerank graph neural network. In International Conference on Learning Representations, 2021.
234
+
235
+ [18] Alex Chin, Yatong Chen, Kristen M. Altenburger, and Johan Ugander. Decoupled smoothing on graphs. In The World Wide Web Conference, pages 263–272, 2019.
236
+
237
+ [19] Djork-Arné Clevert, Thomas Unterthiner, and Sepp Hochreiter. Fast and accurate deep network learning by exponential linear units (elus). arXiv preprint arXiv:1511.07289, 2015.
238
+
239
+ [20] Chenhui Deng, Zhiqiang Zhao, Yongyu Wang, Zhiru Zhang, and Zhuo Feng. Graphzoom: A multi-level spectral approach for accurate and scalable graph embedding. In International Conference on Learning Representations, 2020.
240
+
241
+ [21] Vijay Prakash Dwivedi, Chaitanya K Joshi, Thomas Laurent, Yoshua Bengio, and Xavier Bresson. Benchmarking graph neural networks. arXiv preprint arXiv:2003.00982, 2020.
242
+
243
+ [22] Paul Erdos and Alfréd Rényi. On the evolution of random graphs. ˝ Publ. Math. Inst. Hung. Acad. Sci, 5(1):17–60, 1960.
244
+
245
+ [23] Matthias Fey and Jan Eric Lenssen. Fast graph representation learning with pytorch geometric. arXiv preprint arXiv:1903.02428, 2019.
246
+
247
+ [24] Konrad Gadzicki, Razieh Khamsehashari, and Christoph Zetzsche. Early vs late fusion in multimodal convolutional neural networks. In 2020 IEEE 23rd International Conference on Information Fusion (FUSION), pages 1–6. IEEE, 2020.
248
+
249
+ [25] Wolfgang Gatterbauer. Semi-supervised learning with heterophily. arXiv preprint arXiv:1412.3100, 2014.
250
+
251
+ [26] Ian Goodfellow, Yoshua Bengio, Aaron Courville, and Yoshua Bengio. Deep learning. MIT press Cambridge, 2016.
252
+
253
+ [27] Bronwyn H Hall, Adam B Jaffe, and Manuel Trajtenberg. The nber patent citation data file: Lessons, insights and methodological tools. Working Paper 8498, National Bureau of Economic Research, October 2001. URL http://www.nber.org/papers/w8498.
254
+
255
+ [28] Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. In Advances in neural information processing systems, pages 1024–1034, 2017.
256
+
257
+ [29] Woodrow Hartzog and Frederic Stutzman. The case for online obscurity. California Law Review, 101(1):1–49, 2013. ISSN 00081221. URL http://www.jstor.org/stable/23409387.
258
+
259
+ [30] Michael Hay, Gerome Miklau, David Jensen, Donald Towsley, and Philipp Weis. Resisting structural identification in anonymized social networks. PVLDB, 1:102–114, 08 2008. doi: 10.14778/1453856.1453873.
260
+
261
+ [31] Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and Jure Leskovec. Open graph benchmark: Datasets for machine learning on graphs. arXiv preprint arXiv:2005.00687, 2020.
262
+
263
+ [32] Qian Huang, Horace He, Abhay Singh, Ser-Nam Lim, and Austin Benson. Combining label propagation and simple models out-performs graph neural networks. In International Conference on Learning Representations, 2021.
264
+
265
+ [33] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448–456. PMLR, 2015.
266
+ [34] Junteng Jia and Austion R Benson. Residual correlation in graph neural network regression. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 588–598, 2020.
267
+ [35] Wei Jin, Tyler Derr, Yiqi Wang, Yao Ma, Zitao Liu, and Jiliang Tang. Node similarity preserving graph convolutional networks. In Proceedings of the 14th ACM International Conference on Web Search and Data Mining, pages 148–156, 2021.
268
+ [36] Dongkwan Kim and Alice Oh. How to find your friendly neighborhood: Graph attention design with self-supervision. In International Conference on Learning Representations, 2021.
269
+ [37] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.
270
+ [38] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks, 2017.
271
+ [39] Johannes Klicpera, Aleksandar Bojchevski, and Stephan Günnemann. Predict then propagate: Graph neural networks meet personalized pagerank. In International Conference on Learning Representations, 2019.
272
+ [40] M. Zabovsky L. Takac. Data analysis in public social networks. International Scientific Conference & International Workshop Present Day Trends of Innovations, 2012.
273
+ [41] Jure Leskovec and Andrej Krevl. SNAP Datasets: Stanford large network dataset collection. http://snap.stanford.edu/data, June 2014.
274
+ [42] Jure Leskovec, Jon Kleinberg, and Christos Faloutsos. Graphs over time: densification laws, shrinking diameters and possible explanations. In Proceedings of the eleventh ACM SIGKDD international conference on Knowledge discovery in data mining, pages 177–187, 2005.
275
+ [43] Derek Lim and Austin R Benson. Expertise and dynamics within crowdsourced musical knowledge curation: A case study of the genius platform. In Proceedings of the International AAAI Conference on Web and Social Media, 2021.
276
+ [44] Meng Liu, Zhengyang Wang, and Shuiwang Ji. Non-local graph neural networks, 2020.
277
+ [45] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018.
278
+ [46] Miller McPherson, Lynn Smith-Lovin, and James M Cook. Birds of a feather: Homophily in social networks. Annual review of sociology, 27(1):415–444, 2001.
279
+ [47] Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. A survey on bias and fairness in machine learning, 2019.
280
+ [48] Tom Mitchell et al. The 4 universities data set, 1997. URL http://www.cs.cmu.edu/afs/ cs.cmu.edu/project/theo-20/www/data/.
281
+ [49] Arvind Narayanan and Vitaly Shmatikov. Robust de-anonymization of large sparse datasets. In 2008 IEEE Symposium on Security and Privacy (sp 2008), pages 111–125, 2008. doi: 10.1109/SP.2008.33.
282
+ [50] Arvind Narayanan and Vitaly Shmatikov. De-anonymizing social networks. In 2009 30th IEEE Symposium on Security and Privacy, pages 173–187, 2009. doi: 10.1109/SP.2009.22.
283
+ [51] Mark EJ Newman. Mixing patterns in networks. Physical review E, 67(2):026126, 2003.
284
+ [52] Helen Nissenbaum. A contextual approach to privacy online. Daedalus, 140:32–48, 2011.
285
+
286
+ [53] Hoang Nt and Takanori Maehara. Revisiting graph neural networks: All we have is low-pass filters. arXiv preprint arXiv:1905.09550, 2019.
287
+
288
+ [54] Hoang NT, Takanori Maehara, and Tsuyoshi Murata. Stacked graph filter. arXiv preprint arXiv:2011.10988, 2020.
289
+
290
+ [55] Shashank Pandit, Duen Horng Chau, Samuel Wang, and Christos Faloutsos. Netprobe: a fast and scalable system for fraud detection in online auction networks. In Proceedings of the 16th international conference on World Wide Web, pages 201–210, 2007.
291
+
292
+ [56] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems, 32:8026–8037, 2019.
293
+
294
+ [57] Leto Peel. Graph-based semi-supervised learning for relational networks. In Proceedings of the 2017 SIAM International Conference on Data Mining, pages 435–443. SIAM, 2017.
295
+
296
+ [58] Hongbin Pei, Bingzhe Wei, Kevin Chen-Chuan Chang, Yu Lei, and Bo Yang. Geom-gcn: Geometric graph convolutional networks. In International Conference on Learning Representations, 2019.
297
+
298
+ [59] Jeffrey Pennington, Richard Socher, and Christopher Manning. GloVe: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1532–1543, Doha, Qatar, October 2014. Association for Computational Linguistics. doi: 10.3115/v1/D14-1162.
299
+
300
+ [60] Benedek Rozemberczki and Rik Sarkar. Twitch gamers: a dataset for evaluating proximity preserving and structural role-based node embeddings. arXiv preprint arXiv:2101.03091, 2021.
301
+
302
+ [61] Benedek Rozemberczki, Carl Allen, and Rik Sarkar. Multi-scale attributed node embedding. Journal of Complex Networks, 9(2):cnab014, 2021.
303
+
304
+ [62] Prithviraj Sen, Galileo Namata, Mustafa Bilgic, Lise Getoor, Brian Galligher, and Tina EliassiRad. Collective classification in network data. AI magazine, 29(3):93–93, 2008.
305
+
306
+ [63] Oleksandr Shchur, Maximilian Mumme, Aleksandar Bojchevski, and Stephan Günnemann. Pitfalls of graph neural network evaluation. arXiv preprint arXiv:1811.05868, 2018.
307
+
308
+ [64] Neil JA Sloane. The on-line encyclopedia of integer sequences. In Towards mechanized mathematical assistants, pages 130–130. Springer, 2007.
309
+
310
+ [65] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929–1958, 2014.
311
+
312
+ [66] Jie Tang, Jimeng Sun, Chi Wang, and Zi Yang. Social influence analysis in large-scale networks. In Proceedings of the 15th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 807–816, 2009.
313
+
314
+ [67] Amanda L Traud, Peter J Mucha, and Mason A Porter. Social structure of facebook networks. Physica A: Statistical Mechanics and its Applications, 391(16):4165–4180, 2012.
315
+
316
+ [68] Nate Veldt, Austin R Benson, and Jon Kleinberg. Higher-order homophily is combinatorially impossible. arXiv preprint arXiv:2103.11818, 2021.
317
+
318
+ [69] Petar Velickovi ˇ c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua ´ Bengio. Graph attention networks. In International Conference on Learning Representations, 2018.
319
+
320
+ [70] Kuansan Wang, Zhihong Shen, Chiyuan Huang, Chieh-Han Wu, Yuxiao Dong, and Anshul Kanakia. Microsoft Academic Graph: When experts are not enough. Quantitative Science Studies, 1(1):396–413, 02 2020. ISSN 2641-3337. doi: 10.1162/qss_a_00021. URL https: //doi.org/10.1162/qss_a_00021.
321
+
322
+ [71] Felix Wu, Amauri Souza, Tianyi Zhang, Christopher Fifty, Tao Yu, and Kilian Weinberger. Simplifying graph convolutional networks. In International conference on machine learning, pages 6861–6871. PMLR, 2019.
323
+ [72] Keyulu Xu, Chengtao Li, Yonglong Tian, Tomohiro Sonobe, Ken-ichi Kawarabayashi, and Stefanie Jegelka. Representation learning on graphs with jumping knowledge networks. In International Conference on Machine Learning, pages 5453–5462. PMLR, 2018.
324
+ [73] Yujun Yan, Milad Hashemi, Kevin Swersky, Yaoqing Yang, and Danai Koutra. Two sides of the same coin: Heterophily and oversmoothing in graph convolutional neural networks. arXiv preprint arXiv:2102.06462, 2021.
325
+ [74] Zhilin Yang, William Cohen, and Ruslan Salakhudinov. Revisiting semi-supervised learning with graph embeddings. In International conference on machine learning, pages 40–48. PMLR, 2016.
326
+ [75] Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L Hamilton, and Jure Leskovec. Graph convolutional neural networks for web-scale recommender systems. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 974–983, 2018.
327
+ [76] Hanqing Zeng, Hongkuan Zhou, Ajitesh Srivastava, Rajgopal Kannan, and Viktor Prasanna. Accurate, efficient and scalable graph embedding. In 2019 IEEE International Parallel and Distributed Processing Symposium (IPDPS), pages 462–471. IEEE, 2019.
328
+ [77] Hanqing Zeng, Hongkuan Zhou, Ajitesh Srivastava, Rajgopal Kannan, and Viktor Prasanna. Graphsaint: Graph sampling based inductive learning method. In International Conference on Learning Representations, 2019.
329
+ [78] Jin Zeng, Yanfeng Tong, Yunmu Huang, Qiong Yan, Wenxiu Sun, Jing Chen, and Yongtian Wang. Deep surface normal estimation with hierarchical rgb-d fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6153–6162, 2019.
330
+ [79] Elena Zheleva and Lise Getoor. To join or not to join: the illusion of privacy in social networks with mixed public and private user profiles. In Proceedings of the 18th international conference on World wide web, pages 531–540, 2009.
331
+ [80] Dengyong Zhou, Olivier Bousquet, Thomas N Lal, Jason Weston, and Bernhard Schölkopf. Learning with local and global consistency. In Advances in neural information processing systems, pages 321–328, 2004.
332
+ [81] Jiong Zhu, Ryan A Rossi, Anup Rao, Tung Mai, Nedim Lipka, Nesreen K Ahmed, and Danai Koutra. Graph neural networks with heterophily. arXiv preprint arXiv:2009.13566, 2020.
333
+ [82] Jiong Zhu, Yujun Yan, Lingxiao Zhao, Mark Heimann, Leman Akoglu, and Danai Koutra. Beyond homophily in graph neural networks: Current limitations and effective designs. Advances in Neural Information Processing Systems, 33, 2020.
334
+ [83] Michael Zimmer. Facebook data of 1.2 million users from 2005 released: Limited exposure, but very problematic, Feb 2011. URL https://michaelzimmer.org/2011/02/15/ facebook-data-of-1-2-million-users-from-2005-released/.
335
+
336
+ # Checklist
337
+
338
+ 1. For all authors...
339
+
340
+ (a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes] We have supported the claims in the main paper and appendix.
341
+ (b) Did you describe the limitations of your work? [Yes] See discussion / conclusion Section $6 .$
342
+ (c) Did you discuss any potential negative societal impacts of your work? [Yes] See discussion / conclusion Section 6.
343
+ (d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes]
344
+
345
+ 2. If you are including theoretical results...
346
+
347
+ (a) Did you state the full set of assumptions of all theoretical results? [N/A] (b) Did you include complete proofs of all theoretical results? [N/A]
348
+
349
+ 3. If you ran experiments...
350
+
351
+ (a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] We include our proposed datasets and our codes for reproducing the experimental results in the supplemental material.
352
+ (b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] See Section 5.1 and Appendix Section B.
353
+ (c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [Yes] We include standard deviation of performance metrics across multiple runs in Section 5.
354
+ (d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] We provide GPU information in Appendix Section B.
355
+
356
+ 4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...
357
+
358
+ (a) If your work uses existing assets, did you cite the creators? [Yes] We cite the creators of existing datasets and methods in the main paper. Also, in Appendix B, we cite codes that we used.
359
+ (b) Did you mention the license of the assets? [Yes] In Appendix D, we note the licenses of datasets, if provided in published work. Also, we include the license of open-source codes we build off of in Appendix B.
360
+ (c) Did you include any new assets either in the supplemental material or as a URL? [Yes] We include our proposed datasets along with codes for reproducing our results in the supplemental material.
361
+ (d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [Yes] We discuss this in Appendix D. Most of the datasets were constructed and presented in previously published academic work. This mostly does not apply to the wiki dataset that we collect, as it is encyclopedic in nature.
362
+ (e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [Yes] In Appendix D, we discuss this. Our datasets are primarily numerical, so they do not contain offensive content. To the best of our knowledge, our datasets do not contain personally identifiable content.
363
+
364
+ 5. If you used crowdsourcing or conducted research with human subjects...
365
+
366
+ (a) Did you include the full text of instructions given to participants and screenshots, if applicable? [N/A]
367
+ (b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A]
368
+ (c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [N/A]
parse/train/DfGu8WwT0d/DfGu8WwT0d_content_list.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/DfGu8WwT0d/DfGu8WwT0d_middle.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/DfGu8WwT0d/DfGu8WwT0d_model.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/Hk4_qw5xe/Hk4_qw5xe.md ADDED
@@ -0,0 +1,505 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # TOWARDS PRINCIPLED METHODS FOR TRAINING GENERATIVE ADVERSARIAL NETWORKS
2
+
3
+ Martin Arjovsky Courant Institute of Mathematical Sciences martinarjovsky@gmail.com
4
+
5
+ Leon Bottou´ Facebook AI Research leonb@fb.com
6
+
7
+ # ABSTRACT
8
+
9
+ The goal of this paper is not to introduce a single algorithm or method, but to make theoretical steps towards fully understanding the training dynamics of generative adversarial networks. In order to substantiate our theoretical analysis, we perform targeted experiments to verify our assumptions, illustrate our claims, and quantify the phenomena. This paper is divided into three sections. The first section introduces the problem at hand. The second section is dedicated to studying and proving rigorously the problems including instability and saturation that arize when training generative adversarial networks. The third section examines a practical and theoretically grounded direction towards solving these problems, while introducing new tools to study them.
10
+
11
+ # 1 INTRODUCTION
12
+
13
+ Generative adversarial networks (GANs)(Goodfellow et al., 2014a) have achieved great success at generating realistic and sharp looking images. However, they are widely general methods, now starting to be applied to several other important problems, such as semisupervised learning, stabilizing sequence learning methods for speech and language, and 3D modelling. (Denton et al., 2015; Radford et al., 2015; Salimans et al., 2016; Lamb et al., 2016; Wu et al., 2016)
14
+
15
+ However, they still remain remarkably difficult to train, with most current papers dedicated to heuristically finding stable architectures. (Radford et al., 2015; Salimans et al., 2016)
16
+
17
+ Despite their success, there is little to no theory explaining the unstable behaviour of GAN training. Furthermore, approaches to attacking this problem still rely on heuristics that are extremely sensitive to modifications. This makes it extremely hard to experiment with new variants, or to use them in new domains, which limits their applicability drastically. This paper aims to change that, by providing a solid understanding of these issues, and creating principled research directions towards adressing them.
18
+
19
+ It is interesting to note that the architecture of the generator used by GANs doesn’t differ significantly from other approaches like variational autoencoders (Kingma & Welling, 2013). After all, at the core of it we first sample from a simple prior $z \sim p ( z )$ , and then output our final sample $g _ { \boldsymbol { \theta } } ( z )$ , sometimes adding noise in the end. Always, $g _ { \theta }$ is a neural network parameterized by $\theta$ , and the main difference is how $g _ { \theta }$ is trained.
20
+
21
+ Traditional approaches to generative modeling relied on maximizing likelihood, or equivalently minimizing the Kullback-Leibler (KL) divergence between our unknown data distribution $\mathbb { P } _ { r }$ and our generator’s distribution $\mathbb { P } _ { g }$ (that depends of course on $\theta$ ). If we assume that both distributions are continuous with densities $P _ { r }$ and $P _ { g }$ , then these methods try to minimize
22
+
23
+ $$
24
+ K L ( \mathbb { P } _ { r } \| \mathbb { P } _ { g } ) = \int _ { \mathcal { X } } P _ { r } ( x ) \log { \frac { P _ { r } ( x ) } { P _ { g } ( x ) } } \mathrm { d } x
25
+ $$
26
+
27
+ This cost function has the good property that it has a unique minimum at $\mathbb { P } _ { g } = \mathbb { P } _ { r }$ , and it doesn’t require knowledge of the unknown $P _ { r } ( x )$ to optimize it (only samples). However, it is interesting to see how this divergence is not symetrical between $\boldsymbol { \mathbb { P } } _ { r }$ and $\mathbb { P } _ { g }$ :
28
+
29
+ • If $P _ { r } ( x ) > P _ { g } ( x )$ , then $x$ is a point with higher probability of coming from the data than being a generated sample. This is the core of the phenomenon commonly described as ‘mode dropping’: when there are large regions with high values of $P _ { r }$ , but small or zero values in $P _ { g }$ . It is important to note that when $P _ { r } ( x ) > 0$ but $P _ { g } ( x ) \to 0$ , the integrand inside the KL grows quickly to infinity, meaning that this cost function assigns an extremely high cost to a generator’s distribution not covering parts of the data. If $P _ { r } ( x ) < P _ { g } ( x )$ , then $x$ has low probability of being a data point, but high probability of being generated by our model. This is the case when we see our generator outputting an image that doesn’t look real. In this case, when $P _ { r } ( x ) 0$ and $P _ { g } ( x ) > 0$ , we see that the value inside the KL goes to 0, meaning that this cost function will pay extremely low cost for generating fake looking samples.
30
+
31
+ Clearly, if we would minimize $K L ( \mathbb { P } _ { g } \| \mathbb { P } _ { r } )$ instead, the weighting of these errors would be reversed, meaning that this cost function would pay a high cost for generating not plausibly looking pictures. Generative adversarial networks have been shown to optimize (in its original formulation), the Jensen-shannon divergence, a symmetric middle ground to this two cost functions
32
+
33
+ $$
34
+ J S D ( \mathbb { P } _ { r } \| \mathbb { P } _ { g } ) = \frac { 1 } { 2 } K L ( \mathbb { P } _ { r } \| \mathbb { P } _ { A } ) + \frac { 1 } { 2 } K L ( \mathbb { P } _ { g } \| \mathbb { P } _ { A } )
35
+ $$
36
+
37
+ where $\mathbb { P } _ { A }$ is the ‘average’ distribution, with density Pr+Pg2 . An impressive experimental analysis of the similarities, uses and differences of these divergences in practice can be seen at Theis et al. (2016). It is indeed conjectured that the reason of GANs success at producing reallistically looking images is due to the switch from the traditional maximum likelihood approaches. (Theis et al., 2016; Huszar, 2015). However, the problem is far from closed.
38
+
39
+ Generative adversarial networks are formulated in two steps. We first train a discriminator $D$ to maximize
40
+
41
+ $$
42
+ L ( D , g _ { \theta } ) = \mathbb { E } _ { x \sim \mathbb { P } _ { r } } [ \log D ( x ) ] + \mathbb { E } _ { x \sim \mathbb { P } _ { g } } [ \log ( 1 - D ( x ) ) ]
43
+ $$
44
+
45
+ One can show easily that the optimal discriminator has the shape
46
+
47
+ $$
48
+ D ^ { * } ( x ) = \frac { P _ { r } ( x ) } { P _ { r } ( x ) + P _ { g } ( x ) }
49
+ $$
50
+
51
+ and that $L ( D ^ { * } , g _ { \theta } ) = 2 J S D ( \mathbb { P } _ { r } \| \mathbb { P } _ { g } ) - 2 \log 2$ , so minimizing equation (1) as a function of $\theta$ yields minimizing the Jensen-Shannon divergence when the discriminator is optimal. In theory, one would expect therefore that we would first train the discriminator as close as we can to optimality (so the cost function on $\theta$ better approximates the $J S D$ ), and then do gradient steps on $\theta$ , alternating these two things. However, this doesn’t work. In practice, as the discriminator gets better, the updates to the generator get consistently worse. The original GAN paper argued that this issue arose from saturation, and switched to another similar cost function that doesn’t have this problem. However, even with this new cost function, updates tend to get worse and optimization gets massively unstable. Therefore, several questions arize:
52
+
53
+ • Why do updates get worse as the discriminator gets better? Both in the original and the new cost function.
54
+ • Why is GAN training massively unstable?
55
+ • Is the new cost function following a similar divergence to the JSD? If so, what are its properties?
56
+ • Is there a way to avoid some of these issues?
57
+
58
+ The fundamental contributions of this paper are the answer to all these questions, and perhaps more importantly, to introduce the tools to analyze them properly. We provide a new direction designed to avoid the instability issues in GANs, and examine in depth the theory behind it. Finally, we state a series of open questions and problems, that determine several new directions of research that begin with our methods.
59
+
60
+ # 2 SOURCES OF INSTABILITY
61
+
62
+ The theory tells us that the trained discriminator will have cost at most $2 \log 2 - 2 J S D ( \mathbb { P } _ { r } \Vert \mathbb { P } _ { g } )$ . However, in practice, if we just train $D$ till convergence, its error will go to 0, as observed in Figure 1, pointing to the fact that the $J S D$ between them is maxed out. The only way this can happen is if the distributions are not continuous1, or they have disjoint supports.
63
+
64
+ One possible cause for the distributions not to be continuous is if their supports lie on low dimensional manifolds. There is strong empirical and theoretical evidence to believe that $\mathbb { P } _ { r }$ is indeed extremely concentrated on a low dimensional manifold (Narayanan & Mitter, 2010). As of $\mathbb { P } _ { g }$ , we will prove soon that such is the case as well.
65
+
66
+ In the case of GANs, $\mathbb { P } _ { g }$ is defined via sampling from a simple prior $z \sim p ( z )$ , and then applying a function $g : { \mathcal { Z } } { \mathcal { X } }$ , so the support of $\mathbb { P } _ { g }$ has to be contained in $g ( \mathcal { Z } )$ . If the dimensionality of $\mathcal { Z }$ is less than the dimension of $\mathcal { X }$ (as is typically the case), then it’s imposible for $\mathbb { P } _ { g }$ to be continuous. This is because in most cases $g ( \mathcal { Z } )$ will be contained in a union of low dimensional manifolds, and therefore have measure 0 in $\mathcal { X }$ . Note that while intuitive, this is highly nontrivial, since having an $n$ - dimensional parameterization does absolutely not imply that the image will lie on an $n$ -dimensional manifold. In fact, there are many easy counterexamples, such as Peano curves, lemniscates, and many more. In order to show this for our case, we rely heavily on $g$ being a neural network, since we are able to leverage that $g$ is made by composing very well behaved functions. We now state this properly in the following Lemma:
67
+
68
+ Lemma 1. Let $g : { \mathcal { Z } } { \mathcal { X } }$ be a function composed by affine transformations and pointwise nonlinearities, which can either be rectifiers, leaky rectifiers, or smooth strictly increasing functions (such as the sigmoid, tanh, softplus, etc). Then, $g ( \mathcal { Z } )$ is contained in a countable union of manifolds of dimension at most $\operatorname { d i m } \mathcal { Z }$ . Therefore, if the dimension of $\mathcal { Z }$ is less than the one of $\mathcal { X }$ , $g ( \mathcal { Z } )$ will be $a$ set of measure $O$ in $\mathcal { X }$ .
69
+
70
+ Proof. See Appendix A.
71
+
72
+ Driven by this, this section shows that if the supports of $\mathbb { P } _ { r }$ and $\mathbb { P } _ { g }$ are disjoint or lie in low dimensional manifolds, there is always a perfect discriminator between them, and we explain exactly how and why this leads to an unreliable training of the generator.
73
+
74
+ # 2.1 THE PERFECT DISCRIMINATION THEOREMS
75
+
76
+ For simplicity, and to introduce the methods, we will first explain the case where $\mathbb { P } _ { r }$ and $\mathbb { P } _ { g }$ have disjoint supports. We say that a discriminator $D : \mathcal { X } [ 0 , 1 ]$ has accuracy 1 if it takes the value 1 on a set that contains the support of $\mathbb { P } _ { r }$ and value 0 on a set that contains the support of $\mathbb { P } _ { g }$ . Namely, $\mathbb { P } _ { r } [ D ( x ) = 1 ] = 1$ and $\mathbb { P } _ { g } [ D ( x ) = 0 ] = 1$ .
77
+
78
+ Theorem 2.1. If two distributions $\mathbb { P } _ { r }$ and $\mathbb { P } _ { g }$ have support contained on two disjoint compact subsets $\mathcal { M }$ and $\mathcal { P }$ respectively, then there is a smooth optimal discrimator $D ^ { * } : \mathcal { X } [ 0 , 1 ]$ that has accuracy $I$ and $\nabla _ { x } D ^ { * } ( x ) = 0$ for all $x \in \mathcal { M } \cup \mathcal { P }$ .
79
+
80
+ Proof. The discriminator is trained to maximize
81
+
82
+ $$
83
+ \mathbb { E } _ { x \sim \mathbb { P } _ { r } } [ \log D ( x ) ] + \mathbb { E } _ { x \sim \mathbb { P } _ { g } } [ \log ( 1 - D ( x ) ) ]
84
+ $$
85
+
86
+ Since $\mathcal { M }$ and $\mathcal { P }$ are compact and disjoint, $0 < \delta = d ( \mathcal { P } , \mathcal { M } )$ the distance between both sets. We now define
87
+
88
+ $$
89
+ \begin{array} { c } { { \hat { \mathcal { M } } = \{ x : d ( x , M ) \leq \delta / 3 \} } } \\ { { \hat { \mathcal { P } } = \{ x : d ( x , P ) \leq \delta / 3 \} } } \end{array}
90
+ $$
91
+
92
+ By definition of $\delta$ we have that $\hat { P }$ and $\hat { M }$ are clearly disjoint compact sets. Therefore, by Urysohn’s smooth lemma there exists a smooth function $D ^ { * } : \mathcal { X } [ 0 , 1 ]$ such that $D ^ { * } | _ { \hat { \mathcal { M } } } \equiv 1$ and $D ^ { * } | _ { \hat { \mathcal P } } \equiv 0$ . Since $\log D ^ { * } ( x ) = 0$ for all $x$ in the support of $\mathbb { P } _ { r }$ and $\bar { \log ( 1 - D ^ { * } ( x ) ) } = 0$ for all $x$ in the support of $\mathbb { P } _ { g }$ , the discriminator is completely optimal and has accuracy 1. Furthermore, let $x$ be in ${ \mathcal { M } } \cup { \mathcal { P } }$ . If we assume that $x \in \mathcal { M }$ , there is an open ball $B = B ( x , \delta / 3 )$ on which $D ^ { * } | _ { B }$ is constant. This shows that $\nabla _ { x } D ^ { * } ( x ) \equiv 0$ . Taking $x \in \mathcal { P }$ and working analogously we finish the proof. □
93
+
94
+ ![](images/8504017c82671e0f312984ea8e54505a4aa240e21955bf77b32917cb31edb1ef.jpg)
95
+ Figure 1: First, we trained a DCGAN for 1, 10 and 25 epochs. Then, with the generator fixed we train a discriminator from scratch. We see the error quickly going to 0, even with very few iterations on the discriminator. This even happens after 25 epochs of the DCGAN, when the samples are remarkably good and the supports are likely to intersect, pointing to the non-continuity of the distributions. Note the logarithmic scale. For illustration purposes we also show the accuracy of the discriminator, which goes to 1 in sometimes less than 50 iterations. This is 1 even for numerical precision, and the numbers are running averages, pointing towards even faster convergence.
96
+
97
+ In the next theorem, we take away the disjoint assumption, to make it general to the case of two different manifolds. However, if the two manifolds match perfectly on a big part of the space, no discriminator could separate them. Intuitively, the chances of two low dimensional manifolds having this property is rather dim: for two curves to match in space in a specific segment, they couldn’t be perturbed in any arbitrarilly small way and still satisfy this property. To do this, we will define the notion of two manifolds perfectly aligning, and show that this property never holds with probability 1 under any arbitrarilly small perturbations.
98
+
99
+ Definition 2.1. We first need to recall the definition of transversallity. Let $\mathcal { M }$ and $\mathcal { P }$ be two boundary free regular submanifolds of $\mathcal { F }$ , which in our cases will simply be $\mathcal { F } = \mathbb { R } ^ { d }$ . Let $x \in \mathcal { M } \cap \mathcal { P }$ be an intersection point of the two manifolds. We say that $\mathcal { M }$ and $\mathcal { P }$ intersect transversally in $x$ if $T _ { x } { \mathcal { M } } + T _ { x } { \mathcal { P } } = { \overline { { T } } } _ { x } { \mathcal { F } }$ , where $T _ { x } { \mathcal { M } }$ means the tangent space of $\mathcal { M }$ around $x$ .
100
+
101
+ Definition 2.2. We say that two manifolds without boundary $\mathcal { M }$ and $\mathcal { P }$ perfectly align if there is an $x \in \mathcal { M } \cap \mathcal { P }$ such that $\mathcal { M }$ and $\mathcal { P }$ don’t intersect transversally in $x$ .
102
+
103
+ We shall note the boundary and interior of a manifold $\mathcal { M }$ by $\partial M$ and Int $M$ respectively. We say that two manifolds $\mathcal { M }$ and $\mathcal { P }$ (with or without boundary) perfectly align if any of the boundary free manifold pairs $( \operatorname { I n t } \mathcal { M } , \operatorname { I n t } \mathcal { P } _ { }$ ), (I $\mathrm { n t } \mathcal { M } , \partial \mathcal { P } )$ , $( \partial \mathcal { M } , \mathrm { I n t } \mathcal { P } )$ or $( \partial \mathcal { M } , \partial \mathcal { P } )$ perfectly align.
104
+
105
+ The interesting thing is that we can safely assume in practice that any two manifolds never perfectly align. This can be done since an arbitrarilly small random perturbation on two manifolds will lead them to intersect transversally or don’t intersect at all. This is precisely stated and proven in Lemma 2.
106
+
107
+ As stated by Lemma 3, if two manifolds don’t perfectly align, their intersection $\mathcal { L } = \mathcal { M } \cap \mathcal { P }$ will be a finite union of manifolds with dimensions strictly lower than both the dimension of $\mathcal { M }$ and the one of $\mathcal { P }$ .
108
+
109
+ Lemma 2. Let $\mathcal { M }$ and $\mathcal { P }$ be two regular submanifolds of $\mathbb { R } ^ { d }$ that don’t have full dimension. Let $\eta , \eta ^ { \prime }$ be arbitrary independent continuous random variables. We therefore define the perturbed manifolds as $\tilde { \mathcal { M } } = \tilde { \mathcal { M } } + \eta$ and $\tilde { \mathcal P } = \mathcal P + \eta ^ { \prime }$ . Then
110
+
111
+ $$
112
+ \mathbb { P } _ { \eta , \eta ^ { \prime } } ( \tilde { \mathcal { M } } d o e s n o t p e r f e c t l y a l i g n w i t h \tilde { \mathcal { P } } ) = 1
113
+ $$
114
+
115
+ Proof. See Appendix A.
116
+
117
+ Lemma 3. Let $\mathcal { M }$ and $\mathcal { P }$ be two regular submanifolds of $\mathbb { R } ^ { d }$ that don’t perfectly align and don’t have full dimension. Let $\mathcal { L } = \mathcal { M } \cap \mathcal { P }$ . If $\mathcal { M }$ and $\mathcal { P }$ don’t have boundary, then $\mathcal { L }$ is also a manifold, and has strictly lower dimension than both the one of $\mathcal { M }$ and the one of $\mathcal { P }$ . If they have boundary, $\mathcal { L }$ is a union of at most 4 strictly lower dimensional manifolds. In both cases, $\mathcal { L }$ has measure 0 in both $\mathcal { M }$ and $\mathcal { P }$ .
118
+
119
+ Proof. See Appendix A.
120
+
121
+ We now state our perfect discrimination result for the case of two manifolds.
122
+
123
+ Theorem 2.2. Let $\mathbb { P } _ { r }$ and $\mathbb { P } _ { g }$ be two distributions that have support contained in two closed manifolds $\mathcal { M }$ and $\mathcal { P }$ that don’t perfectly align and don’t have full dimension. We further assume that $\mathbb { P } _ { r }$ and $\mathbb { P } _ { g }$ are continuous in their respective manifolds, meaning that if there is a set $A$ with measure $O$ in $\mathcal { M }$ , then $\mathbb { P } _ { r } ( A ) = 0$ (and analogously for $\mathbb { P } _ { g , \mathbb { \Lambda } }$ . Then, there exists an optimal discriminator $D ^ { * } : \mathcal { X } [ 0 , 1 ]$ that has accuracy $^ { l }$ and for almost any $x$ in $\mathcal { M }$ or $\mathcal { P }$ , $D ^ { * }$ is smooth in a neighbourhood of $x$ and $\nabla _ { x } D ^ { * } ( x ) = 0$ .
124
+
125
+ Proof. By Lemma 3 we know that $\mathcal { L } = \mathcal { M } \cap \mathcal { P }$ is strictly lower dimensional than both $\mathcal { M }$ and $\mathcal { P }$ , and has measure 0 on both of them. By continuity, $\mathbb { P } _ { r } ( \mathcal { L } ) = 0$ and $\mathbb { P } _ { g } ( \mathcal { L } ) = 0$ . Note that this implies the support of $\mathbb { P } _ { r }$ is contained in $\mathcal { M } \backslash \mathcal { L }$ and the support of $\mathbb { P } _ { g }$ is contained in $\mathcal { P } \backslash \mathcal { L }$ .
126
+
127
+ Let $x \in \mathcal { M } \backslash \mathcal { L }$ . Therefore, $x \in \mathcal { P } ^ { c }$ (the complement of $\mathcal { P }$ ) which is an open set, so there exists a ball of radius $\epsilon _ { x }$ such that $B ( x , \epsilon _ { x } ) \cap \mathcal { P } = \varnothing$ . This way, we define
128
+
129
+ $$
130
+ \hat { \mathcal { M } } = \bigcup _ { x \in \mathcal { M } \setminus \mathcal { L } } B ( x , \epsilon _ { x } / 3 )
131
+ $$
132
+
133
+ We define $\hat { \mathcal { P } }$ analogously. Note that by construction these are both open sets on $\mathbb { R } ^ { d }$ . Since $\mathcal { M } \backslash \mathcal { L } \subseteq$ $\hat { \mathcal { M } }$ , and $\mathcal { P } \setminus \mathcal { L } \subseteq \hat { \mathcal { P } }$ , the support of $\mathbb { P } _ { r }$ and $\mathbb { P } _ { g }$ is contained in $\hat { \mathcal { M } }$ and $\hat { \mathcal { P } }$ respectively. As well by construction, $\hat { \mathcal { M } } \cap \hat { \mathcal { P } } = \varnothing$ .
134
+
135
+ Let us define $D ^ { * } ( x ) = 1$ for all $x \in { \hat { \mathcal { M } } }$ , and 0 elsewhere (clearly including $\hat { \mathcal { P } }$ . Since $\log D ^ { * } ( x ) = 0$ for all $x$ in the support of $\mathbb { P } _ { r }$ and $\log ( 1 - D ^ { * } ( x ) ) = 0$ for all $x$ in the support of $\mathbb { P } _ { g }$ , the discriminator is completely optimal and has accuracy 1. Furthermore, let $x \in { \hat { \mathcal { M } } }$ . Since $\hat { \mathcal { M } }$ is an open set and $D ^ { * }$ is constant on $\hat { \mathcal { M } }$ , then $\nabla _ { x } D ^ { * } | _ { \hat { \mathcal { M } } } \equiv 0$ . Analogously, $\nabla _ { x } D ^ { * } | _ { \hat { \mathcal { P } } } \equiv 0$ . Therefore, the set of points where $D ^ { * }$ is non-smooth or has non-zero gradient inside $\mathcal { M } \cup \mathbf { \dot { \mathcal { P } } }$ is contained in $\mathcal { L }$ , which has null-measure in both manifolds, therefore concluding the theorem. □
136
+
137
+ These two theorems tell us that there are perfect discriminators which are smooth and constant almost everywhere in $\mathcal { M }$ and $\mathcal { P }$ . The fact that the discriminator is constant in both manifolds points to the fact that we won’t really be able to learn anything by backproping through it, as we shall see in the next subsection. To conclude this general statement, we state the following theorem on the divergences of $\mathbb { P } _ { r }$ and $\mathbb { P } _ { g }$ , whose proof is trivial and left as an exercise to the reader.
138
+
139
+ Theorem 2.3. Let $\mathbb { P } _ { r }$ and $\mathbb { P } _ { g }$ be two distributions whose support lies in two manifolds $\mathcal { M }$ and $\mathcal { P }$ that don’t have full dimension and don’t perfectly align. We further assume that $\mathbb { P } _ { r }$ and $\mathbb { P } _ { g }$ are continuous in their respective manifolds. Then,
140
+
141
+ $$
142
+ \begin{array} { r } { J S D ( \mathbb { P } _ { r } \| \mathbb { P } _ { g } ) = \log 2 } \\ { K L ( \mathbb { P } _ { r } \| \mathbb { P } _ { g } ) = + \infty } \\ { K L ( \mathbb { P } _ { g } \| \mathbb { P } _ { r } ) = + \infty } \end{array}
143
+ $$
144
+
145
+ Note that these divergences will be maxed out even if the two manifolds lie arbitrarilly close to each other. The samples of our generator might look impressively good, yet both KL divergences will be infinity. Therefore, Theorem 2.3 points us to the fact that attempting to use divergences out of the box to test similarities between the distributions we typically consider might be a terrible idea. Needless to say, if these divergencies are always maxed out attempting to minimize them by gradient descent isn’t really possible. We would like to have a perhaps softer measure, that incorporates a notion of distance between the points in the manifolds. We will come back to this topic later in section 3, where we explain an alternative metric and provide bounds on it that we are able to analyze and optimize.
146
+
147
+ # 2.2 THE CONSEQUENCES, AND THE PROBLEMS OF EACH COST FUNCTION
148
+
149
+ Theorems 2.1 and 2.2 showed one very important fact. If the two distributions we care about have supports that are disjoint or lie on low dimensional manifolds, the optimal discriminator will be perfect and its gradient will be zero almost everywhere.
150
+
151
+ # 2.2.1 THE ORIGINAL COST FUNCTION
152
+
153
+ We will now explore what happens when we pass gradients to the generator through a discriminator. One crucial difference with the typical analysis done so far is that we will develop the theory for an approximation to the optimal discriminator, instead of working with the (unknown) true discriminator. We will prove that as the approximaton gets better, either we see vanishing gradients or the massively unstable behaviour we see in practice, depending on which cost function we use.
154
+
155
+ In what follows, we denote by $\| D \|$ the norm
156
+
157
+ $$
158
+ \| D \| = \operatorname* { s u p } _ { x \in \mathcal { X } } | D ( x ) | + \| \nabla _ { x } D ( x ) \| _ { 2 }
159
+ $$
160
+
161
+ The use of this norm is to make the proofs simpler, but could have been done in another Sobolev norm $\| \cdot \| _ { 1 , p }$ for $p < \infty$ covered by the universal approximation theorem in the sense that we can guarantee a neural network approximation in this norm (Hornik, 1991).
162
+
163
+ Theorem 2.4 (Vanishing gradients on the generator). Let $g _ { \theta } : \mathcal { Z } \mathcal { X }$ be a differentiable function that induces a distribution $\mathbb { P } _ { g }$ . Let $\mathbb { P } _ { r }$ be the real data distribution. Let $D$ be a differentiable discriminator. If the conditions of Theorems 2.1 or 2.2 are satisfied, $\lVert D - D ^ { * } \rVert < \epsilon ,$ , and $\mathbb { E } _ { z \sim p ( z ) } \left[ \| J _ { \theta } g _ { \theta } ( z ) \| _ { 2 } ^ { 2 } \right] \le M ^ { 2 }$ , 2 then
164
+
165
+ $$
166
+ \| \nabla _ { \theta } \mathbb { E } _ { z \sim p ( z ) } [ \log ( 1 - D ( g _ { \theta } ( z ) ) ) ] \| _ { 2 } < M \frac { \epsilon } { 1 - \epsilon }
167
+ $$
168
+
169
+ Proof. In both proofs of Theorems 2.1 and 2.2 we showed that $D ^ { * }$ is locally 0 on the support of $\mathbb { P } _ { g }$ Then, using Jensen’s inequality and the chain rule on this support we have
170
+
171
+ $$
172
+ \begin{array} { r l } & { \| \nabla _ { \theta } \mathbb { E } _ { z \sim p ( z ) } [ \log ( 1 - D ( g \theta ( z ) ) ) ] \| _ { 2 } ^ { 2 } \le \mathbb { E } _ { \pi \sim p ( z ) } \left[ \frac { \| \nabla _ { \theta } D ( g \theta ( z ) ) \| _ { 2 } ^ { 2 } } { \| 1 - D ( g \theta ( z ) ) \| ^ { 2 } } \right] } \\ & { \qquad \le \mathbb { E } _ { \pi \sim p ( z ) } \left[ \frac { \| \nabla _ { x } D ( g \theta ( z ) ) \| _ { 2 } ^ { 2 } \| J _ { g } g \theta ( z ) \| _ { 2 } ^ { 2 } } { \| 1 - D ( g \theta ( z ) ) \| ^ { 2 } } \right] } \\ & { \qquad < \mathbb { E } _ { z \sim p ( z ) } \left[ \frac { \left( \| \nabla _ { x } D ^ { * } ( g _ { \theta } ( z ) ) \| _ { 2 } + \epsilon ^ { 2 } \right) ^ { 2 } \| J _ { g } g _ { \theta } ( z ) \| _ { 2 } ^ { 2 } } { \left( \| 1 - D ^ { * } ( g _ { \theta } ( z ) ) \| _ { - } \epsilon \right) ^ { 2 } } \right] } \\ & { \qquad = \mathbb { E } _ { z \sim p ( z ) } \left[ \frac { \epsilon ^ { 2 } \| J _ { g } g _ { \theta } ( z ) \| _ { 2 } ^ { 2 } } { \left( 1 - \epsilon \right) ^ { 2 } } \right] } \\ & { \qquad \le M ^ { 2 } \frac { \epsilon ^ { 2 } } { \left( 1 - \epsilon \right) ^ { 2 } } } \end{array}
173
+ $$
174
+
175
+ Taking square root of everything we get
176
+
177
+ $$
178
+ \| \nabla _ { \theta } \mathbb { E } _ { z \sim p ( z ) } [ \log ( 1 - D ( g _ { \theta } ( z ) ) ) ] \| _ { 2 } < M \frac { \epsilon } { 1 - \epsilon }
179
+ $$
180
+
181
+ finishing the proof
182
+
183
+ Corollary 2.1. Under the same assumptions of Theorem 2.4
184
+
185
+ $$
186
+ \operatorname* { l i m } _ { \| D - D ^ { * } \| \to 0 } \nabla _ { \theta } \mathbb { E } _ { z \sim p ( z ) } [ \log ( 1 - D ( g _ { \theta } ( z ) ) ) ] = 0
187
+ $$
188
+
189
+ ![](images/eb72ca1b4c80fde1b1bdaea98cf2f02a6ae9f63aee0c5c94dec9e0abd28e7ef1.jpg)
190
+ Figure 2: First, we trained a DCGAN for 1, 10 and 25 epochs. Then, with the generator fixed we train a discriminator from scratch and measure the gradients with the original cost function. We see the gradient norms decay quickly, in the best case 5 orders of magnitude after 4000 discriminator iterations. Note the logarithmic scale.
191
+
192
+ This shows that as our discriminator gets better, the gradient of the generator vanishes. For completeness, this was experimentally verified in Figure 2. The fact that this happens is terrible, since the fact that the generator’s cost function being close to the Jensen Shannon divergence depends on the quality of this approximation. This points us to a fundamental: either our updates to the discriminator will be inacurate, or they will vanish. This makes it difficult to train using this cost function, or leave up to the user to decide the precise amount of training dedicated to the discriminator, which can make GAN training extremely hard.
193
+
194
+ # 2.2.2 $\mathrm { T H E } - \log D$ ALTERNATIVE
195
+
196
+ To avoid gradients vanishing when the discriminator is very confident, people have chosen to use a different gradient step for the generator.
197
+
198
+ $$
199
+ \Delta \theta = \nabla _ { \theta } \mathbb { E } _ { z \sim p ( z ) } \left[ - \log D ( g _ { \theta } ( z ) ) \right]
200
+ $$
201
+
202
+ We now state and prove for the first time which cost function is being optimized by this gradient step. Later, we prove that while this gradient doesn’t necessarily suffer from vanishing gradients, it does cause massively unstable updates (that have been widely experienced in practice) under the prescence of a noisy approximation to the optimal discriminator.
203
+
204
+ Theorem 2.5. Let $\mathbb { P } _ { r }$ and $\mathbb { P } _ { g _ { \theta } }$ be two continuous distributions, with densities $P _ { r }$ and $P _ { g _ { \theta } }$ respectively. Let $\begin{array} { r } { D ^ { * } = \frac { P _ { r } } { { P _ { g _ { \theta _ { 0 } } } } + P _ { r } } } \end{array}$ be the optimal discriminator, fixed for a value ${ \theta _ { 0 } } ^ { 3 }$ . Therefore,
205
+
206
+ $$
207
+ \mathbb { E } _ { z \sim p ( z ) } \left[ - \nabla _ { \theta } \log D ^ { * } ( g _ { \theta } ( z ) ) | _ { \theta = \theta _ { 0 } } \right] = \nabla _ { \theta } \left[ K L ( \mathbb { P } _ { g _ { \theta } } \| \mathbb { P } _ { r } ) - 2 J S D ( \mathbb { P } _ { g _ { \theta } } \| \mathbb { P } _ { r } ) \right] | _ { \theta = \theta _ { 0 } }
208
+ $$
209
+
210
+ Before diving into the proof, let’s look at equation (3) for a second. This is the inverted $\mathrm { K L }$ minus two JSD. First of all, the JSDs are in the opposite sign, which means they are pushing for the distributions to be different, which seems like a fault in the update. Second, the KL appearing in the equation is $K L ( \mathbb { P } _ { g } \| \mathbb { P } _ { r } )$ , not the one equivalent to maximum likelihood. As we know, this KL assigns an extremely high cost to generating fake looking samples, and an extremely low cost on mode dropping; and the JSD is symetrical so it shouldn’t alter this behaviour. This explains what we see in practice, that GANs (when stabilized) create good looking samples, and justifies what is commonly conjectured, that GANs suffer from an extensive amount of mode dropping.
211
+
212
+ Proof. We already know by Goodfellow et al. (2014a) that
213
+
214
+ $$
215
+ \mathbb { E } _ { z \sim p ( z ) } \left[ \nabla _ { \theta } \log ( 1 - D ^ { * } ( g _ { \theta } ( z ) ) ) | _ { \theta = \theta _ { 0 } } \right] = \nabla _ { \theta } 2 J S D ( \mathbb { P } _ { g _ { \theta } } \| \mathbb { P } _ { r } ) | _ { \theta = \theta _ { 0 } }
216
+ $$
217
+
218
+ Furthermore, as remarked by Huszar (2016),
219
+
220
+ $$
221
+ \begin{array} { l } { { \displaystyle K L ( \mathbb { P } _ { g _ { \theta } } \| \mathbb { P } _ { r } ) = \mathbb { E } _ { x \sim \mathbb { P } _ { g _ { \theta } } } \left[ \log \frac { P _ { g _ { \theta } } ( x ) } { P _ { r } ( x ) } \right] } } \\ { { \displaystyle \qquad = \mathbb { E } _ { x \sim \mathbb { P } _ { g _ { \theta } } } \left[ \log \frac { P _ { g _ { \theta _ { 0 } } } ( x ) } { P _ { r } ( x ) } \right] - \mathbb { E } _ { x \sim \mathbb { P } _ { g _ { \theta } } } \left[ \log \frac { P _ { g _ { \theta } } ( x ) } { P _ { g _ { \theta _ { 0 } } } ( x ) } \right] } } \\ { { \displaystyle \qquad = - \mathbb { E } _ { x \sim \mathbb { P } _ { g _ { \theta } } } \left[ \log \frac { D ^ { * } ( x ) } { 1 - D ^ { * } ( x ) } \right] - K L ( \mathbb { P } _ { g _ { \theta } } \| \mathbb { P } _ { g _ { \theta _ { 0 } } } ) } } \\ { { \displaystyle \qquad = - \mathbb { E } _ { z \sim p ( z ) } \left[ \log \frac { D ^ { * } ( g _ { \theta } ( z ) ) } { 1 - D ^ { * } ( g _ { \theta } ( z ) ) } \right] - K L ( \mathbb { P } _ { g _ { \theta } } \| \mathbb { P } _ { g _ { \theta _ { 0 } } } ) } } \end{array}
222
+ $$
223
+
224
+ Taking derivatives in $\theta$ at $\theta _ { 0 }$ we get
225
+
226
+ $$
227
+ \begin{array} { r l } & { \nabla _ { \theta } K L ( \mathbb { P } _ { g _ { \theta } } \| \mathbb { P } _ { r } ) | _ { \theta = \theta _ { 0 } } = - \nabla _ { \theta } \mathbb { E } _ { z \sim p ( z ) } \left[ \log \frac { D ^ { * } \left( g _ { \theta } ( z ) \right) } { 1 - D ^ { * } \left( g _ { \theta } ( z ) \right) } \right] | _ { \theta = \theta _ { 0 } } - \nabla _ { \theta } K L ( \mathbb { P } _ { g _ { \theta } } \| \mathbb { P } _ { g _ { \theta _ { 0 } } } ) | _ { \theta = \theta _ { 0 } } } \\ & { \qquad = \mathbb { E } _ { z \sim p ( z ) } \left[ - \nabla _ { \theta } \log \frac { D ^ { * } \left( g _ { \theta } ( z ) \right) } { 1 - D ^ { * } \left( g _ { \theta } ( z ) \right) } \right] | _ { \theta = \theta _ { 0 } } } \end{array}
228
+ $$
229
+
230
+ Substracting this last equation with the result for the JSD, we obtain our desired result.
231
+
232
+ We now turn to our result regarding the instability of a noisy version of the true distriminator.
233
+
234
+ Theorem 2.6 (Instability of generator gradient updates). Let $g _ { \theta } : \mathcal { Z } \mathcal { X }$ be a differentiable function that induces a distribution $\mathbb { P } _ { g }$ . Let $\mathbb { P } _ { r }$ be the real data distribution, with either conditions of Theorems 2.1 or 2.2 satisfied. Let $D$ be a discriminator such that $D ^ { * } - D = \epsilon$ is a centered Gaussian process indexed by $x$ and independent for every $x$ (popularly known as white noise) and $\nabla _ { x } D ^ { * } - \nabla _ { x } D = r$ another independent centered Gaussian process indexed by $x$ and independent for every $x$ . Then, each coordinate of
235
+
236
+ $$
237
+ \mathbb { E } _ { z \sim p ( z ) } \left[ - \nabla _ { \theta } \log D ( g _ { \theta } ( z ) ) \right]
238
+ $$
239
+
240
+ is a centered Cauchy distribution with infinite expectation and variance.4
241
+
242
+ Proof. Let us remember again that in this case $D$ is locally constant equal to 0 on the support of $\mathbb { P } _ { g }$ . We denote $r ( z ) , \epsilon ( z )$ the random variables $r ( g _ { \boldsymbol { \theta } } ( z ) ) , \epsilon ( g _ { \boldsymbol { \theta } } ( z ) )$ . By the chain rule and the definition of $r , \epsilon$ , we get
243
+
244
+ $$
245
+ \begin{array} { r l } & { \mathbb { E } _ { z \sim p ( z ) } \left[ - \nabla _ { \theta } \log D ( g _ { \theta } ( z ) ) \right] = \mathbb { E } _ { z \sim p ( z ) } \left[ - \frac { J _ { \theta } g _ { \theta } ( z ) \nabla _ { x } D ( g _ { \theta } ( z ) ) } { D ( g _ { \theta } ( z ) ) } \right] } \\ & { \qquad = \mathbb { E } _ { z \sim p ( z ) } \left[ - \frac { J _ { \theta } g _ { \theta } ( z ) r ( z ) } { \epsilon ( z ) } \right] } \end{array}
246
+ $$
247
+
248
+ Since $r ( z )$ is a centered Gaussian distribution, multiplying by a matrix doesn’t change this fact. Furthermore, when we divide by $\epsilon ( z )$ , a centered Gaussian independent from the numerator, we get a centered Cauchy random variable on every coordinate. Averaging over $z$ the different independent Cauchy random variables again yields a centered Cauchy distribution. 5 □
249
+
250
+ ![](images/39ab35c5dcb1a5d5d9a1ea53e045d86866f0112f499873250eacc342cd1ae020.jpg)
251
+ Figure 3: First, we trained a DCGAN for 1, 10 and 25 epochs. Then, with the generator fixed we train a discriminator from scratch and measure the gradients with the $- \log D$ cost function. We see the gradient norms grow quickly. Furthermore, the noise in the curves shows that the variance of the gradients is also increasing. All these gradients lead to updates that lower sample quality notoriously.
252
+
253
+ Note that even if we ignore the fact that the updates have infinite variance, we still arrive to the fact that the distribution of the updates is centered, meaning that if we bound the updates the expected update will be 0, providing no feedback to the gradient.
254
+
255
+ Since the assumption that the noises of $D$ and $\nabla D$ are decorrelated is albeit too strong, we show in Figure 3 how the norm of the gradient grows drastically as we train the discriminator closer to optimality, at any stage in training of a well stabilized DCGAN except when it has already converged. In all cases, using this updates lead to a notorious decrease in sample quality. The noise in the curves also shows that the variance of the gradients is increasing, which is known to delve into slower convergence and more unstable behaviour in the optimization (Bottou et al., 2016).
256
+
257
+ # 3 TOWARDS SOFTER METRICS AND DISTRIBUTIONS
258
+
259
+ An important question now is how to fix the instability and vanishing gradients issues. Something we can do to break the assumptions of these theorems is add continuous noise to the inputs of the discriminator, therefore smoothening the distribution of the probability mass.
260
+
261
+ Theorem 3.1. If $X$ has distribution $\mathbb { P } _ { X }$ with support on $\mathcal { M }$ and $\epsilon$ is an aboslutely continuous distribution with density $P _ { \epsilon }$ , then $\mathbb { P } _ { X + \epsilon }$ is absolutely continuous with density
262
+
263
+ $$
264
+ \begin{array} { r l } & { P _ { X + \epsilon } ( x ) = \mathbb { E } _ { y \sim \mathbb { P } _ { X } } \left[ P _ { \epsilon } ( x - y ) \right] } \\ & { \quad \quad = \displaystyle \int _ { \mathcal { M } } P _ { \epsilon } ( x - y ) \mathrm { d } \mathbb { P } _ { X } ( y ) } \end{array}
265
+ $$
266
+
267
+ Proof. See Appendix A.
268
+
269
+ Corollary 3.1.
270
+
271
+ • $I f \epsilon \sim { \mathcal { N } } ( 0 , \sigma ^ { 2 } I )$ then
272
+
273
+ $$
274
+ P _ { X + \epsilon } ( x ) = \frac { 1 } { Z } \int _ { \mathcal { M } } e ^ { - \frac { \| y - x \| ^ { 2 } } { 2 \sigma ^ { 2 } } } \mathrm { d } \mathbb { P } _ { X } ( y )
275
+ $$
276
+
277
+ • If $\epsilon \sim \mathcal { N } ( 0 , \Sigma )$ then
278
+
279
+ $$
280
+ P _ { X + \epsilon } ( x ) = \frac { 1 } { Z } \mathbb { E } _ { y \sim \mathbb { P } _ { X } } \left[ e ^ { - \frac { 1 } { 2 } \left\| y - x \right\| _ { \Sigma ^ { - 1 } } ^ { 2 } } \right]
281
+ $$
282
+
283
+ • If $\begin{array} { r } { P _ { \epsilon } ( x ) \propto \frac { 1 } { \| x \| ^ { d + 1 } } } \end{array}$ then
284
+
285
+ $$
286
+ P _ { X + \epsilon } ( x ) = \frac { 1 } { Z } \mathbb { E } _ { y \sim \mathbb { P } _ { X } } \left[ \frac { 1 } { \| x - y \| ^ { d + 1 } } \right]
287
+ $$
288
+
289
+ This theorem therefore tells us that the density $P _ { X + \epsilon } ( x )$ is inversely proportional to the average distance to points in the support of $\mathbb { P } _ { X }$ , weighted by the probability of these points. In the case of the support of $\mathbb { P } _ { X }$ being a manifold, we will have the weighted average of the distance to the points along the manifold. How we choose the distribution of the noise $\epsilon$ will impact the notion of distance we are choosing. In our corolary, for example, we can see the effect of changing the covariance matrix by altering the norm inside the exponential. Different noises with different types of decays can therefore be used.
290
+
291
+ Now, the optimal discriminator between $\mathbb { P } _ { g + \epsilon }$ and $\mathbb { P } _ { r + \epsilon }$ is
292
+
293
+ $$
294
+ D ^ { * } ( x ) = \frac { P _ { r + \epsilon } ( x ) } { P _ { r + \epsilon } ( x ) + P _ { g + \epsilon } ( x ) }
295
+ $$
296
+
297
+ and we want to calculate what the gradient passed to the generator is.
298
+
299
+ Theorem 3.2. Let $\mathbb { P } _ { r }$ and $\mathbb { P } _ { g }$ be two distributions with support on $\mathcal { M }$ and $\mathcal { P }$ respectively, with $\epsilon \sim \mathcal { N } ( 0 , \sigma ^ { 2 } I )$ . Then, the gradient passed to the generator has the form
300
+
301
+ $$
302
+ \begin{array} { r l } & { \mathbb { E } _ { z \sim p ( z ) } \left[ \nabla _ { \theta } \log ( 1 - D ^ { * } ( g _ { \theta } ( z ) ) ) \right] } \\ & { \qquad = \mathbb { E } _ { z \sim p ( z ) } \bigg [ a ( z ) \int _ { \mathcal { M } } P _ { \epsilon } ( g _ { \theta } ( z ) - y ) \nabla _ { \theta } \| g _ { \theta } ( z ) - y \| ^ { 2 } \mathrm { d } \mathbb { P } _ { r } ( y ) } \\ & { \qquad - b ( z ) \int _ { \mathcal { P } } P _ { \epsilon } ( g _ { \theta } ( z ) - y ) \nabla _ { \theta } \| g _ { \theta } ( z ) - y \| ^ { 2 } \mathrm { d } \mathbb { P } _ { g } ( y ) \bigg ] } \end{array}
303
+ $$
304
+
305
+ where $a ( z )$ and $b ( z )$ are positive functions. Furthermore, $b > a$ if and only if $P _ { r + \epsilon } > P _ { g + \epsilon }$ , and $b < a$ if and only if $P _ { r + \epsilon } < P _ { g + \epsilon }$ .
306
+
307
+ This theorem proves that we will drive our samples $g _ { \boldsymbol { \theta } } ( z )$ towards points along the data manifold, weighted by their probability and the distance from our samples. Furthermore, the second term drives our points away from high probability samples, again, weighted by the sample manifold and distance to these samples. This is similar in spirit to contrastive divergence, where we lower the free energy of our samples and increase the free energy of data points. The importance of this term is seen more clearly when we have samples that have higher probability of coming from $\mathbb { P } _ { g }$ than from $\mathbb { P } _ { r }$ . In this case, we will have $b > a$ and the second term will have the strength to lower the probability of this too likely samples. Finally, if there’s an area around $x$ that has the same probability to come from $\mathbb { P } _ { g }$ than $\mathbb { P } _ { r }$ , the gradient contributions between the two terms will cancel, therefore stabilizing the gradient when $\mathbb { P } _ { r }$ is similar to $\mathbb { P } _ { g }$ .
308
+
309
+ There is one important problem with taking gradient steps exactly of the form (4), which is that in that case, $D$ will disregards errors that lie exactly in $g ( \mathcal { Z } )$ , since this is a set of measure 0. However, $g$ will be optimizing its cost only on that space. This will make the discriminator extremely susceptible to adversarial examples, and will render low cost on the generator without high cost on the discriminator, and lousy meaningless samples. This is easilly seen when we realize the term inside the expectation of equation (4) will be a positive scalar times $\nabla _ { x } \log ( 1 - D ^ { * } ( x ) ) \nabla _ { \theta } g _ { \theta } ( z )$ , which is the directional derivative towards the exact adversarial term of Goodfellow et al. (2014b). Because of this, it is important to backprop through noisy samples in the generator as well. This will yield a crucial benefit: the generator’s backprop term will be through samples on a set of positive measure that the discriminator will care about. Formalizing this notion, the actual gradient through the generator will now be proportional to $\nabla _ { \theta } J S D ( \mathbb { P } _ { r + \epsilon } \| \mathbb { P } _ { g + \epsilon } )$ , which will make the two noisy distributions match. As we anneal the noise, this will make $\mathbb { P } _ { r }$ and $\mathbb { P } _ { g }$ match as well. For completeness, we show the smooth gradient we get in this case. The proof is identical to the one of Theorem 3.2, so we leave it to the reader.
310
+
311
+ Corollary 3.2. Let $\epsilon , \epsilon ^ { \prime } \sim \mathcal { N } ( 0 , \sigma ^ { 2 } I )$ and $\tilde { g } _ { \boldsymbol { \theta } } ( z ) = g _ { \boldsymbol { \theta } } ( z ) + \epsilon ^ { \prime }$ , then
312
+
313
+ $$
314
+ \begin{array} { r l r } { { \mathbb { E } _ { z \sim p ( z ) , \epsilon ^ { \prime } } \bigl [ \nabla _ { \theta } \log ( 1 - D ^ { * } ( \tilde { g } _ { \theta } ( z ) ) ) \bigr ] } } \\ & { } & { = \mathbb { E } _ { z \sim p ( z ) , \epsilon ^ { \prime } } \bigg [ a ( z ) \int _ { \mathcal { M } } P _ { \epsilon } ( \tilde { g } _ { \theta } ( z ) - y ) \nabla _ { \theta } \| \tilde { g } _ { \theta } ( z ) - y \| ^ { 2 } \mathrm { d } \mathbb { P } _ { r } ( y ) } \\ & { } & { - b ( z ) \int _ { \mathcal { P } } P _ { \epsilon } ( \tilde { g } _ { \theta } ( z ) - y ) \nabla _ { \theta } \| \tilde { g } _ { \theta } ( z ) - y \| ^ { 2 } \mathrm { d } \mathbb { P } _ { g } ( y ) \bigg ] } \\ & { } & { = 2 \nabla _ { \theta } J S D ( \mathbb { P } _ { r + \epsilon } \| \mathbb { P } _ { g + \epsilon } ) } \end{array}
315
+ $$
316
+
317
+ In the same as with Theorem 3.2, $a$ and $b$ will have the same properties. The main difference is that we will be moving all our noisy samples towards the data manifold, which can be thought of as moving a small neighbourhood of samples towards it. This will protect the discriminator against measure 0 adversarial examples.
318
+
319
+ Proof of theorem 3.2. Since the discriminator is assumed fixed when backproping to the generator, the only thing that depends on $\theta$ is $g _ { \theta } ( z )$ for every $z$ . By taking derivatives on our cost function
320
+
321
+ $$
322
+ \begin{array} { r l } & { \mathbb { E } _ { z \sim p ( z ) } [ \nabla _ { \theta } \log ( 1 - D ^ { * } ( g _ { \theta } ( z ) ) ) ] } \\ & { \quad = \mathbb { E } _ { z \sim p ( z ) } [ \nabla _ { \theta } \log \frac { P _ { g + \epsilon } ( g _ { \theta } ( z ) ) } { P _ { r + \epsilon } ( g _ { \theta } ( z ) ) + P _ { g + \epsilon } ( g _ { \theta } ( z ) ) } ] } \\ & { \quad \quad = \mathbb { E } _ { z \sim p ( z ) } [ \nabla _ { \theta } \log P _ { g + \epsilon } ( g _ { \theta } ( z ) ) - \nabla _ { \theta } \log ( P _ { g + \epsilon } ( g _ { \theta } ( z ) ) + P _ { r + \epsilon } ( g _ { \theta } ( z ) ) ) ] } \\ & { \quad \quad = \mathbb { E } _ { z \sim p ( z ) } [ \frac { \nabla _ { \theta } P _ { g + \epsilon } ( g _ { \theta } ( z ) ) } { P _ { g + \epsilon } ( g _ { \theta } ( z ) ) } - \frac { \nabla _ { \theta } P _ { g + \epsilon } ( g _ { \theta } ( z ) ) + \nabla _ { \theta } P _ { r + \epsilon } ( g _ { \theta } ( z ) ) } { P _ { g + \epsilon } ( g _ { \theta } ( z ) ) + P _ { r + \epsilon } ( g _ { \theta } ( z ) ) } ] } \\ & { \quad \quad = \mathbb { E } _ { z \sim p ( z ) } [ \frac { 1 } { P _ { g + \epsilon } ( g _ { \theta } ( z ) ) + P _ { r + \epsilon } ( g _ { \theta } ( z ) ) } \nabla _ { \theta } [ \cdot \ : \ : \ : \ : \ : \ : } \\ & { \quad \quad \frac { 1 } { P _ { g + \epsilon } ( g _ { \theta } ( z ) ) + P _ { r + \epsilon } ( g _ { \theta } ( z ) ) } \nabla _ { \theta } [ \cdot \ : \ : \ : \ : \ : \ : } \\ & { \quad \ : \ : \frac { 1 } { P _ { g + \epsilon } ( g _ { \theta } ( z ) ) + P _ { r + \epsilon } ( g _ { \theta } ( z ) ) } \nabla _ { \theta } [ \cdot \ : \ : \ : \ : \ : \ : ( g _ { \theta } ( z ) ) ] - } \\ & \quad \ : \ : \frac { 1 } P _ { g + \epsilon } ( g _ { \theta } ( z ) ) + P _ { r + \epsilon } \end{array}
323
+ $$
324
+
325
+ Let the density of $\epsilon$ be $\frac { 1 } { Z } e ^ { - \frac { \| x \| ^ { 2 } } { 2 \sigma ^ { 2 } } }$ . We now define
326
+
327
+ $$
328
+ \begin{array} { l } { { a ( z ) = \displaystyle \frac { 1 } { 2 \sigma ^ { 2 } } \frac { 1 } { P _ { g + \epsilon } ( g _ { \theta } ( z ) ) + P _ { r + \epsilon } ( g _ { \theta } ( z ) ) } } } \\ { { b ( z ) = \displaystyle \frac { 1 } { 2 \sigma ^ { 2 } } \frac { 1 } { P _ { g + \epsilon } ( g _ { \theta } ( z ) ) + P _ { r + \epsilon } ( g _ { \theta } ( z ) ) } \frac { P _ { r + \epsilon } ( g _ { \theta } ( z ) ) } { P _ { g + \epsilon } ( g _ { \theta } ( z ) ) } } } \end{array}
329
+ $$
330
+
331
+ Trivially, $a$ and $b$ are positive functions. Since $\begin{array} { r } { b \ = \ a \frac { P _ { r + \epsilon } } { P _ { g + \epsilon } } } \end{array}$ a Pr+P , we know that b > a if and only if $P _ { r + \epsilon } > P _ { g + \epsilon }$ , and $b < a$ if and only if $P _ { r + \epsilon } < P _ { g + \epsilon }$ as we wanted. Continuing the proof, we know
332
+
333
+ $$
334
+ \begin{array} { r l } { \mathbb { E } _ { z \sim p ( z ) } \big [ \nabla _ { \theta } \log ( 1 - D ^ { * } ( g _ { \theta } ( z ) ) ) \big ] } & { } \\ & { = \mathbb { E } _ { z \sim p ( z ) } \left[ 2 \sigma ^ { 2 } a ( z ) \nabla _ { \theta } \big ( - P r _ { + } ( g _ { \theta } ( z ) ) \big ) - 2 \sigma ^ { 2 } b ( z ) \nabla _ { \theta } \big [ - P _ { g } + ( g _ { \theta } ( z ) ) \big ] \right. } \\ & { = \mathbb { E } _ { z \sim p ( z ) } \left[ 2 \sigma ^ { 2 } a ( z ) \int _ { M } - \nabla _ { \theta } \frac { 1 } { Z } e ^ { - \frac { \mathrm { l i g n } ( z ) - \mathrm { s i g n } ( z ) } { 2 q } } \mathrm { d } \mathrm { P r } _ { \theta } ( g ) - 2 \sigma ^ { 2 } b ( z ) \int _ { p ^ { - } } - \nabla _ { \theta } \frac { 1 } { Z } e ^ { - \frac { \mathrm { l i g n } ( z ) - \mathrm { s i g n } ( z ) } { 2 q } } \mathrm { d } \mathrm { P } _ { g } \right. } \\ & { = \mathbb { E } _ { \sim \sim p ( z ) } \bigg [ a ( z ) \int _ { M } \frac { 1 } { Z } e ^ { - \frac { \mathrm { l i g n } ( z ) - \mathrm { s i g n } ( z ) } { 2 q } } \nabla _ { \theta } \big \| g _ { \theta } ( z ) - y \big \| ^ { 2 } \mathrm { d } \mathrm { P r } _ { \theta } ( y ) } \\ & { - b ( z ) \int _ { p } \frac { 1 } { Z } e ^ { - \frac { \mathrm { l i g n } ( z ) - \mathrm { s i g n } ( z ) } { 2 q } } \nabla _ { \theta } \big \| g _ { \theta } ( z ) - y \big \| ^ { 2 } \mathrm { d } \mathrm { P } _ { g } ( y ) \bigg ] } \\ & { = \mathbb { E } _ { \sim p ( z ) } \bigg [ a ( z ) \int _ { M } P _ { + } ( g _ { \theta } ( z ) - y ) \nabla _ { \theta } \big \| g _ { \theta } ( z ) - y \big \| ^ { 2 } \mathrm { d } \mathrm { P r } _ { \theta } ( y ) } \\ & { - b ( z ) \int _ { p ^ { \star } } ( g _ { \theta } ( z ) - y ) \nabla _ { \theta } \big \| g _ { \theta } ( z ) - y \big \| ^ { 2 } \mathrm { d } \mathrm { P r } _ { \theta } ( y ) \bigg ] } \end{array}
335
+ $$
336
+
337
+ Finishing the proof.
338
+
339
+ An interesting observation is that if we have two distributions $\mathbb { P } _ { r }$ and $\mathbb { P } _ { g }$ with support on manifolds that are close, the noise terms will make the noisy distributions $\mathbb { P } _ { r + \epsilon }$ and $\mathbb { P } _ { g + \epsilon }$ almost overlap, and the JSD between them will be small. This is in drastic contrast to the noiseless variants $\mathbb { P } _ { r }$ and $\mathbb { P } _ { g }$ , where all the divergences are maxed out, regardless of the closeness of the manifolds. We could argue to use the JSD of the noisy variants to measure a similarity between the original distributions, but this would depend on the amount of noise, and is not an intrinsic measure of $\mathbb { P } _ { r }$ and $\mathbb { P } _ { g }$ . Luckilly, there are alternatives.
340
+
341
+ Definition 3.1. We recall the definition of the Wasserstein metric $W ( P , Q )$ for $P$ and $Q$ two distributions over $\mathcal { X }$ . Namely,
342
+
343
+ $$
344
+ W ( P , Q ) = \operatorname* { i n f } _ { \gamma \in \Gamma } \int _ { \mathcal { X } \times \mathcal { X } } \| x - y \| _ { 2 } d \gamma ( x , y )
345
+ $$
346
+
347
+ where $\Gamma$ is the set of all possible joints on $\mathcal { X } \times \mathcal { X }$ that have marginals $P$ and $Q$
348
+
349
+ The Wasserstein distance also goes by other names, most commonly the transportation metric and the earth mover’s distance. This last name is most explicative: it’s the minimum cost of transporting the whole probability mass of $P$ from its support to match the probability mass of $Q$ on $Q$ ’s support. This identification of transporting points from $P$ to $Q$ is done via the coupling $\gamma$ . We refer the reader to Villani (2009) for an in-depth explanation of these ideas. It is easy to see now that the Wasserstein metric incorporates the notion of distance (as also seen inside the integral) between the elements in the support of $P$ and the ones in the support of $Q$ , and that as the supports of $P$ and $Q$ get closer and closer, the metric will go to 0, inducing as well a notion of distance between manifolds.
350
+
351
+ Intuitively, as we decrease the noise, $\mathbb { P } _ { X }$ and $\mathbb { P } _ { X + \epsilon }$ become more similar. However, it is easy to see again that $J S D ( \mathbb { P } _ { X } \| \mathbb { P } _ { X + \epsilon } )$ is maxed out, regardless of the amount of noise. The following Lemma shows that this is not the case for the Wasserstein metric, and that it goes to 0 smoothly when we decrease the variance of the noise.
352
+
353
+ Lemma 4. If  is a random vector with mean $O$ , then we have
354
+
355
+ $$
356
+ W ( \mathbb { P } _ { X } , \mathbb { P } _ { X + \epsilon } ) \leq V ^ { \frac { 1 } { 2 } }
357
+ $$
358
+
359
+ where $V = \mathbb { E } [ \| \epsilon \| _ { 2 } ^ { 2 } ]$ is the variance of .
360
+
361
+ Proof. Let $x \sim \mathbb { P } _ { X }$ , and $y = x + \epsilon$ with $\epsilon$ independent from $x$ . We call $\gamma$ the joint of $( x , y )$ , which clearly has marginals $\mathbb { P } _ { X }$ and $\mathbb { P } _ { X + \epsilon }$ . Therefore,
362
+
363
+ $$
364
+ \begin{array} { r l } & { W ( \mathbb { P } _ { X } , \mathbb { P } _ { X + \epsilon } ) \leq \int \| x - y \| _ { 2 } d \gamma ( x , y ) } \\ & { \qquad = \mathbb { E } _ { x \sim \mathbb { P } _ { X } } \mathbb { E } _ { y \sim x + \epsilon } [ \| x - y \| _ { 2 } ] } \\ & { \qquad = \mathbb { E } _ { x \sim \mathbb { P } _ { X } } \mathbb { E } _ { y \sim x + \epsilon } [ \| \epsilon \| _ { 2 } ] } \\ & { \qquad = \mathbb { E } _ { x \sim \mathbb { P } _ { X } } \mathbb { E } _ { \epsilon } [ \| \epsilon \| _ { 2 } ] } \\ & { \qquad = \mathbb { E } _ { \epsilon } [ \| \epsilon \| _ { 2 } ] } \\ & { \qquad \leq \mathbb { E } _ { \epsilon } [ \| \epsilon \| _ { 2 } ^ { 2 } ] ^ { \frac { 1 } { 2 } } = V ^ { \frac { 1 } { 2 } } } \end{array}
365
+ $$
366
+
367
+ where the last inequality was due to Jensen.
368
+
369
+ We now turn to one of our main results. We are interested in studying the distance between $\mathbb { P } _ { r }$ and $\mathbb { P } _ { g }$ without any noise, even when their supports lie on different manifolds, since (for example) the closer these manifolds are, the closer to actual points on the data manifold the samples will be. Furthermore, we eventually want a way to evaluate generative models, regardless of whether they are continuous (as in a VAE) or not (as in a GAN), a problem that has for now been completely unsolved. The next theorem relates the Wasserstein distance of $\mathbb { P } _ { r }$ and $\mathbb { P } _ { g }$ , without any noise or modification, to the divergence of $\mathbb { P } _ { r + \epsilon }$ and $\mathbb { P } _ { g + \epsilon }$ , and the variance of the noise. Since $\mathbb { P } _ { r + \epsilon }$ and $\mathbb { P } _ { g + \epsilon }$ are continuous distributions, this divergence is a sensible estimate, which can even be attempted to minimize, since a discriminator trained on those distributions will approximate the JSD between them, and provide smooth gradients as per Corolary 3.2.
370
+
371
+ Theorem 3.3. Let $\mathbb { P } _ { r }$ and $\mathbb { P } _ { g }$ be any two distributions, and $\epsilon$ be a random vector with mean 0 and variance $V$ . If $\mathbb { P } _ { r + \epsilon }$ and $\mathbb { P } _ { g + \epsilon }$ have support contained on a ball of diameter $C$ , then 6
372
+
373
+ $$
374
+ W ( \mathbb { P } _ { r } , \mathbb { P } _ { g } ) \leq 2 V ^ { \frac { 1 } { 2 } } + 2 C \sqrt { J S D ( \mathbb { P } _ { r + \epsilon } \| \mathbb { P } _ { g + \epsilon } ) }
375
+ $$
376
+
377
+ Proof.
378
+
379
+ $$
380
+ \begin{array} { r l } & { W ( \mathbb { P } _ { r } , \mathbb { P } _ { g } ) \leq W ( \mathbb { P } _ { r } , \mathbb { P } _ { r + \epsilon } ) + W ( \mathbb { P } _ { r + \epsilon } , \mathbb { P } _ { g + \epsilon } ) + W ( \mathbb { P } _ { g + \epsilon } , \mathbb { P } _ { g } ) } \\ & { \qquad \leq 2 V ^ { \frac { 1 } { 2 } } + W ( \mathbb { P } _ { r + \epsilon } , \mathbb { P } _ { g + \epsilon } ) } \\ & { \qquad \leq 2 V ^ { \frac { 1 } { 2 } } + C \delta ( \mathbb { P } _ { r + \epsilon } , \mathbb { P } _ { g + \epsilon } ) } \\ & { \qquad \leq 2 V ^ { \frac { 1 } { 2 } } + C \left( \delta ( \mathbb { P } _ { r + \epsilon } , \mathbb { P } _ { m } ) + \delta ( \mathbb { P } _ { g + \epsilon } , \mathbb { P } _ { m } ) \right) } \\ & { \qquad \leq 2 V ^ { \frac { 1 } { 2 } } + C \left( \sqrt { \frac { 1 } { 2 } K L ( \mathbb { P } _ { r + \epsilon } \| \mathbb { P } _ { m } ) } + \sqrt { \frac { 1 } { 2 } K L ( \mathbb { P } _ { g + \epsilon } \| \mathbb { P } _ { m } ) } \right) } \\ & { \qquad \leq 2 V ^ { \frac { 1 } { 2 } } + 2 C \sqrt { J S D ( \mathbb { P } _ { r + \epsilon } \| \mathbb { P } _ { g + \epsilon } ) } } \\ & { \qquad \leq 2 V ^ { \frac { 1 } { 2 } } + 2 C \sqrt { J S D ( \mathbb { P } _ { r + \epsilon } \| \mathbb { P } _ { g + \epsilon } ) } } \end{array}
381
+ $$
382
+
383
+ We first used the Lemma 4 to bound everything but the middle term as a function of $V$ . After that, we followed by the fact that $W ( P , Q ) \leq C \delta ( P , Q )$ wih $\delta$ the total variation, which is a popular Lemma arizing from the Kantorovich-Rubinstein duality. After that, we used the triangular inequality on $\delta$ and $\mathbb { P } _ { m }$ the mixture distribution between $\mathbb { P } _ { g + \epsilon }$ and $\mathbb { P } _ { r + \epsilon }$ . Finally, we used Pinsker’s inequality and later the fact that each individual $K L$ is only one of the non-negative sumands of the $J S D$ . □
384
+
385
+ Theorem 3.3 points us to an interesting idea. The two terms in equation (6) can be controlled. The first term can be decreased by annealing the noise, and the second term can be minimized by a GAN when the discriminator is trained on the noisy inputs, since it will be approximating the JSD between the two continuous distributions. One great advantage of this is that we no longer have to worry about training schedules. Because of the noise, we can train the discriminator till optimality without any problems and get smooth interpretable gradients by Corollary 3.2. All this while still minimizing the distance between $\mathbb { P } _ { r }$ and $\mathbb { P } _ { g }$ , the two noiseless distributions we in the end care about.
386
+
387
+ # ACKNOWLEDGMENTS
388
+
389
+ The first author would like to especially thank Luis Scoccola for help with the proof of Lemma 1.
390
+
391
+ The authors would also like to thank Ishmael Belghazi, Yoshua Bengio, Gerry Che, Soumith Chintala, Caglar Gulcehre, Daniel Jiwoong Im, Alex Lamb, Luis Scoccola, Pablo Sprechmann, Arthur Szlam, Jake Zhao for insightful comments and advice.
392
+
393
+ # REFERENCES
394
+
395
+ Leon Bottou, Frank E. Curtis, and Jorge Nocedal. Optimization methods for large-scale machine ´ learning. CoRR, abs/1606.04838, 2016.
396
+
397
+ Emily L. Denton, Soumith Chintala, Arthur Szlam, and Rob Fergus. Deep generative image models using a laplacian pyramid of adversarial networks. In Advances in Neural Information Processing Systems 28, pp. 1486–1494. Curran Associates, Inc., 2015.
398
+
399
+ Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems 27, pp. 2672–2680. Curran Associates, Inc., 2014a.
400
+
401
+ Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. CoRR, abs/1412.6572, 2014b.
402
+
403
+ Kurt Hornik. Approximation capabilities of multilayer feedforward networks. Neural Networks, 4 (2):251 – 257, 1991.
404
+
405
+ Ferenc Huszar. How (not) to train your generative model: Scheduled sampling, likelihood, adversary? CoRR, abs/1511.05101, 2015.
406
+
407
+ Ferenc Huszar. An alternative update rule for generative adversarial networks. Blogpost, 2016. URL http://www.inference.vc/ an-alternative-update-rule-for-generative-adversarial-networks/.
408
+
409
+ Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. CoRR, abs/1312.6114, 2013.
410
+
411
+ Alex Lamb, Anirudh Goyal, Ying Zhang, Saizheng Zhang, Aaron Courville, and Yoshua Bengio. Professor forcing: A new algorithm for training recurrent networks. Corr, abs/1610.09038, 2016.
412
+
413
+ Hariharan Narayanan and Sanjoy Mitter. Sample complexity of testing the manifold hypothesis. In Advances in Neural Information Processing Systems 23, pp. 1786–1794. Curran Associates, Inc., 2010.
414
+
415
+ Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. CoRR, abs/1511.06434, 2015.
416
+
417
+ Tim Salimans, Ian J. Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. CoRR, abs/1606.03498, 2016.
418
+
419
+ Lucas Theis, Aaron van den Oord, and Matthias Bethge. A note on the evaluation of generative models. In International Conference on Learning Representations, Apr 2016.
420
+
421
+ Cedric Villani. ´ Optimal Transport: Old and New. Grundlehren der mathematischen Wissenschaften. Springer, Berlin, 2009. ISBN 978-3-540-71049-3. URL http://opac.inria. fr/record $\underline { { \underline { { \mathbf { \Pi } } } } } =$ b1129524.
422
+
423
+ Jiajun Wu, Chengkai Zhang, Tianfan Xue, William T. Freeman, and Joshua B. Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Corr, abs/1610.07584, 2016.
424
+
425
+ # A PROOFS OF THINGS
426
+
427
+ Proof of Lemma 1. We first consider the case where the nonlinearities are rectifiers or leaky rectifiers of the form $\sigma ( x ) ~ = ~ \mathbb { 1 } [ x ~ < ~ 0 ] c _ { 1 } x + \mathbb { 1 } [ x ~ \geq ~ 0 ] c _ { 2 } x$ for some $c _ { 1 } , c _ { 2 } \ \in \ \mathbb { R }$ . In this case, $g ( z ) = \mathbf { D } _ { n } \mathbf { W } _ { n } \ldots \mathbf { D } _ { 1 } \mathbf { W } _ { 1 } z$ , where $\mathbf { W } _ { i }$ are affine transformations and $\mathbf { D } _ { i }$ are some diagonal matrices dependent on $z$ that have diagonal entries $c _ { 1 }$ or $c _ { 2 }$ . If we consider $\mathcal { D }$ to be the (finite) set of all diagonal matrices with diagonal entries $c _ { 1 }$ or $c _ { 2 }$ , then $g ( \mathcal { Z } ) \subseteq \bigcup _ { D _ { i } \in \mathcal { D } } \mathbf { D } _ { n } \mathbf { W } _ { n } \ldots \mathbf { D } _ { 1 } \mathbf { W } _ { 1 } \mathcal { Z }$ , which is a finite union of linear manifolds.
428
+
429
+ The proof for the second case is technical and slightly more involved. When $\sigma$ is a pointwise smooth strictly increasing nonlinearity, then applying it vectorwise it’s a diffeomorphism to its image. Therefore, it sends a countable union of manifolds of dimension $d$ to a countable union of manifolds of dimension $d$ . If we can prove the same thing for affine transformations we will be finished, since $g ( \mathcal { Z } )$ is just a composition of these applied to a $\mathrm { d i m } \mathcal { Z }$ dimensional manifold. Of course, it suffices to prove that an affine transformation sends a manifold to a countable union of manifolds without increasing dimension, since a countable union of countable unions is still a countable union. Furthermore, we only need to show this for linear transformations, since applying a bias term is a diffeomorphism.
430
+
431
+ Let $\mathbf { W } \in \mathbb { R } ^ { n \times m }$ be a matrix. Note that by the singular value decomposition, $\mathbf { W } = \mathbf { U } \pmb { \Sigma } \mathbf { V }$ , where $\pmb { \Sigma }$ is a square diagonal matrix with diagonal positive entries and $\mathbf { U } , \mathbf { V }$ are compositions of changes of basis, inclusions (meaning adding 0s to new coordinates) and projections to a subset of the coordinates. Multiplying by $\Sigma$ and applying a change of basis are diffeomorphisms, and adding 0s to new coordinates is a manifold embedding, so we only need to prove our statement for projections onto a subset of the coordinates. Let $\pi : \bar { \mathbb { R } } ^ { n + k } \bar { \mathbb { R } } ^ { \bar { n } }$ , where $\pi ( x _ { 1 } , \ldots , x _ { n + k } ) = ( x _ { 1 } , \ldots , x _ { n } )$ be our projection and $\mathcal { M } \subseteq \mathbb { R } ^ { n + k }$ our $d$ -dimensional manifold. If $n \leq d$ , we are done since the image of $\pi$ is contained in all $\mathbb { R } ^ { n }$ , a manifold with at most dimension $d$ . We now turn to the case where $n > d$ . Let $\pi _ { i } ( x ) = x _ { i }$ be the projection onto the $i$ -th coordinate. If $x$ is a critical point of $\pi$ , since the coordinates of $\pi$ are independent, then $x$ has to be a critical point of a $\pi _ { i }$ . By a consequence of the Morse Lemma, the critical points of $\pi _ { i }$ are isolated, and therefore so are the ones of $\pi$ , meaning that there is at most a countable number of them. Since $\pi$ maps the non-critical points onto a $d$ dimensional manifold (because it acts as an embedding) and the countable number of critical points into a countable number of points (or 0 dimensional manifolds), the proof is finished. □
432
+
433
+ Proof of Lemma 2. For now we assume that $\mathcal { M }$ and $\mathcal { P }$ are without boundary. If $\dim { \mathcal { M } } + \dim { \mathcal { P } } \geq$ $d$ it is known that under arbitrarilly small perturbations defined as the ones in the statement of this Lemma, the two dimensions will intersect only transversally with probability 1 by the General Position Lemma. If $\dim { \mathcal { M } } + \dim { \mathcal { P } } < d$ , we will show that with probability 1, $\mathcal { M } + \eta$ and $\mathcal { P } + \eta ^ { \prime }$ will not intersect, thereby getting our desired result. Let us then assume $\dim { \mathcal { M } } + \dim { \mathcal { P } } < d$ . Note that $\hat { \mathcal { M } } \cap \hat { \mathcal { P } } \neq \varnothing$ if and only if there are $x \in { \mathcal { M } } , y \in { \mathcal { P } }$ such that $x + \eta = y + \eta ^ { \prime }$ , or equivalently $x - y = \eta ^ { \prime } - \eta$ . Therefore, $\hat { \mathcal { M } }$ and $\hat { \mathcal { P } }$ intersect if and only if $\eta ^ { \prime } - \eta \in \mathcal { M } - \mathcal { P }$ . Since $\eta , \eta ^ { \prime }$ are independent continuous random variables, the difference is also continuous. If $\mathcal { M } - \mathcal { P }$ has measure 0 in $\mathbf { \bar { \mathbb { R } } } ^ { d }$ then $\mathbb { P } ( \eta ^ { \prime } - \eta \in \mathcal { M } - \mathcal { P } ) = 0$ , concluding the proof. We will therefore show that $\mathcal { M } - \mathcal { P }$ has measure 0. Let $f : \mathcal { M } \times \mathcal { P } \to \mathbb { R } ^ { d }$ be $f ( x , y ) = x - y$ . If $m$ and $p$ are the dimensions of $\mathcal { M }$ and $\mathcal { P }$ , then $f$ is a smooth function between an $m + p$ -dimensional manifold and a $d$ dimensional one. Clearly, the image of $f$ is $\mathcal { M } - \mathcal { P }$ . Therefore,
434
+
435
+ $$
436
+ \begin{array} { r l } & { \mathcal { M } - \mathcal { P } = f ( \{ z \in \mathcal { M } \times \mathcal { P } | \mathrm { r a n k } ( d _ { z } f ) < m + p \} ) } \\ & { \qquad \cup \ f ( \{ z \in \mathcal { M } \times \mathcal { P } | \mathrm { r a n k } ( d _ { z } f ) = m + p \} ) } \end{array}
437
+ $$
438
+
439
+ The first set is the image of the critical points, namely the critical values. By Sard’s Lemma, this set has measure 0. Let’s call $A = \{ z \in \mathcal { M } \times \mathcal { P } | \mathrm { r a n k } ( d _ { z } f ) = m + p \}$ . Let $z$ be an element of $A$ . By the inverse function theorem, there is a neighbourhood $U _ { z } \subseteq \mathcal { M } \times \mathcal { P }$ of $z$ such that $f | _ { U _ { z } }$ is an embedding. Since every manifold has a countable topological basis, we can cover $A$ by countable sets $U _ { z _ { n } }$ , where $n \in \mathbb N$ . We will just note them by $U _ { n }$ . Since $f | _ { U _ { n } }$ is an embedding, $f ( U _ { n } )$ is an $m + p$ -dimensional manifold, and since $m + p < d$ , this set has measure 0 in $\mathbb { R } ^ { d }$ . Now, $\textstyle f ( A ) = \bigcup _ { n \in \mathbb { N } } f ( U _ { n } )$ , which therefore has measure 0 in $\mathbb { R } ^ { d }$ , finishing the proof of the boundary free case.
440
+
441
+ Now we consider the case where $\mathcal { M }$ and $\mathcal { P }$ are manifolds with boundary. By a simple union bound,
442
+
443
+ $$
444
+ \begin{array} { r l } { \mathbb { P } _ { \eta , \eta ^ { \prime } } \big ( \tilde { \mathcal { M } } \mathrm { p e r f e c t l y ~ a l i g n s ~ w i t h ~ } \tilde { \mathcal { P } } \big ) \le \mathbb { P } _ { \eta , \eta ^ { \prime } } \big ( \mathrm { I n t } \tilde { \mathcal { M } } \mathrm { p e r f e c t l y ~ a l i g n s ~ w i t h ~ I n t } \tilde { \mathcal { P } } \big ) } & { } \\ { + \mathbb { P } _ { \eta , \eta ^ { \prime } } \big ( \mathrm { I n t } \tilde { \mathcal { M } } \mathrm { p e r f e c t l y ~ a l i g n s ~ w i t h ~ } \partial \tilde { \mathcal { P } } \big ) } & { } \\ { + \mathbb { P } _ { \eta , \eta ^ { \prime } } \big ( \partial \tilde { \mathcal { M } } \mathrm { p e r f e c t l y ~ a l i g n s ~ w i t h ~ I n t } \tilde { \mathcal { P } } \big ) } & { } \\ { + \mathbb { P } _ { \eta , \eta ^ { \prime } } \big ( \partial \tilde { \mathcal { M } } \mathrm { p e r f e c t l y ~ a l i g n s ~ w i t h ~ } \partial \tilde { \mathcal { P } } \big ) } & { } \\ { = 0 } \end{array}
445
+ $$
446
+
447
+ where the last equality arizes when combining the facts that Int $\mathcal { M } = \eta + \mathrm { I n t } \mathcal { M } = \mathrm { I n t } \left( \eta + \mathcal { M } \right) =$ Int $\tilde { \mathcal { M } }$ (and analogously for the boundary and $\mathcal { P }$ ), that the boundary and interiors of $\mathcal { M }$ and $\mathcal { P }$ are boundary free regular submanifolds of ${ \bar { \mathbb { R } } } ^ { d }$ without full dimension, and then applying the boundary free case of the proof. □
448
+
449
+ Proof of Lemma 3. Let $m = \dim { \mathcal { M } }$ and $p = \dim { \mathcal { P } }$ . We again consider first the case where $\mathcal { M }$ and $\mathcal { P }$ are manifolds without boundary. If $m + p < d$ , then ${ \mathcal { L } } = \emptyset$ so the statement is obviously true. If $m + p \geq d$ , then $\mathcal { M }$ and $\mathcal { P }$ intersect transversally. This implies that $\mathcal { L }$ is a manifold of dimension $m + p - d < m , p$ . Since $\mathcal { L }$ is a submanifold of both $\mathcal { M }$ and $\mathcal { P }$ that has lower dimension, it has measure 0 on both of them.
450
+
451
+ We now tackle the case where $\mathcal { M }$ and $\mathcal { P }$ have boundaries. Let us remember that ${ \mathcal { M } } = \operatorname { I n t } { \mathcal { M } } \cup { \partial { \mathcal { M } } }$ and the union is disjoint (and analogously for $\mathcal { P }$ ). By using elementary properties of sets, we can trivially see that
452
+
453
+ $$
454
+ { \mathcal { L } } = { \mathcal { M } } \cap { \mathcal { P } } = ( { \mathrm { I n t } } { \mathcal { M } } \cap { \mathrm { I n t } } \mathscr { P } ) \cup ( { \mathrm { I n t } } { \mathcal { M } } \cap \partial { \mathcal { P } } ) \cup ( \partial { \mathcal { M } } \cap { \mathrm { I n t } } \mathscr { P } ) \cup ( \partial { \mathcal { M } } \cap \partial { \mathcal { P } } )
455
+ $$
456
+
457
+ where the unions are disjoint. This is the disjoint union of 4 strictly lower dimensional manifolds, by using the first part of the proof. Since each one of these intersections has measure 0 on either the interior or boundary of $\mathcal { M }$ (again, by the first part of the proof), and interior and boundary are contained in $\mathcal { M }$ , each one of the four intersections has measure 0 in $\mathcal { M }$ . Analogously, they have measure 0 in $\mathcal { P }$ , and by a simple union bound we see that $\mathcal { L }$ has measure 0 in $\mathcal { M }$ and $\mathcal { P }$ finishing the remaining case of the proof. □
458
+
459
+ Proof of Theorem 3.1. We first need to show that $\mathbb { P } _ { X + \epsilon }$ is absolutely continuous. Let $A$ be a Borel set with Lebesgue measure 0. Then, by the fact that $\epsilon$ and $X$ are independent, we know by Fubini
460
+
461
+ $$
462
+ \begin{array} { l l l } { \mathbb { P } _ { X + \epsilon } \mathopen { } \mathclose \bgroup \left( A \aftergroup \egroup \right) = \biggr \rvert _ { \mathbb { R } ^ { d } } \mathbb { P } _ { \epsilon } \mathopen { } \mathclose \bgroup \left( A - x \aftergroup \egroup \right) \mathrm { d } \mathbb { P } _ { X } \mathopen { } \mathclose \bgroup \left( x \aftergroup \egroup \right) } \\ { \qquad = \biggr \rvert _ { \mathbb { R } ^ { d } } 0 \mathrm { d } \mathbb { P } _ { X } \mathopen { } \mathclose \bgroup \left( x \aftergroup \egroup \right) = 0 } \end{array}
463
+ $$
464
+
465
+ Where we used the fact that if $A$ has Lebesgue measure zero, then so does $A - x$ and since $\mathbb { P } _ { \epsilon }$ is absolutely continuous, $\mathbb { P } _ { \epsilon } ( A - x ) = 0$ .
466
+
467
+ Now we calculate the density of $\mathbb { P } _ { X + \epsilon }$ . Again, by using the independence of $X$ and $\epsilon$ , for any Borel set $B$ we know
468
+
469
+ $$
470
+ \begin{array} { l l } { \displaystyle \mathbb { P } _ { X + \epsilon } ( B ) = \int _ { \mathbb { R } ^ { d } } \mathbb { P } _ { \epsilon } ( B - y ) \mathrm { d } \mathbb { P } _ { X } ( y ) } \\ { \displaystyle \quad \quad = \mathbb { E } _ { y \sim \mathbb { P } _ { X } } \left[ \mathbb { P } _ { \epsilon } ( B - y ) \right] } \\ { \displaystyle \quad = \mathbb { E } _ { y \sim \mathbb { P } _ { \epsilon } } \left[ \int _ { B - y } P _ { \epsilon } ( x ) d x \right] } \\ { \displaystyle \quad \quad = \mathbb { E } _ { y \sim \mathbb { P } _ { z } } \left[ \int _ { B } P _ { \epsilon } ( x - y ) d x \right] } \\ { \displaystyle \quad = \int _ { B } \mathbb { E } _ { y \sim \mathbb { P } _ { z } } \left[ P _ { \epsilon } ( x - y ) \right] d x } \end{array}
471
+ $$
472
+
473
+ Therefore, $\begin{array} { r } { \mathbb { P } _ { X + \epsilon } ( B ) = \int _ { B } P _ { X + \epsilon } ( x ) d x } \end{array}$ for our proposed $P _ { X + \epsilon }$ and all Borel sets $B$ . By the uniqueness of the Radon-Nikodym theorem, this implies the proposed $P _ { X + \epsilon }$ is the density of $\mathbb { P } _ { X + \epsilon }$ . The equivalence of the formula changing the expectation for $\ b { \int _ { \mathcal { M } } \mathbb { P } _ { X } }$ is trivial by the definition of expectation and the fact that the support of $\mathbb { P } _ { X }$ lies on $\mathcal { M }$ . □
474
+
475
+ # B FURTHER CLARIFICATIONS
476
+
477
+ In this appendix we further explain some of the terms and ideas mentioned in the paper, which due to space constrains, and to keep the flow of the paper, couldn’t be extremely developed in the main text. Some of these have to do with notation, others with technical elements of the proofs. On the latter case, we try to convey more intuition than we previously could. We present these clarifications in a very informal fashion in the following item list.
478
+
479
+ • There are two different but very related properties a random variable can have. A random variable $X$ is said to be continuous if $P ( { \bar { X } } { \bar { = } } x ) = 0$ for all single points $x \in \mathcal { X }$ . Note that a random variable concentrated on a low dimensional manifold such as a plane can have this property. However, an absolutely continuous random variable has the following property: if a set $A$ has Lebesgue measure 0, then $P ( X \in A ) = 0$ . Since points have measure 0 with the Lebesgue measure, absolute continuity implies continuity. A random variable that’s supported on a low dimensional manifold therefore will not be absolutely continuous: let $\mathcal { M }$ a low dimensional manifold be the support of $X$ . Since a low dimensional manifold has 0 Lebesgue measure, this would imply $P ( X \in M ) = 0$ , which is an absurd since $\mathcal { M }$ was the support of $X$ . The property of $X$ being absolutely continuous can be shown to be equivalent to $X$ having a density: the existence of a function $f : \mathcal { X } \mathbb { R }$ such that $\textstyle P ( X \in { \bar { A } } ) = \int _ { A } f ( x ) \operatorname { d } x$ (this is a consequence of the Radon-Nikodym theorem).
480
+
481
+ The annoying part is that in everyday paper writing when we talk about continuous random variables, we omit the ”absolutely” word to keep the text concise and actually talk about absolutely continuous random variables (ones that have a density), this is done through almost all sciences and throughout mathematics as well, annoying as it is. However we made the clarification in here since it’s relevant to our paper not to mistake the two terms.
482
+
483
+ • The notation $\mathbb { P } _ { r } [ D ( x ) = 1 ] = 1$ is the abbreviation of $\mathbb { P } _ { r } [ \{ x \in \mathcal { X } : D ( x ) = 1 \} ] = 1$ for a measure $\mathbb { P } _ { r }$ . Another way of expressing this more formally is $\mathbb { P } _ { r } [ D ^ { - 1 } ( 1 ) ] = 1$ .
484
+
485
+ • In the proof of Theorem 2.1, the distance between sets $d ( A , B )$ is defined as the usual distance between sets in a metric space
486
+
487
+ $$
488
+ d ( A , B ) = \operatorname* { i n f } _ { x \in A , y \in B } d ( x , y )
489
+ $$
490
+
491
+ where $d ( x , y )$ is the distance between points (in our case the Euclidean distance).
492
+
493
+ • Note that not everything that’s outside of the support of $\mathbb { P } _ { r }$ has to be a generated image. Generated images are only things that lie in the support of $\mathbb { P } _ { g }$ , and there are things that don’t need to be in the support of either $\mathbb { P } _ { r }$ or $\mathbb { P } _ { g }$ (these could be places where $0 < D < 1$ for example). This is because the discriminator is not trained to discriminate $\mathbb { P } _ { r }$ from all things that are not $\mathbb { P } _ { r }$ , but to distinguish $\mathbb { P } _ { r }$ from $\mathbb { P } _ { g }$ . Points that don’t lie in the support of $\mathbb { P } _ { r }$ or $\mathbb { P } _ { g }$ are not important to the performance of the discriminator (as is easily evidenced in its cost). Why we define accuracy 1 as is done in the text is to avoid the identification of a single ‘tight’ support, since this typically leads to problems (if I take a measure 0 set from any support it still is the support of the distribution). In the end, what we aim for is:
494
+
495
+ – We want $D ( x ) = 1$ with probability 1 when $x \sim \mathbb { P } _ { r }$ .
496
+ – We want $D ( x ) = 0$ with probability 1 when $x \sim \mathbb { P } _ { g }$ .
497
+ – Whatever happens elsewhere is irrelevant (as it is also reflected by the cost of the discriminator)
498
+
499
+ • We say that a discriminator $D ^ { * }$ is optimal for $g _ { \theta }$ (or its corresponding $\mathbb { P } _ { g }$ ) if for all measurable functions $D : \mathcal { X } [ 0 , 1 ]$ we have
500
+
501
+ $$
502
+ L ( D ^ { * } , g _ { \boldsymbol { \theta } } ) \geq L ( D , g _ { \boldsymbol { \theta } } )
503
+ $$
504
+
505
+ for $L$ defined as in equation (1).
parse/train/Hk4_qw5xe/Hk4_qw5xe_content_list.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/Hk4_qw5xe/Hk4_qw5xe_middle.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/Hk4_qw5xe/Hk4_qw5xe_model.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/Hk9Xc_lR-/Hk9Xc_lR-.md ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/Hk9Xc_lR-/Hk9Xc_lR-_content_list.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/Hk9Xc_lR-/Hk9Xc_lR-_middle.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/Hk9Xc_lR-/Hk9Xc_lR-_model.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/HkeGhoA5FX/HkeGhoA5FX.md ADDED
@@ -0,0 +1,352 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # RESIDUAL NON-LOCAL ATTENTION NETWORKS FOR IMAGE RESTORATION
2
+
3
+ Yulun Zhang1, Kunpeng $\mathbf { L i } ^ { 1 }$ , Kai $\mathbf { L i } ^ { 1 }$ , Bineng Zhong2 & Yun $\mathbf { F u ^ { 1 , 3 } }$
4
+
5
+ 1Department of ECE, Northeastern University, Boston, MA 02115, USA 2School of Computer Science and Technology, Huaqiao University, Xiamen 362100, China 3College of CIS, Northeastern University, Boston, MA 02115, USA {yulun100,kinpeng.li.1994,li.kai.gml}@gmail.com, bnzhong@hqu.edu.cn,yunfu@ece.neu.edu
6
+
7
+ # ABSTRACT
8
+
9
+ In this paper, we propose a residual non-local attention network for high-quality image restoration. Without considering the uneven distribution of information in the corrupted images, previous methods are restricted by local convolutional operation and equal treatment of spatial- and channel-wise features. To address this issue, we design local and non-local attention blocks to extract features that capture the long-range dependencies between pixels and pay more attention to the challenging parts. Specifically, we design trunk branch and (non-)local mask branch in each (non-)local attention block. The trunk branch is used to extract hierarchical features. Local and non-local mask branches aim to adaptively rescale these hierarchical features with mixed attentions. The local mask branch concentrates on more local structures with convolutional operations, while non-local attention considers more about long-range dependencies in the whole feature map. Furthermore, we propose residual local and non-local attention learning to train the very deep network, which further enhance the representation ability of the network. Our proposed method can be generalized for various image restoration applications, such as image denoising, demosaicing, compression artifacts reduction, and super-resolution. Experiments demonstrate that our method obtains comparable or better results compared with recently leading methods quantitatively and visually.
10
+
11
+ # 1 INTRODUCTION
12
+
13
+ Image restoration aims to recover high-quality (HQ) images from their corrupted low-quality (LQ) observations and plays a fundamental role in various high-level vision tasks. It is a typical ill-posed problem due to the irreversible nature of the image degradation process. Some most widely studied image restoration tasks include image denoising, demosaicing, and compression artifacts reduction. By distinctively modelling the restoration process from LQ observations to HQ objectives, i.e., without assumption for a specific restoration task when modelling, these tasks can be uniformly addressed in the same framework. Recently, deep convolutional neural network (CNN) has shown extraordinary capability of modelling various vision problems, ranging from low-level (e.g., image denoising (Zhang et al., 2017a), compression artifacts reduction (Dong et al., 2015), and image super-resolution (Kim et al., 2016; Lai et al., 2017; Tai et al., 2017; Lim et al., 2017; Zhang et al., 2018a; Haris et al., 2018; Wang et al., 2018c; Zhang et al., 2018b; Wang et al., 2018b)) to high-level (e.g., image recognition (He et al., 2016)) vision applications.
14
+
15
+ Stacked denoising auto-encoder (Vincent et al., 2008) is one of the best well-known CNN based model for image restoration. Dong et al. proposed SRCNN (Dong et al., 2014) for image superresolution and ARCNN (Dong et al., 2015) for image compression artifacts reduction. Both SRCNN and ARCNN achieved superior performance against previous works. By introducing residual learning to ease the training difficulty for deeper network, Zhang et al. proposed DnCNN (Zhang et al., 2017a) for image denoising and compression artifacts reduction. The denoiser prior was lately introduced in IRCNN (Zhang et al., 2017b) for fast image restoration. Mao et al. proposed a very deep fully convolutional encoding-decoding framework with symmetric skip connections for image restoration (Mao et al., 2016). Tai et al. later proposed a very deep end-to-end persistent memory network (MemNet) for image restoration (Tai et al., 2017) and achieved promising results. These CNN based methods have demonstrated the great ability of CNN for image restoration tasks.
16
+
17
+ However, there are mainly three issues in the existing CNN based methods above. First, the receptive field size of these networks is relatively small. Most of them extract features in a local way with convolutional operation, which fails to capture the long-range dependencies between pixels in the whole image. A larger receptive field size allows to make better use of training inputs and more context information. This would be very helpful to capture the latent degradation model of LQ images, especially when the images suffer from heavy corruptions. Second, distinctive ability of these networks is also limited. Let’s take image denoising as an example. For a noisy image, the noise may appear in both the plain and textural regions. Noise removal would be easier in the plain area than that in the textural one. It is desired to make the denoising model focus on textual area more. However, most previous denoising methods neglect to consider different contents in the noisy input and treat them equally. This would result in over-smoothed outputs and some textural details would also fail to be recovered. Third, all channel-wise features are treated equally in those networks. This naive treatment lacks flexibility in dealing with different types of information (e.g., low- and high-frequency information). For a set of features, some contain more information related to HQ image and the others may contain more information related to corruptions. The interdependencies among channels should be considered for more accurate image restoration.
18
+
19
+ To address the above issues, we propose the very deep residual non-local attention networks (RNAN) for high-quality image restoration. We design residual local and non-local attention blocks as the basic building modules for the very deep network. Each attention block consists of trunk and mask branches. We introduce residual block (He et al., 2016; Lim et al., 2017) for trunk branch and extract hierarchical features. For mask branch, we conduct feature downscaling and upscaling with largestride convolution and deconvolution to enlarge receptive field size. Furthermore, we incorporate non-local block in the mask branch to obtain residual non-local mixed attention. We apply RNAN for various restoration tasks, including image denoising, demosaicing, and compression artifacts reduction. Extensive experiments show that our proposed RNAN achieves state-of-the-art results compared with other recent leading methods in all tasks. To the best of our knowledge, this is the first time to consider residual non-local attention for image restoration problems.
20
+
21
+ The main contributions of this work are three-fold:
22
+
23
+ • We propose the very deep residual non-local networks for high-quality image restoration. The powerful networks are based on our proposed residual local and non-local attention blocks, which consist of trunk and mask branches. The network obtains non-local mixed attention with non-local block in the mask branch. Such attention mechanis helps to learn local and non-local information from the hierarchical features. • We propose residual non-local attention learning to train very deep networks by preserving more low-level features, being more suitable for image restoration. Using non-local lowlevel and high-level attention from the very deep network, we can pursue better network representational ability and finally obtain high-quality image restoration results. We demonstrate with extensive experiments that our RNAN is powerful for various image restoration tasks. RNAN achieves superior results over leading methods for image denoising, demosaicing, compression artifacts reduction, and super-resolution. In addition, RNAN achieves superior performance with moderate model size and performs very fast.
24
+
25
+ # 2 RELATED WORK
26
+
27
+ Non-local prior. As a classical filtering algorithm, non-local means (Buades et al., 2005) is computed as weighted mean of all pixels of an image. Such operation allows distant pixels to contribute to the response of a position at a time. It was lately introduced in BM3D (Dabov et al., 2007b) for image denoising. Recently, Wang et al. (2018a) proposed non-local neural network by incorporating non-local operations in deep neural network for video classification. We can see that those methods mainly introduce non-local information in the trunk pipeline. Liu et al. (2018) proposed non-local recurrent network for image restoration. However, in this paper, we mainly focus on learning nonlocal attention to better guide feature extraction in trunk branch.
28
+
29
+ Attention mechanisms. Generally, attention can be viewed as a guidance to bias the allocation of available processing resources towards the most informative components of an input (Hu et al.,
30
+
31
+ 2017). Recently, tentative works have been proposed to apply attention into deep neural networks (Wang et al., 2017; Hu et al., 2017). It’s usually combined with a gating function (e.g., sigmoid) to rescale the feature maps. Wang et al. (2017) proposed residual attention network for image classification with a trunk-and-mask attention mechanism. Hu et al. (2017) proposed squeezeand-excitation (SE) block to model channel-wise relationships to obtain significant performance improvement for image classification. In all, these works mainly aim to guide the network pay more attention to the regions of interested. However, few works have been proposed to investigate the effect of attention for image restoration tasks. Here, we want to enhances the network with distinguished power for noise and image content.
32
+
33
+ Image restoration architectures. Stacked denoising auto-encoder (Vincent et al., 2008) is one of the most well-known CNN-based image restoration method. (Dong et al., 2015) proposed ARCNN for image compression artifact reduction with several stacked convolutional layers. With the help of residual learning and batch normalization (Ioffe & Szegedy, 2015), Zhang et al. proposed DnCNN (Zhang et al., 2017a) for accurate image restoration and denoiser priors for image restoration in IRCNN (Zhang et al., 2017b). Recently, great progresses have been made in image restoration community, where Timofte et al. (Timofte et al., 2017), Ancuti et al. (Ancuti et al., 2018), and Blau et al. (Blau et al., 2018) lead the main competitions recently and achieved new research status and records. For example, Wang et al. (Wang et al., 2018c) proposed a fully progressive image SR approach. However, most methods are plain networks and neglect to use non-local information.
34
+
35
+ ![](images/0f225b1c9125a405f571f642589e7556b20551f09c44b346a2d16174ccaaba64.jpg)
36
+ Figure 1: The framework of our proposed residual non-local attention network for image restoration. ‘Conv’, ‘RNAB’, and ‘RAB’ denote convolutional layer, residual non-local attention block, and residual local attention block respectively. Here, we take image denoising as a task of interest.
37
+
38
+ # 3 RESIDUAL NON-LOCAL ATTENTION NETWORK FOR IMAGE RESTORATION
39
+
40
+ # 3.1 FRAMEWORK
41
+
42
+ The framework of our proposed residual non-local attention network (RNAN) is shown in Figure 1. Let’s denote $I _ { L }$ and $I _ { H }$ as the low-quality (e.g., noisy, blurred, or compressed images) and highquality images. The reconstructed image $I _ { R }$ can be obtained by
43
+
44
+ $$
45
+ I _ { R } = H _ { R N A N } \left( I _ { L } \right) ,
46
+ $$
47
+
48
+ where $H _ { R N A N }$ denotes the function of our proposed RNAN. With the usage of global residual learning in pixel space, the main part of our network can concentrate on learning the degradation components (e.g., noisy, blurring, or compressed artifacts).
49
+
50
+ The first and last convolutional layers are shallow feature extractor and reconstruction layer respectively. We propose residual local and non-local attention blocks to extract hierarchical attentionaware features. In addition to making the main network learn degradation components, we further concentrate on more challenging areas by using local and non-local attention. We only incorporate residual non-local attention block in low-level and high-level feature space. This is mainly because a few non-local modules can well offer non-local ability to the network for image restoration.
51
+
52
+ Then RNAN is optimized with loss function. Several loss functions have been investigated, such as $L _ { 2 }$ (Mao et al., 2016; Zhang et al., 2017a; Tai et al., 2017; Zhang et al., 2017b), $L _ { 1 }$ (Lim et al., 2017; Zhang et al., 2018c), perceptual and adversarial losses (Ledig et al., 2017). To show the effectiveness of our RNAN, we choosprevious works. Given a training set the same loss fu, which contains ionlo .g., qua $L _ { 2 }$ loss function) as inputs and their $\left\{ I _ { L } ^ { i } , I _ { H } ^ { i } \right\} _ { i = 1 } ^ { N }$ $N$ high-quality counterparts. The goal of training RNAN is to minimize the $L _ { 2 }$ loss function
53
+
54
+ $$
55
+ L \left( \Theta \right) = \frac { 1 } { N } \sum _ { i = 1 } ^ { N } \left| \left| H _ { R N A N } \left( I _ { L } ^ { i } \right) - I _ { H } ^ { i } \right| \right| _ { 2 } ,
56
+ $$
57
+
58
+ where $\lVert \cdot \rVert _ { 2 }$ denotes $l _ { 2 }$ norm. As detailed in Section 4, we use the same loss function as that in other compared methods. Such choice makes it clearer and more fair to see the effectiveness of our proposed RNAN. Then we give more details to residual local and non-local attention blocks.
59
+
60
+ ![](images/494e6b27464158d51e705512b95e910603ff6e61ef0a304f1aa91a5f2fc2cb70.jpg)
61
+ Figure 2: Residual (non-)local attention block. It mainly consists of trunk branch (labelled with gray dashed) and mask branch (labelled with red dashed). The trunk branch consists of $t$ RBs. The mask branch is used to learning mixed attention maps in channel- and spatial-wise simultaneously.
62
+
63
+ # 3.2 RESIDUAL NON-LOCAL ATTENTION BLOCK
64
+
65
+ Our residual non-local attention network is constructed by stacking several residual local and nonlocal attention blocks shown in Figure 2. Each attention block is divided into two parts: $q$ residual blocks (RBs) in the beginning and end of attention block. Two branches in the middle part: trunk branch and mask branch. For non-local attention block, we incorporate non-local block (NLB) in the mask branch, resulting non-local attention. Then we give more details to those components.
66
+
67
+ # 3.2.1 TRUNK BRANCH
68
+
69
+ As shown in Figure 2, the trunk branch includes $t$ residual blocks (RBs). Different from the original residual block in ResNet (He et al., 2016), we adopt the simplified RB from (Lim et al., 2017). The simplified RB (labelled with blue dashed) only consists of two convolutional layers and one ReLU (Nair & Hinton, 2010), omitting unnecessary components, such as maxpooling and batch normalization (Ioffe & Szegedy, 2015) layers. We find that such simplified RB not only contributes to image super-resolution (Lim et al., 2017), but also helps to construct very deep network for other image restoration tasks.
70
+
71
+ Feature maps from trunk branch of different depths serve as hierarchical features. If attention mechanism is not considered, the proposed network would become a simplified ResNet. With mask branch, we can take channel and spatial attention to adaptively rescale hierarchical features. Then we give more details about local and non-local attention.
72
+
73
+ # 3.2.2 MASK BRANCH
74
+
75
+ As labelled with red dashed in Figure 2, the mask branches used in our network include local and non-local ones. Here, we mainly focus on local mask branch, which can become a non-local one by using non-local block (NLB, labelled with green dashed arrow).
76
+
77
+ The key point in mask branch is how to grasp information of larger scope, namely larger receptive field size, so that it’s possible to obtain more sophisticated attention map. One possible solution is to perform maxpooling several times, as used in (Wang et al., 2017) for image classification. However, more pixel-level accurate results are desired in image restoration. Maxpooling would lose lots of details of the image, resulting in bad performance. To alleviate such drawbacks, we choose to use large-stride convolution and deconvolution to enlarge receptive field size. Another way is considering non-local information across the whole inputs, which will be discussed in the next subsection.
78
+
79
+ From the input, large-stride (stride $\geq 2$ ) convolutional layer increases the receptive field size after $m$ RBs. After additional $2 m$ RBs, the downscaled feature maps are then expanded by a deconvolutional layer (also known as transposed convolutional layer). The upscaled features are further forwarded through $m$ RBs and one $1 \times 1$ convolutional layer. Then a sigmoid layer normalizes the output values, ranging in $[ 0 , 1 ]$ . Although the receptive field size of the mask branch is much larger than that of the trunk branch, it cannot cover the whole features at a time. This can be achieved by using non-local block (NLB), resulting in non-local mixed attention.
80
+
81
+ ![](images/365a1350703ea1f5655a316d99e4c086706111a79546eb9eace6e96051f79ee6.jpg)
82
+ Figure 3: Non-local block. Red dashed line denotes matrix reshaping. $H \times W \times C$ means $C$ features with height $H$ and width W. $\otimes$ denotes matrix multiplication. $\oplus$ denotes element-wise addition.
83
+
84
+ # 3.2.3 NON-LOCAL MIXED ATTENTION
85
+
86
+ As discussed above, convolution operation processes one local neighbourhood at a time. In order to obtain better attention maps, here we seek to take all the positions into consideration at a time. Inspired by classical non-local means method (Buades et al., 2005) and non-local neural networks (Wang et al., 2018a), we incorporate non-local block (NLB) into the mask branch to obtain non-local mixed attention (shown in Figure 3). The non-local operation can be defined as
87
+
88
+ $$
89
+ y _ { i } = \left( \sum _ { \forall j } f ( x _ { i } , x _ { j } ) g ( x _ { j } ) \right) / \sum _ { \forall j } f ( x _ { i } , x _ { j } ) ,
90
+ $$
91
+
92
+ where $i$ is the output feature position index and $j$ is the index that enumerates all possible positions. $x$ and $y$ are the input and output of non-local operation. The pairwise function $f ( x _ { i } , x _ { j } )$ computes relationship between $x _ { i }$ and $x _ { j }$ . The function $g ( x _ { j } )$ computes a representation of the input at the position $j$ .
93
+
94
+ As shown in Figure 3, we use embedded Gaussian function to evaluate the pairwise relationship
95
+
96
+ $$
97
+ f ( x _ { i } , x _ { j } ) = e x p ( u ( x _ { i } ) ^ { T } v ( x _ { j } ) ) = e x p ( ( W _ { u } x _ { i } ) ^ { T } W _ { v } x _ { j } ) ,
98
+ $$
99
+
100
+ where $W _ { u }$ and $W _ { v }$ are weight matrices. As investigated in (Wang et al., 2018a), there are several versions of $f$ , such as Gaussian function, dot product similarity, and feature concatenation. We also consider a linear embedding for $g$ : $g ( x _ { j } ) = W _ { g } x _ { j }$ with weight matrix $W _ { g }$ . Then the output $z$ at position $i$ of non-local block (NLB) is calculated as
101
+
102
+ $$
103
+ z _ { i } = W _ { z } y _ { i } + x _ { i } = W _ { z } s o f t m a x ( ( W _ { u } x _ { i } ) ^ { T } W _ { v } x _ { j } ) g ( x _ { j } ) + x _ { i } ,
104
+ $$
105
+
106
+ where $W _ { z }$ is a weight matrix. For a given $i$ , $\textstyle \sum _ { \forall j } f ( x _ { i } , x _ { j } ) / \sum _ { \forall j } f ( x _ { i } , x _ { j } )$ in Eq. 3 becomes the softmax computation along dimension $j$ . The residual connection allows us to insert the NLB into pretrained networks (Wang et al., 2018a) by initializing $W _ { z }$ as zero.
107
+
108
+ With non-local and local attention computation, feature maps in the mask branch are finally mapped by sigmoid function
109
+
110
+ $$
111
+ f _ { m i x } ( x _ { i , c } ) = \frac { 1 } { 1 + e x p ( - x _ { i , c } ) } ,
112
+ $$
113
+
114
+ where $i$ ranges over spatial positions and $c$ ranges over feature channel positions. Such simple sigmoid operation is applied to each channel and spatial position, resulting mixed attention (Wang et al., 2017). As a result, the mask branch with non-local block can produce non-local mixed attention. However, simple multiplication between features from trunk and mask branches is not powerful enough or proper to form very deep trainable network. We propose residual non-local attention learning to solve those problems.
115
+
116
+ # 3.3 RESIDUAL NON-LOCAL ATTENTION LEARNING
117
+
118
+ How to train deep image restoration network with non-local mixed attention remains unclear. Here we only consider the trunk and mask branches, and residual connection with them (Figure 2). We focus on obtaining non-local attention information from the input feature $x$ . It should be noted that one form of attention residual learning was proposed in Wang et al. (2017), whose formulation is
119
+
120
+ $$
121
+ H _ { R N A } ( x ) = H _ { t r u n k } ( x ) ( H _ { m a s k } ( x ) + 1 ) .
122
+ $$
123
+
124
+ We find that this form of attention learning is not suitable for image restoration tasks. This is mainly because Eq. 7 is more suitable for high-level vision tasks (e.g., image classification), where low-level features are not preserved too much. However, low-level features are more important for image restoration. As a result, we propose a simple yet more suitable residual attention learning method by introducing input feature $x$ directly. We compute its output $H _ { R N A } ( x )$ as
125
+
126
+ $$
127
+ H _ { R N A } ( x ) = H _ { t r u n k } ( x ) H _ { m a s k } ( x ) + x ,
128
+ $$
129
+
130
+ where $H _ { t r u n k } ( x )$ and $H _ { m a s k } ( x )$ denote the functions of trunk and mask branches respectively. Such residual learning tends to preserve more low-level features and allows us to form very deep networks for high-quality image restoration tasks with stronger representation ability.
131
+
132
+ # 3.4 IMPLEMENTATION DETAILS
133
+
134
+ Now, we specify the implementation details of our proposed RNAN. We use 10 residual local and non-local attention blocks (2 non-local one). In each residual (non-)local block, we set $q , t , m =$ $2 , 2 , 1$ . We set $3 \times 3$ as the size of all convolutional layers except for those in non-local block and convolutional layer before sigmoid function, where the kernel size is $1 \times 1$ . Features in RBs have 64 filters, except for that in the non-local block (see Figure 3), where $C = 3 2$ . In each training batch, 16 low-quality (LQ) patches with the size of $4 8 \times 4 8$ are extracted as inputs. Our model is trained by ADAM optimizer (Kingma & Ba, 2014) with $\beta _ { 1 } = 0 . 9$ , $\beta _ { 2 } = 0 . 9 9 9$ , and $\epsilon = 1 0 ^ { - 8 }$ . The initial learning rate is set to $1 0 ^ { - 4 }$ and then decreases to half every $2 \times 1 0 ^ { 5 }$ iterations of back-propagation. We use PyTorch (Paszke et al., 2017) to implement our models with a Titan $\mathrm { X p }$ GPU.
135
+
136
+ # 4 EXPERIMENTS
137
+
138
+ We apply our proposed RNAN to three classical image restoration tasks: image denoising, demosaicing, and compression artifacts reduction. For image denoising and demosaicing, we follow the same setting as IRCNN (Zhang et al., 2017b). For image compression artifacts reduction, we follow the same setting as ARCNN (Dong et al., 2015). We use 800 training images in DIV2K (Timofte et al., 2017; Agustsson & Timofte, 2017) to train all of our models. For each task, we use commonly used dataset for testing and report PSNR and/or SSIM (Wang et al., 2004) to evaluate the results of each method. More results are shown in Appendix A.
139
+
140
+ Table 1: Ablation study of different components. PSNR values are based on Urban100 $\scriptstyle { \overbrace { \boldsymbol { \sigma } } = 3 0 } $ .
141
+
142
+ <table><tr><td>Case Index</td><td>1</td><td>2</td><td>3</td><td>4</td><td>5</td><td>6</td><td>7</td><td>8</td></tr><tr><td>Mask Branch</td><td>X</td><td>X</td><td>X</td><td>?</td><td>√</td><td>√</td><td>?</td><td>√</td></tr><tr><td>Non-local Block</td><td>×</td><td></td><td>√</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>RAB Number</td><td>7</td><td>7</td><td>5</td><td>5</td><td>1</td><td>2</td><td>5</td><td>8</td></tr><tr><td>RNABNumber</td><td>0</td><td>0</td><td>2</td><td>2</td><td>1</td><td>2</td><td>1</td><td>2</td></tr><tr><td>PSNR (dB)</td><td>30.96</td><td>31.17</td><td>31.20</td><td>31.32</td><td>30.78</td><td>30.99</td><td>31.27</td><td>31.50</td></tr></table>
143
+
144
+ # 4.1 ABLATION STUDY
145
+
146
+ We show ablation study in Table 1 to investigate the effects of different components in RNAN.
147
+
148
+ Non-local Mixed Attention. In cases 1, all the mask branches and non-local blocks are removed. In case 4, we enable non-local mixed attention with same block number as in case 1. The positive effect of non-local mixed attention is demonstrated by its obvious performance improvement.
149
+
150
+ Mask branch. We also learn that mask branch contributes to performance improvement, no matter non-local blocks are used (cases 3 and 4) or not (cases 1 and 2). This’s mainly because mask branch provides informative attention to the network, gaining better representational ability.
151
+
152
+ Non-local block. Non-local block also contributes to the network ability obviously, no matter mask branches are used (cases 2 and 4) or not (cases 1 and 3). With non-local information from low-level and high-level features, RNAN performs better image restoration.
153
+
154
+ Block Number. When comparing cases 2, 4, and 7, we learn that more non-local blocks achieve better results. However, the introduction of non-local block consumes much time. So we use 2 non-local blocks by considering low- and high-level features. When RNAB number is fixed in cases 5 and 7 or cases 6 and 8, performance also benefits from more RABs.
155
+
156
+ Table 2: Quantitative results about color image denoising. Best results are highlighted.
157
+
158
+ <table><tr><td rowspan=2 colspan=1>Method</td><td rowspan=1 colspan=4>Kodak24</td><td rowspan=1 colspan=4>BSD68</td><td rowspan=1 colspan=4>Urban100</td></tr><tr><td rowspan=1 colspan=1>10</td><td rowspan=1 colspan=1>30</td><td rowspan=1 colspan=1>50</td><td rowspan=1 colspan=1>70</td><td rowspan=1 colspan=1>10</td><td rowspan=1 colspan=1>30</td><td rowspan=1 colspan=1>50</td><td rowspan=1 colspan=1>70</td><td rowspan=1 colspan=1>10</td><td rowspan=1 colspan=1>30</td><td rowspan=1 colspan=1>50</td><td rowspan=1 colspan=1>70</td></tr><tr><td rowspan=1 colspan=1>CBM3D</td><td rowspan=1 colspan=1>36.57</td><td rowspan=1 colspan=1>30.89</td><td rowspan=1 colspan=1>28.63</td><td rowspan=1 colspan=1>27.27</td><td rowspan=1 colspan=1>35.91</td><td rowspan=1 colspan=1>29.73</td><td rowspan=1 colspan=1>27.38</td><td rowspan=1 colspan=1>26.00</td><td rowspan=1 colspan=1>36.00</td><td rowspan=1 colspan=1>30.36</td><td rowspan=1 colspan=1>27.94</td><td rowspan=1 colspan=1>26.31</td></tr><tr><td rowspan=1 colspan=1>TNRD</td><td rowspan=1 colspan=1>34.33</td><td rowspan=1 colspan=1>28.83</td><td rowspan=1 colspan=1>27.17</td><td rowspan=1 colspan=1>24.94</td><td rowspan=1 colspan=1>33.36</td><td rowspan=1 colspan=1>27.64</td><td rowspan=1 colspan=1>25.96</td><td rowspan=1 colspan=1>23.83</td><td rowspan=1 colspan=1>33.60</td><td rowspan=1 colspan=1>27.40</td><td rowspan=1 colspan=1>25.52</td><td rowspan=1 colspan=1>22.63</td></tr><tr><td rowspan=1 colspan=1>RED</td><td rowspan=1 colspan=1>34.91</td><td rowspan=1 colspan=1>29.71</td><td rowspan=1 colspan=1>27.62</td><td rowspan=1 colspan=1>26.36</td><td rowspan=1 colspan=1>33.89</td><td rowspan=1 colspan=1>28.46</td><td rowspan=1 colspan=1>26.35</td><td rowspan=1 colspan=1>25.09</td><td rowspan=1 colspan=1>34.59</td><td rowspan=1 colspan=1>29.02</td><td rowspan=1 colspan=1>26.40</td><td rowspan=1 colspan=1>24.74</td></tr><tr><td rowspan=1 colspan=1>DnCNN</td><td rowspan=1 colspan=1>36.98</td><td rowspan=1 colspan=1>31.39</td><td rowspan=1 colspan=1>29.16</td><td rowspan=1 colspan=1>27.64</td><td rowspan=1 colspan=1>36.31</td><td rowspan=1 colspan=1>30.40</td><td rowspan=1 colspan=1>28.01</td><td rowspan=1 colspan=1>26.56</td><td rowspan=1 colspan=1>36.21</td><td rowspan=1 colspan=1>30.28</td><td rowspan=1 colspan=1>28.16</td><td rowspan=1 colspan=1>26.17</td></tr><tr><td rowspan=1 colspan=1>MemNet</td><td rowspan=1 colspan=1>N/A</td><td rowspan=1 colspan=1>29.67</td><td rowspan=1 colspan=1>27.65</td><td rowspan=1 colspan=1>26.40</td><td rowspan=1 colspan=1>N/A</td><td rowspan=1 colspan=1>28.39</td><td rowspan=1 colspan=1>26.33</td><td rowspan=1 colspan=1>25.08</td><td rowspan=1 colspan=1>N/A</td><td rowspan=1 colspan=1>28.93</td><td rowspan=1 colspan=1>26.53</td><td rowspan=1 colspan=1>24.93</td></tr><tr><td rowspan=1 colspan=1>IRCNN</td><td rowspan=1 colspan=1>36.70</td><td rowspan=1 colspan=1>31.24</td><td rowspan=1 colspan=1>28.93</td><td rowspan=1 colspan=1>N/A</td><td rowspan=1 colspan=1>36.06</td><td rowspan=1 colspan=1>30.22</td><td rowspan=1 colspan=1>27.86</td><td rowspan=1 colspan=1>N/A</td><td rowspan=1 colspan=1>35.81</td><td rowspan=1 colspan=1>30.28</td><td rowspan=1 colspan=1>27.69</td><td rowspan=1 colspan=1>N/A</td></tr><tr><td rowspan=1 colspan=1>FFDNet</td><td rowspan=1 colspan=1>36.81</td><td rowspan=1 colspan=1>31.39</td><td rowspan=1 colspan=1>29.10</td><td rowspan=1 colspan=1>27.68</td><td rowspan=1 colspan=1>36.14</td><td rowspan=1 colspan=1>30.31</td><td rowspan=1 colspan=1>27.96</td><td rowspan=1 colspan=1>26.53</td><td rowspan=1 colspan=1>35.77</td><td rowspan=1 colspan=1>30.53</td><td rowspan=1 colspan=1>28.05</td><td rowspan=1 colspan=1>26.39</td></tr><tr><td rowspan=1 colspan=1>RNAN (ours)</td><td rowspan=1 colspan=1>37.24</td><td rowspan=1 colspan=1>31.86</td><td rowspan=1 colspan=1>29.58</td><td rowspan=1 colspan=1>28.16</td><td rowspan=1 colspan=1>36.43</td><td rowspan=1 colspan=1>30.63</td><td rowspan=1 colspan=1>28.27</td><td rowspan=1 colspan=1>26.83</td><td rowspan=1 colspan=1>36.59</td><td rowspan=1 colspan=1>31.50</td><td rowspan=1 colspan=1>29.08</td><td rowspan=1 colspan=1>27.45</td></tr></table>
159
+
160
+ Table 3: Quantitative results about gray-scale image denoising. Best results are highlighted.
161
+
162
+ <table><tr><td rowspan=2 colspan=1>Method</td><td rowspan=1 colspan=4>Kodak24</td><td rowspan=1 colspan=4>BSD68</td><td rowspan=1 colspan=4>Urban100</td></tr><tr><td rowspan=1 colspan=1>10</td><td rowspan=1 colspan=1>30</td><td rowspan=1 colspan=1>50</td><td rowspan=1 colspan=1>70</td><td rowspan=1 colspan=1>10</td><td rowspan=1 colspan=1>30</td><td rowspan=1 colspan=1>50</td><td rowspan=1 colspan=1>70</td><td rowspan=1 colspan=1>10</td><td rowspan=1 colspan=1>30</td><td rowspan=1 colspan=1>50</td><td rowspan=1 colspan=1>70</td></tr><tr><td rowspan=1 colspan=1>BM3D</td><td rowspan=1 colspan=1>34.39</td><td rowspan=1 colspan=1>29.13</td><td rowspan=1 colspan=1>26.99</td><td rowspan=1 colspan=1>25.73</td><td rowspan=1 colspan=1>33.31</td><td rowspan=1 colspan=1>27.76</td><td rowspan=1 colspan=1>25.62</td><td rowspan=1 colspan=1>24.44</td><td rowspan=1 colspan=1>34.47</td><td rowspan=1 colspan=1>28.75</td><td rowspan=1 colspan=1>25.94</td><td rowspan=1 colspan=1>24.27</td></tr><tr><td rowspan=1 colspan=1>TNRD</td><td rowspan=1 colspan=1>34.41</td><td rowspan=1 colspan=1>28.87</td><td rowspan=1 colspan=1>27.20</td><td rowspan=1 colspan=1>24.95</td><td rowspan=1 colspan=1>33.41</td><td rowspan=1 colspan=1>27.66</td><td rowspan=1 colspan=1>25.97</td><td rowspan=1 colspan=1>23.83</td><td rowspan=1 colspan=1>33.78</td><td rowspan=1 colspan=1>27.49</td><td rowspan=1 colspan=1>25.59</td><td rowspan=1 colspan=1>22.67</td></tr><tr><td rowspan=1 colspan=1>RED</td><td rowspan=1 colspan=1>35.02</td><td rowspan=1 colspan=1>29.77</td><td rowspan=1 colspan=1>27.66</td><td rowspan=1 colspan=1>26.39</td><td rowspan=1 colspan=1>33.99</td><td rowspan=1 colspan=1>28.50</td><td rowspan=1 colspan=1>26.37</td><td rowspan=1 colspan=1>25.10</td><td rowspan=1 colspan=1>34.91</td><td rowspan=1 colspan=1>29.18</td><td rowspan=1 colspan=1>26.51</td><td rowspan=1 colspan=1>24.82</td></tr><tr><td rowspan=1 colspan=1>DnCNN</td><td rowspan=1 colspan=1>34.90</td><td rowspan=1 colspan=1>29.62</td><td rowspan=1 colspan=1>27.51</td><td rowspan=1 colspan=1>26.08</td><td rowspan=1 colspan=1>33.88</td><td rowspan=1 colspan=1>28.36</td><td rowspan=1 colspan=1>26.23</td><td rowspan=1 colspan=1>24.90</td><td rowspan=1 colspan=1>34.73</td><td rowspan=1 colspan=1>28.88</td><td rowspan=1 colspan=1>26.28</td><td rowspan=1 colspan=1>24.36</td></tr><tr><td rowspan=1 colspan=1>MemNet</td><td rowspan=1 colspan=1>N/A</td><td rowspan=1 colspan=1>29.72</td><td rowspan=1 colspan=1>27.68</td><td rowspan=1 colspan=1>26.42</td><td rowspan=1 colspan=1>N/A</td><td rowspan=1 colspan=1>28.43</td><td rowspan=1 colspan=1>26.35</td><td rowspan=1 colspan=1>25.09</td><td rowspan=1 colspan=1>N/A</td><td rowspan=1 colspan=1>29.10</td><td rowspan=1 colspan=1>26.65</td><td rowspan=1 colspan=1>25.01</td></tr><tr><td rowspan=1 colspan=1>IRCNN</td><td rowspan=1 colspan=1>34.76</td><td rowspan=1 colspan=1>29.53</td><td rowspan=1 colspan=1>27.45</td><td rowspan=1 colspan=1>N/A</td><td rowspan=1 colspan=1>33.74</td><td rowspan=1 colspan=1>28.26</td><td rowspan=1 colspan=1>26.15</td><td rowspan=1 colspan=1>N/A</td><td rowspan=1 colspan=1>34.60</td><td rowspan=1 colspan=1>28.85</td><td rowspan=1 colspan=1>26.24</td><td rowspan=1 colspan=1>N/A</td></tr><tr><td rowspan=1 colspan=1>FFDNet</td><td rowspan=1 colspan=1>34.81</td><td rowspan=1 colspan=1>29.70</td><td rowspan=1 colspan=1>27.63</td><td rowspan=1 colspan=1>26.34</td><td rowspan=1 colspan=1>33.76</td><td rowspan=1 colspan=1>28.39</td><td rowspan=1 colspan=1>26.29</td><td rowspan=1 colspan=1>25.04</td><td rowspan=1 colspan=1>34.45</td><td rowspan=1 colspan=1>29.03</td><td rowspan=1 colspan=1>26.52</td><td rowspan=1 colspan=1>24.86</td></tr><tr><td rowspan=1 colspan=1>RNAN (ours)</td><td rowspan=1 colspan=1>35.20</td><td rowspan=1 colspan=1>30.04</td><td rowspan=1 colspan=1>27.93</td><td rowspan=1 colspan=1>26.60</td><td rowspan=1 colspan=1>34.04</td><td rowspan=1 colspan=1>28.61</td><td rowspan=1 colspan=1>26.48</td><td rowspan=1 colspan=1>25.18</td><td rowspan=1 colspan=1>35.52</td><td rowspan=1 colspan=1>30.20</td><td rowspan=1 colspan=1>27.65</td><td rowspan=1 colspan=1>25.89</td></tr></table>
163
+
164
+ # 4.2 COLOR AND GRAY IMAGE DENOISING
165
+
166
+ We compare our RNAN with state-of-the-art denoising methods: BM3D (Dabov et al., 2007b), CBM3D (Dabov et al., 2007a), TNRD (Chen & Pock, 2017), RED (Mao et al., 2016), DnCNN (Zhang et al., 2017a), MemNet (Tai et al., 2017), IRCNN (Zhang et al., 2017b), and FFDNet (Zhang et al., 2017c). Kodak24 (http://r0k.us/graphics/kodak/), BSD68 (Martin et al., 2001), and Urban100 (Huang et al., 2015) are used for color and gray-scale image denoising. AWGN noises of different levels (e.g., 10, 30, 50, and 70) are added to clean images.
167
+
168
+ Quantitative results are shown in Tables 2 and 3. As we can see that our proposed RNAN achieves the best results on all the datasets with all noise levels. Our proposed non-local attention covers the information from the whole image, which should be effective for heavy image denoising. To demonstrate this analysis, we take noise level $\sigma = 7 0$ as an example. We can see that our proposed RNAN achieves 0.48, 0.30, and 1.06 dB PSNR gains over the second best method FFDNet. This comparison strongly shows the effectiveness of our proposed non-local mixed attention.
169
+
170
+ We also show visual results in Figures 4 and 5. With the learned non-local mixed attention, RNAN treats different image parts distinctively, alleviating over-smoothing artifacts obviously.
171
+
172
+ Table 4: Quantitative results about color image demosaicing. Best results are highlighted.
173
+
174
+ <table><tr><td rowspan=2 colspan=1>Method</td><td rowspan=1 colspan=2>McMaster18</td><td rowspan=1 colspan=2>Kodak24</td><td rowspan=1 colspan=2>BSD68</td><td rowspan=1 colspan=2>Urban100</td></tr><tr><td rowspan=1 colspan=1>PSNR</td><td rowspan=1 colspan=1>SSIM</td><td rowspan=1 colspan=1>PSNR</td><td rowspan=1 colspan=1>SSIM</td><td rowspan=1 colspan=1>PSNR</td><td rowspan=1 colspan=1>SSIM</td><td rowspan=1 colspan=1>PSNR</td><td rowspan=1 colspan=1>SSIM</td></tr><tr><td rowspan=1 colspan=1>Mosaiced</td><td rowspan=1 colspan=1>9.17</td><td rowspan=1 colspan=1>0.1674</td><td rowspan=1 colspan=1>8.56</td><td rowspan=1 colspan=1>0.0682</td><td rowspan=1 colspan=1>8.43</td><td rowspan=1 colspan=1>0.0850</td><td rowspan=1 colspan=1>7.48</td><td rowspan=1 colspan=1>0.1195</td></tr><tr><td rowspan=1 colspan=1>IRCNN</td><td rowspan=1 colspan=1>37.47</td><td rowspan=1 colspan=1>0.9615</td><td rowspan=1 colspan=1>40.41</td><td rowspan=1 colspan=1>0.9807</td><td rowspan=1 colspan=1>39.96</td><td rowspan=1 colspan=1>0.9850</td><td rowspan=1 colspan=1>36.64</td><td rowspan=1 colspan=1>0.9743</td></tr><tr><td rowspan=1 colspan=1>RNAN (ours)</td><td rowspan=1 colspan=1>39.71</td><td rowspan=1 colspan=1>0.9725</td><td rowspan=1 colspan=1>43.09</td><td rowspan=1 colspan=1>0.9902</td><td rowspan=1 colspan=1>42.50</td><td rowspan=1 colspan=1>0.9929</td><td rowspan=1 colspan=1>39.75</td><td rowspan=1 colspan=1>0.9848</td></tr></table>
175
+
176
+ # 4.3 IMAGE DEMOSAICING
177
+
178
+ Following the same setting in IRCNN (Zhang et al., 2017b), we compare image demosaicing results with those of IRCNN on McMaster (Zhang et al., 2017b), Kodak24, BSD68, and Urban100. Since IRCNN has been one of the best methods for image demosaicing and limited space, we only compare with IRCNN in Table 4. As we can see, mosaiced images have very poor quality, resulting in very low PSNR and SSIM values. IRCNN can enhance the low-quality images and achieve relatively high values of PSNR and SSIM. Our RNAN can still make significant improvements over IRCNN. Using local and non-local attention, our RNAN can better handle the degradation situation.
179
+
180
+ Visual results are shown in Figure 6. Although IRCNN can remove mosaicing effect greatly, there’re still some artifacts in its results (e.g., blocking artifacts in ‘img 026’). However, RNAN recovers more faithful color and alliciates blocking artifacts.
181
+
182
+ # 4.4 IMAGE COMPRESSION ARTIFACTS REDUCTION
183
+
184
+ We further apply our RNAN to reduce image compression artifacts. We compare our RNAN with SA-DCT (Foi et al., 2007), ARCNN (Dong et al., 2015), TNRD (Chen & Pock, 2017), and DnCNN (Zhang et al., 2017a). We apply the standard JPEG compression scheme to obtain the compressed images by following (Dong et al., 2015). Four JPEG quality settings $q = 1 0$ , 20, 30, 40 are used in Matlab JPEG encoder. Here, we only focus on the restoration of $\mathrm { Y }$ channel (in YCbCr space) to keep fair comparison with other methods. We use the same datasets LIVE1 (Sheikh et al., 2005) and Classic5 (Foi et al., 2007) in ARCNN and report PSNR/SSIM values in Table 5. As we can see, our RNAN achieves the best PSNR and SSIM values on LIVE1 and Classic5 with all JPEG qualities.
185
+
186
+ We further shown visual comparisons in Figure 7. We provide comparisons under very low image quality $\scriptstyle ( q = 1 0 )$ . The blocking artifacts can be removed to some degree, but ARCNN, TNRD, and DnCNN would also over-smooth some structures. RNAN obtains more details with consistent structures by considering non-local mixed attention.
187
+
188
+ Table 5: Quantitative results about compression artifacts reduction. Best results are highlighted.
189
+
190
+ <table><tr><td rowspan=2 colspan=1>Dataset</td><td rowspan=2 colspan=1>q</td><td rowspan=1 colspan=2>JPEG</td><td rowspan=1 colspan=3>SA-DCT</td><td rowspan=1 colspan=2>ARCNN</td><td rowspan=1 colspan=2>TNRD</td><td rowspan=1 colspan=2>DnCNN</td><td rowspan=1 colspan=1>RNAN</td><td rowspan=1 colspan=1>(ours)</td></tr><tr><td rowspan=1 colspan=1>PSNR</td><td rowspan=1 colspan=1>SSIM</td><td rowspan=1 colspan=2>PSNR</td><td rowspan=1 colspan=1>SSIM</td><td rowspan=1 colspan=1>PSNR</td><td rowspan=1 colspan=1>SSIM</td><td rowspan=1 colspan=1>PSNR</td><td rowspan=1 colspan=1>SSIM</td><td rowspan=1 colspan=1>PSNR</td><td rowspan=1 colspan=1>SSIM</td><td rowspan=1 colspan=1>PSNR</td><td rowspan=1 colspan=1>SSIM</td></tr><tr><td rowspan=4 colspan=1>LIVE1</td><td rowspan=4 colspan=1>10203040</td><td rowspan=4 colspan=1>27.7730.0731.4132.35</td><td rowspan=4 colspan=1>0.79050.86830.90000.9173</td><td rowspan=3 colspan=2>28.6530.8132.08</td><td rowspan=1 colspan=1>0.8093</td><td rowspan=1 colspan=1>28.98</td><td rowspan=1 colspan=1>0.8217</td><td rowspan=2 colspan=1>29.1531.46</td><td rowspan=2 colspan=1>0.81110.8769</td><td rowspan=2 colspan=1>29.1931.59</td><td rowspan=2 colspan=1>0.81230.8802</td><td rowspan=2 colspan=1>29.6332.03</td><td rowspan=2 colspan=1>0.82390.8877</td></tr><tr><td rowspan=1 colspan=1>0.8781</td><td rowspan=1 colspan=1>31.29</td><td rowspan=1 colspan=1>0.8871</td></tr><tr><td rowspan=1 colspan=1>.08</td><td rowspan=1 colspan=1>0.9078</td><td rowspan=1 colspan=1>32.69</td><td rowspan=1 colspan=1>0.9166</td><td rowspan=1 colspan=1>32.84</td><td rowspan=1 colspan=1>0.9059</td><td rowspan=1 colspan=1>32.98</td><td rowspan=2 colspan=1>0.90900.9247</td><td rowspan=2 colspan=1>33.4534.47</td><td rowspan=2 colspan=1>0.91490.9299</td></tr><tr><td rowspan=1 colspan=2>32.99</td><td rowspan=1 colspan=1>0.9240</td><td rowspan=1 colspan=1>33.63</td><td rowspan=1 colspan=1>0.9306</td><td rowspan=1 colspan=1>N/A</td><td rowspan=1 colspan=1>N/A</td><td rowspan=1 colspan=1>33.96</td></tr><tr><td rowspan=3 colspan=1>Classic5</td><td rowspan=3 colspan=1>10203040</td><td rowspan=3 colspan=1>27.8230.1231.4832.43</td><td rowspan=3 colspan=1>0.78000.85410.88440.9011</td><td rowspan=2 colspan=2>28.8830.9232.14</td><td rowspan=1 colspan=1>0.80710.8663</td><td rowspan=1 colspan=1>29.0431.16</td><td rowspan=1 colspan=1>0.81110.8694</td><td rowspan=1 colspan=1>29.2831.47</td><td rowspan=1 colspan=1>0.79920.8576</td><td rowspan=1 colspan=1>29.4031.63</td><td rowspan=1 colspan=1>0.80260.8610</td><td rowspan=1 colspan=1>29.9632.11</td><td rowspan=3 colspan=1>0.81780.86930.89240.9061</td></tr><tr><td rowspan=1 colspan=1>0.8914</td><td rowspan=1 colspan=1>32.52</td><td rowspan=1 colspan=1>0.8967</td><td rowspan=1 colspan=1>32.78</td><td rowspan=1 colspan=1>0.8837</td><td rowspan=1 colspan=1>32.91</td><td rowspan=2 colspan=1>0.88610.9003</td><td rowspan=2 colspan=1>33.3834.27</td></tr><tr><td rowspan=1 colspan=2>33.00</td><td rowspan=1 colspan=1>0.9055</td><td rowspan=1 colspan=1>33.34</td><td rowspan=1 colspan=1>0.9101</td><td rowspan=1 colspan=1>N/A</td><td rowspan=1 colspan=1>N/A</td><td rowspan=1 colspan=1>33.77</td></tr></table>
191
+
192
+ # 4.5 IMAGE SUPER-RESOLUTION
193
+
194
+ We further compare our RNAN with state-of-the-art SR methods: EDSR (Lim et al., 2017), SRMDNF (Zhang et al., 2018a), D-DBPN (Haris et al., 2018), and RCAN (Zhang et al., 2018b). Similar to (Lim et al., 2017; Zhang et al., 2018c), we also introduce self-ensemble strategy to further improve our RNAN and denote the self-ensembled one as ${ \mathrm { R N A N } } +$ .
195
+
196
+ As shown in Table 6, our ${ \mathrm { R N A N } } +$ achieves the second best performance among benchmark datasets: Set5 (Bevilacqua et al., 2012), Set14 (Zeyde et al., 2010), B100 (Martin et al., 2001), Urban100 (Huang et al., 2015), and Manga109 (Matsui et al., 2017). Even without self-ensemble, our RNAN achieves third best results in most cases. Such improvements are notable, because the parameter number of RNAN is $7 . 5 { \bf M }$ , far smaller than $4 3 \mathrm { { M } }$ in EDSR and $1 6 \mathbf { M }$ in RCAN. The network depth of our RNAN (about 120 convolutional layers) is also far shallower than that of RCAN, which has about 400 convolutional layers. It indicates that non-local attention make better use of main network, saving much network parameter.
197
+
198
+ In Figure 8, we conduct image SR $( \times 4 )$ with several state-of-the-art methods. We can see that our RNAN obtains better visually pleasing results with finer structures. These comparisons further demonstrate the effectiveness of our proposed RNAN with the usage of non-local mixed attention.
199
+
200
+ Table 6: Quantitative image SR results. Best and second best results are highlighted and underlined
201
+
202
+ <table><tr><td rowspan=2 colspan=1>Method</td><td rowspan=2 colspan=1>Scale</td><td rowspan=1 colspan=2>Set5</td><td rowspan=1 colspan=2>Set14</td><td rowspan=1 colspan=1>B</td><td rowspan=1 colspan=1>00</td><td rowspan=1 colspan=2>Urban100</td><td rowspan=1 colspan=2>Manga109</td></tr><tr><td rowspan=1 colspan=1>PSNR</td><td rowspan=1 colspan=1>SSIM</td><td rowspan=1 colspan=1>PSNR</td><td rowspan=1 colspan=1>SSIM</td><td rowspan=1 colspan=1>PSNR</td><td rowspan=1 colspan=1>SSIM</td><td rowspan=1 colspan=1>PSNR</td><td rowspan=1 colspan=1>SSIM</td><td rowspan=1 colspan=1>PSNR</td><td rowspan=1 colspan=1>SSIM</td></tr><tr><td rowspan=6 colspan=1>BicubicEDSRSRMDNFD-DBPNRCANRNAN (ours)RNAN+ (ours)</td><td rowspan=2 colspan=1>×2×2</td><td rowspan=1 colspan=1>33.66</td><td rowspan=1 colspan=1>0.9299</td><td rowspan=1 colspan=1>30.24</td><td rowspan=1 colspan=1>0.8688</td><td rowspan=1 colspan=1>29.56</td><td rowspan=1 colspan=1>0.8431</td><td rowspan=1 colspan=1>26.88</td><td rowspan=1 colspan=1>0.8403</td><td rowspan=1 colspan=1>30.80</td><td rowspan=1 colspan=1>0.9339</td></tr><tr><td rowspan=1 colspan=1>38.11</td><td rowspan=1 colspan=1>0.9602</td><td rowspan=1 colspan=1>33.92</td><td rowspan=1 colspan=1>0.9195</td><td rowspan=1 colspan=1>32.32</td><td rowspan=1 colspan=1>0.9013</td><td rowspan=1 colspan=1>32.93</td><td rowspan=1 colspan=1>0.9351</td><td rowspan=1 colspan=1>39.10</td><td rowspan=1 colspan=1>0.9773</td></tr><tr><td rowspan=1 colspan=1>×2</td><td rowspan=1 colspan=1>37.79</td><td rowspan=1 colspan=1>0.9601</td><td rowspan=1 colspan=1>33.32</td><td rowspan=1 colspan=1>0.9159</td><td rowspan=1 colspan=1>32.05</td><td rowspan=1 colspan=1>0.8985</td><td rowspan=1 colspan=1>31.33</td><td rowspan=1 colspan=1>0.9204</td><td rowspan=1 colspan=1>38.07</td><td rowspan=1 colspan=1>0.9761</td></tr><tr><td rowspan=3 colspan=1>×2×2×2×2</td><td rowspan=1 colspan=1>38.09</td><td rowspan=1 colspan=1>0.9600</td><td rowspan=1 colspan=1>33.85</td><td rowspan=1 colspan=1>0.9190</td><td rowspan=1 colspan=1>32.27</td><td rowspan=1 colspan=1>0.9000</td><td rowspan=1 colspan=1>32.55</td><td rowspan=1 colspan=1>0.9324</td><td rowspan=1 colspan=1>38.89</td><td rowspan=1 colspan=1>0.9775</td></tr><tr><td rowspan=2 colspan=1>38.2738.1738.22</td><td rowspan=2 colspan=1>0.96140.96110.9613</td><td rowspan=1 colspan=1>34.12</td><td rowspan=1 colspan=1>0.9216</td><td rowspan=1 colspan=1>32.41</td><td rowspan=1 colspan=1>0.9027</td><td rowspan=1 colspan=1>33.34</td><td rowspan=1 colspan=1>0.9384</td><td rowspan=1 colspan=1>39.44</td><td rowspan=1 colspan=1>0.9786</td></tr><tr><td rowspan=1 colspan=1>33.8733.97</td><td rowspan=1 colspan=1>0.92070.9216</td><td rowspan=1 colspan=1>32.3232.36</td><td rowspan=1 colspan=1>0.90140.9018</td><td rowspan=1 colspan=1>32.7332.90</td><td rowspan=1 colspan=1>0.93400.9351</td><td rowspan=1 colspan=1>39.2339.41</td><td rowspan=1 colspan=1>0.97850.9789</td></tr><tr><td rowspan=7 colspan=1>BicubicEDSRSRMDNFD-DBPNRCANRNAN (ours)RNAN+ (ours)</td><td rowspan=2 colspan=1>×4×4</td><td rowspan=1 colspan=1>28.42</td><td rowspan=1 colspan=1>0.8104</td><td rowspan=1 colspan=1>26.00</td><td rowspan=1 colspan=1>0.7027</td><td rowspan=1 colspan=1>25.96</td><td rowspan=1 colspan=1>0.6675</td><td rowspan=1 colspan=1>23.14</td><td rowspan=1 colspan=1>0.6577</td><td rowspan=1 colspan=1>24.89</td><td rowspan=1 colspan=1>0.7866</td></tr><tr><td rowspan=1 colspan=1>32.46</td><td rowspan=1 colspan=1>0.8968</td><td rowspan=1 colspan=1>28.80</td><td rowspan=1 colspan=1>0.7876</td><td rowspan=1 colspan=1>27.71</td><td rowspan=1 colspan=1>0.7420</td><td rowspan=1 colspan=1>26.64</td><td rowspan=1 colspan=1>0.8033</td><td rowspan=1 colspan=1>31.02</td><td rowspan=1 colspan=1>0.9148</td></tr><tr><td rowspan=1 colspan=1>×4</td><td rowspan=1 colspan=1>31.96</td><td rowspan=1 colspan=1>0.8925</td><td rowspan=1 colspan=1>28.35</td><td rowspan=1 colspan=1>0.7787</td><td rowspan=1 colspan=1>27.49</td><td rowspan=1 colspan=1>0.7337</td><td rowspan=1 colspan=1>25.68</td><td rowspan=1 colspan=1>0.7731</td><td rowspan=1 colspan=1>30.09</td><td rowspan=1 colspan=1>0.9024</td></tr><tr><td rowspan=1 colspan=1>×4</td><td rowspan=1 colspan=1>32.47</td><td rowspan=1 colspan=1>0.8980</td><td rowspan=1 colspan=1>28.82</td><td rowspan=1 colspan=1>0.7860</td><td rowspan=1 colspan=1>27.72</td><td rowspan=1 colspan=1>0.7400</td><td rowspan=1 colspan=1>26.38</td><td rowspan=1 colspan=1>0.7946</td><td rowspan=1 colspan=1>30.91</td><td rowspan=1 colspan=1>0.9137</td></tr><tr><td rowspan=1 colspan=1>×4</td><td rowspan=1 colspan=1>32.63</td><td rowspan=1 colspan=1>0.9002</td><td rowspan=1 colspan=1>28.87</td><td rowspan=1 colspan=1>0.7889</td><td rowspan=1 colspan=1>27.77</td><td rowspan=1 colspan=1>0.7436</td><td rowspan=1 colspan=1>26.82</td><td rowspan=1 colspan=1>0.8087</td><td rowspan=1 colspan=1>31.22</td><td rowspan=1 colspan=1>0.9173</td></tr><tr><td rowspan=2 colspan=1>×4×4</td><td rowspan=1 colspan=1>32.49</td><td rowspan=1 colspan=1>0.8982</td><td rowspan=1 colspan=1>28.83</td><td rowspan=1 colspan=1>0.7878</td><td rowspan=1 colspan=1>27.72</td><td rowspan=1 colspan=1>0.7421</td><td rowspan=1 colspan=1>26.61</td><td rowspan=1 colspan=1>0.8023</td><td rowspan=2 colspan=1>31.0931.37</td><td rowspan=2 colspan=1>0.91490.9175</td></tr><tr><td rowspan=1 colspan=1>32.56</td><td rowspan=1 colspan=1>0.8992</td><td rowspan=1 colspan=1>28.90</td><td rowspan=1 colspan=1>0.7883</td><td rowspan=1 colspan=1>27.77</td><td rowspan=1 colspan=1>0.7424</td><td rowspan=1 colspan=1>26.75</td><td rowspan=1 colspan=1>0.8052</td></tr></table>
203
+
204
+ # 4.6 PARAMETERS AND RUNNING TIME ANALYSES
205
+
206
+ We also compare parameters, running time, and performance based on color image denoising in Table 7. PSNR values are tested on Urban100 $\scriptstyle ( \sigma = 5 0 )$ ). RNAN with 10 blocks achieves the best performance with the highest parameter number, which can be reduced to only 2 blocks and obtains second best performance. Here, we report running time for reference, because the time is related to implementation platform and code.
207
+
208
+ Table 7: Parameter and time comparisons. ‘\*’ means being applied channel-wise for color images.
209
+
210
+ <table><tr><td rowspan=1 colspan=1>Methods</td><td rowspan=1 colspan=1>RED*</td><td rowspan=1 colspan=1>DnCNN</td><td rowspan=1 colspan=1>MemNet*</td><td rowspan=1 colspan=1>RNAN (1LB+1NLB)</td><td rowspan=1 colspan=1>RNAN (8LBs+2NLBs)</td></tr><tr><td rowspan=1 colspan=1>Parameter Number</td><td rowspan=1 colspan=1>4,131 K</td><td rowspan=1 colspan=1>672K</td><td rowspan=1 colspan=1>677K</td><td rowspan=1 colspan=1>1,494k</td><td rowspan=1 colspan=1>7,409K</td></tr><tr><td rowspan=1 colspan=1>PSNR (dB)</td><td rowspan=1 colspan=1>26.40</td><td rowspan=1 colspan=1>28.16</td><td rowspan=1 colspan=1>26.53</td><td rowspan=1 colspan=1>28.36</td><td rowspan=1 colspan=1>29.15</td></tr><tr><td rowspan=1 colspan=1>Time (s) (Platform)</td><td rowspan=1 colspan=1>58.7(Caffe)</td><td rowspan=1 colspan=1>16.5(Matlab+GPU Array)</td><td rowspan=1 colspan=1>239.0 (Caffe)</td><td rowspan=1 colspan=1>2.6 (PyTorch)</td><td rowspan=1 colspan=1>6.5 (PyTorch)</td></tr></table>
211
+
212
+ # 5 CONCLUSIONS
213
+
214
+ In this paper, we propose residual non-local attention networks for high-quality image restoration. The networks are built by stacking local and non-local attention blocks, which extract local and non-local attention-aware features and consist of trunk and (non-)local mask branches. They’re used to extract hierarchical features and adaptively rescale hierarchical features with soft weights. We further generate non-local attention by considering the whole feature map. Furthermore, we propose residual local and non-local attention learning to train very deep networks. We introduce the input feature into attention computation, being more suitable for image restoration. RNAN achieves state-of-the-art image restoration results with moderate model size and running time.
215
+
216
+ Acknowledgements. This research is supported in part by the NSF IIS award 1651902 and U.S.
217
+ Army Research Office Award W911NF-17-1-0367.
218
+
219
+ ![](images/49e8dcc1bb12b52cfa87da7eaafe6ed65e0d5a7613c8327985c5ffa007e66516.jpg)
220
+ Figure 4: Color image denoising results with noise level $\sigma = 5 0$ .
221
+
222
+ ![](images/693050a598ad308005a9ec92c609b884160ae8b156f1293db9add7e2ccb65b0a.jpg)
223
+ Figure 5: Gray image denoising results with noise level $\sigma = 5 0$ .
224
+
225
+ ![](images/d13af9cec5ccfb201984522a8c167b26debf5a933879a1c5cfed49b692f3b7f2.jpg)
226
+ Figure 6: Image demosaicing results.
227
+
228
+ ![](images/97ce416f989d8996f5ef0ac8769757519389af1e5b149e87e177f373df5c12e3.jpg)
229
+ Figure 7: Image compression artifacts reduction results with JPEG quality $q = 1 0$ .
230
+
231
+ ![](images/9007fc5a051d4dc8bd191045ea3b8bf62613697c486b8c7a931948eef3246998.jpg)
232
+ Figure 8: Image super-resolution results with scaling factor $s = 4$ .
233
+
234
+ # REFERENCES
235
+
236
+ Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In CVPRW, 2017.
237
+
238
+ Cosmin Ancuti, Codruta O Ancuti, Radu Timofte, Luc Van Gool, Lei Zhang, Ming-Hsuan Yang, Vishal M Patel, He Zhang, Vishwanath A Sindagi, Ruhao Zhao, et al. Ntire 2018 challenge on image dehazing: Methods and results. In CVPRW, 2018.
239
+
240
+ Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Lowcomplexity single-image super-resolution based on nonnegative neighbor embedding. In BMVC, 2012.
241
+
242
+ Yochai Blau, Roey Mechrez, Radu Timofte, Tomer Michaeli, and Lihi Zelnik-Manor. 2018 pirm challenge on perceptual image super-resolution. In ECCVW, 2018.
243
+
244
+ Antoni Buades, Bartomeu Coll, and J-M Morel. A non-local algorithm for image denoising. In CVPR, 2005.
245
+
246
+ Yunjin Chen and Thomas Pock. Trainable nonlinear reaction diffusion: A flexible framework for fast and effective image restoration. TPAMI, 2017.
247
+
248
+ Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Color image denoising via sparse 3d collaborative filtering with grouping constraint in luminance-chrominance space. In ICIP, 2007a.
249
+
250
+ Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-d transform-domain collaborative filtering. TIP, 2007b.
251
+
252
+ Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009.
253
+
254
+ Chao Dong, Chen Change Loy, Kaiming He, and Xiaoou Tang. Learning a deep convolutional network for image super-resolution. In ECCV, 2014.
255
+
256
+ Chao Dong, Yubin Deng, Chen Change Loy, and Xiaoou Tang. Compression artifacts reduction by a deep convolutional network. In ICCV, 2015.
257
+
258
+ Chao Dong, Chen Change Loy, Kaiming He, and Xiaoou Tang. Image super-resolution using deep convolutional networks. TPAMI, 2016.
259
+
260
+ Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Pointwise shape-adaptive dct for highquality denoising and deblocking of grayscale and color images. TIP, May 2007.
261
+
262
+ Muhammad Haris, Greg Shakhnarovich, and Norimichi Ukita. Deep back-projection networks for super-resolution. In CVPR, 2018.
263
+
264
+ Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016.
265
+
266
+ Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. arXiv preprint arXiv:1709.01507, 2017.
267
+
268
+ Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In CVPR, 2015.
269
+
270
+ Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015.
271
+
272
+ Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In CVPR, 2016.
273
+
274
+ Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2014.
275
+
276
+ Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In CVPR, 2017.
277
+
278
+ Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro ´ Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. Photorealistic single image super-resolution using a generative adversarial network. In CVPR, 2017.
279
+
280
+ Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, 2017.
281
+
282
+ Ding Liu, Bihan Wen, Yuchen Fan, Chen Change Loy, and Thomas S Huang. Non-local recurrent network for image restoration. In NIPS, 2018.
283
+
284
+ Kede Ma, Zhengfang Duanmu, Qingbo Wu, Zhou Wang, Hongwei Yong, Hongliang Li, and Lei Zhang. Waterloo exploration database: New challenges for image quality assessment models. TIP, 2017.
285
+
286
+ Xiaojiao Mao, Chunhua Shen, and Yu-Bin Yang. Image restoration using very deep convolutional encoder-decoder networks with symmetric skip connections. In NIPS, 2016.
287
+
288
+ David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In ICCV, 2001.
289
+
290
+ Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multimedia Tools and Applications, 2017.
291
+
292
+ Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In ICML, 2010.
293
+
294
+ Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017.
295
+
296
+ Hamid R Sheikh, Zhou Wang, Lawrence Cormack, and Alan C Bovik. Live image quality assessment database release 2 (2005), 2005.
297
+
298
+ Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In ICCV, 2017.
299
+
300
+ Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, Lei Zhang, Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, Kyoung Mu Lee, et al. Ntire 2017 challenge on single image super-resolution: Methods and results. In CVPRW, 2017.
301
+
302
+ Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In ICML, 2008.
303
+
304
+ Fei Wang, Mengqing Jiang, Chen Qian, Shuo Yang, Cheng Li, Honggang Zhang, Xiaogang Wang, and Xiaoou Tang. Residual attention network for image classification. In CVPR, 2017.
305
+
306
+ Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In CVPR, 2018a.
307
+
308
+ Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Chen Change Loy, Yu Qiao, and Xiaoou Tang. Esrgan: Enhanced super-resolution generative adversarial networks. In ECCVW, 2018b.
309
+
310
+ Yifan Wang, Federico Perazzi, Brian McWilliams, Alexander Sorkine-Hornung, Olga SorkineHornung, and Christopher Schroers. A fully progressive approach to single-image superresolution. In CVPRW, 2018c.
311
+
312
+ Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. TIP, 2004.
313
+
314
+ Jianchao Yang, John Wright, Thomas Huang, and Yi Ma. Image super-resolution as sparse representation of raw image patches. In CVPR, 2008.
315
+
316
+ Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparserepresentations. In Proc. 7th Int. Conf. Curves Surf., 2010.
317
+ Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. TIP, 2017a.
318
+ Kai Zhang, Wangmeng Zuo, Shuhang Gu, and Lei Zhang. Learning deep cnn denoiser prior for image restoration. In CVPR, 2017b.
319
+ Kai Zhang, Wangmeng Zuo, and Lei Zhang. Ffdnet: Toward a fast and flexible solution for cnn based image denoising. arXiv preprint arXiv:1710.04026, 2017c.
320
+ Kai Zhang, Wangmeng Zuo, and Lei Zhang. Learning a single convolutional super-resolution network for multiple degradations. In CVPR, 2018a.
321
+ Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image superresolution using very deep residual channel attention networks. In ECCV, 2018b.
322
+ Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In CVPR, 2018c.
323
+
324
+ # A APPENDIX
325
+
326
+ # A.1 TRAIN RNAN WITH SMALL TRAINING DATA
327
+
328
+ All the results shown in the main paper are based on DIV2K training data. Here, we retrain our RNAN on small training sets for 5 tasks. In Table 8, for each task, we only refer the second best method from our main paper to compare. For training data, we use BSD400 (Martin et al., 2001) for color/gray-scale image denoising and demosaicing. We use 91 images in Yang et al. (2008) and 200 images in Martin et al. (2001) (denoted as SR291) for image compression artifacts reduction and super-resolution. FFDNet used BSD400 (Martin et al., 2001), 400 images from ImageNet Deng et al. (2009), and 4,744 images in Waterloo Exploration Database Ma et al. (2017). Here, $\mathrm { \cdot B S D 4 0 0 + } ^ { \mathrm { \cdot } }$ ’ is used to denote $\mathbf { \cdot B S D 4 0 0 + I m a g e N e t 4 0 0 + W E D 4 7 4 4 } ^ { \prime }$ . According to Table 8, we use the same or even smaller training set for our RNAN and obtain better results for 5 tasks. These experiments demonstrate the effectiveness of our RNAN for general image restoration tasks.
329
+
330
+ Table 8: Quantitative results about image restoration. Best results are highlighted.
331
+
332
+ <table><tr><td rowspan=1 colspan=14>Task1:Color image denoising.o denotes noise level.</td></tr><tr><td rowspan=2 colspan=1>Data</td><td rowspan=2 colspan=1>Method</td><td rowspan=1 colspan=4>Kodak24</td><td rowspan=1 colspan=4>BSD68</td><td rowspan=1 colspan=4>Urban100</td></tr><tr><td rowspan=1 colspan=1>g=10</td><td rowspan=1 colspan=1>=30</td><td rowspan=1 colspan=1>=50</td><td rowspan=1 colspan=1>g=70</td><td rowspan=1 colspan=1>g=10</td><td rowspan=1 colspan=1>=30</td><td rowspan=1 colspan=1>=50</td><td rowspan=1 colspan=1>=70</td><td rowspan=1 colspan=1>=10</td><td rowspan=1 colspan=1>=30</td><td rowspan=1 colspan=1>=50</td><td rowspan=1 colspan=1>g=70</td></tr><tr><td rowspan=1 colspan=1>BSD400+</td><td rowspan=1 colspan=1>FFDNet</td><td rowspan=1 colspan=1>36.81</td><td rowspan=1 colspan=1>31.39</td><td rowspan=1 colspan=1>29.10</td><td rowspan=1 colspan=1>27.68</td><td rowspan=1 colspan=1>36.14</td><td rowspan=1 colspan=1>30.31</td><td rowspan=1 colspan=1>27.96</td><td rowspan=1 colspan=1>26.53</td><td rowspan=1 colspan=1>35.77</td><td rowspan=1 colspan=1>30.53</td><td rowspan=1 colspan=1>28.05</td><td rowspan=1 colspan=1>26.39</td></tr><tr><td rowspan=1 colspan=1>BSD400</td><td rowspan=1 colspan=1>RNAN</td><td rowspan=1 colspan=1>37.16</td><td rowspan=1 colspan=1>31.81</td><td rowspan=1 colspan=1>29.55</td><td rowspan=1 colspan=1>28.15</td><td rowspan=1 colspan=1>36.60</td><td rowspan=1 colspan=1>30.73</td><td rowspan=1 colspan=1>28.35</td><td rowspan=1 colspan=1>26.88</td><td rowspan=1 colspan=1>36.39</td><td rowspan=1 colspan=1>30.90</td><td rowspan=1 colspan=1>28.65</td><td rowspan=1 colspan=1>27.11</td></tr><tr><td rowspan=1 colspan=14>Task 2: Gray image denoising. g denotes noise level.</td></tr><tr><td rowspan=2 colspan=1>Data</td><td rowspan=2 colspan=1>Method</td><td rowspan=1 colspan=4>Kodak24</td><td rowspan=1 colspan=4>BSD68</td><td rowspan=1 colspan=4>Urban100</td></tr><tr><td rowspan=1 colspan=1>g=10</td><td rowspan=1 colspan=1>g=30</td><td rowspan=1 colspan=1>=50</td><td rowspan=1 colspan=1>g=70</td><td rowspan=1 colspan=1>g=10</td><td rowspan=1 colspan=1>g=30</td><td rowspan=1 colspan=1>g=50</td><td rowspan=1 colspan=1>g=70</td><td rowspan=1 colspan=1>g=10</td><td rowspan=1 colspan=1>g=30</td><td rowspan=1 colspan=1>σ=50</td><td rowspan=1 colspan=1>g=70</td></tr><tr><td rowspan=1 colspan=1>BSD400+</td><td rowspan=1 colspan=1>FFDNet</td><td rowspan=1 colspan=1>34.81</td><td rowspan=1 colspan=1>29.70</td><td rowspan=1 colspan=1>27.63</td><td rowspan=1 colspan=1>26.34</td><td rowspan=1 colspan=1>33.76</td><td rowspan=1 colspan=1>28.39</td><td rowspan=1 colspan=1>26.29</td><td rowspan=1 colspan=1>25.04</td><td rowspan=1 colspan=1>34.45</td><td rowspan=1 colspan=1>29.03</td><td rowspan=1 colspan=1>26.52</td><td rowspan=1 colspan=1>24.86</td></tr><tr><td rowspan=1 colspan=1>BSD400</td><td rowspan=1 colspan=1>RNAN</td><td rowspan=1 colspan=1>35.15</td><td rowspan=1 colspan=1>29.87</td><td rowspan=1 colspan=1>27.92</td><td rowspan=1 colspan=1>26.50</td><td rowspan=1 colspan=1>34.18</td><td rowspan=1 colspan=1>28.72</td><td rowspan=1 colspan=1>26.61</td><td rowspan=1 colspan=1>25.34</td><td rowspan=1 colspan=1>35.08</td><td rowspan=1 colspan=1>29.82</td><td rowspan=1 colspan=1>27.30</td><td rowspan=1 colspan=1>25.57</td></tr><tr><td rowspan=1 colspan=6>Task3:Image demosaicing</td><td rowspan=1 colspan=3>Task4:Imagecompress</td><td rowspan=1 colspan=5>Task 4: Image compression artifacts reduction. q denotes JPEG quality.</td></tr><tr><td rowspan=2 colspan=1>Data</td><td rowspan=2 colspan=1>Method</td><td rowspan=2 colspan=2>Kodak24</td><td rowspan=2 colspan=2>BSD68</td><td rowspan=2 colspan=3>Data</td><td rowspan=2 colspan=1>Method</td><td rowspan=1 colspan=1>LI</td><td rowspan=1 colspan=1>E1</td><td rowspan=1 colspan=2>Classic5</td></tr><tr><td rowspan=1 colspan=1>q=20</td><td rowspan=1 colspan=1>q=40</td><td rowspan=1 colspan=1>q=20</td><td rowspan=1 colspan=1>q=40</td></tr><tr><td rowspan=1 colspan=1>BSD400</td><td rowspan=1 colspan=1>IRCNN</td><td rowspan=1 colspan=2>40.41</td><td rowspan=1 colspan=2>39.96</td><td rowspan=1 colspan=3>SR291</td><td rowspan=1 colspan=1>DnCNN</td><td rowspan=1 colspan=1>31.59</td><td rowspan=1 colspan=1>33.96</td><td rowspan=1 colspan=1>31.63</td><td rowspan=1 colspan=1>33.77</td></tr><tr><td rowspan=1 colspan=1>BSD400</td><td rowspan=1 colspan=1>RNAN</td><td rowspan=1 colspan=2>42.86</td><td rowspan=1 colspan=2>42.61</td><td rowspan=1 colspan=3>SR291</td><td rowspan=1 colspan=1>RNAN</td><td rowspan=1 colspan=1>31.87</td><td rowspan=1 colspan=1>34.32</td><td rowspan=1 colspan=1>31.99</td><td rowspan=1 colspan=1>34.03</td></tr><tr><td rowspan=1 colspan=1></td><td rowspan=1 colspan=10>Task 5:Image super-resolution (SR).Low-resolution (LR) images are obtained by Bicubic downsampling.</td><td rowspan=1 colspan=3>mpling.</td></tr><tr><td rowspan=2 colspan=1>Data</td><td rowspan=2 colspan=1>Method</td><td rowspan=1 colspan=3>Set5</td><td rowspan=1 colspan=3>Set14</td><td rowspan=1 colspan=3>B100</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=2>Urban100</td></tr><tr><td rowspan=1 colspan=1>×2</td><td rowspan=1 colspan=1>×3</td><td rowspan=1 colspan=1>×4</td><td rowspan=1 colspan=1>×2</td><td rowspan=1 colspan=1>×3</td><td rowspan=1 colspan=1>×4</td><td rowspan=1 colspan=1>×2</td><td rowspan=1 colspan=1>×3</td><td rowspan=1 colspan=1>×4</td><td rowspan=1 colspan=1>×2</td><td rowspan=1 colspan=1>×3</td><td rowspan=1 colspan=1>×4</td></tr><tr><td rowspan=1 colspan=1>SR291</td><td rowspan=1 colspan=1>MemNet</td><td rowspan=1 colspan=1>37.78</td><td rowspan=1 colspan=1>34.09</td><td rowspan=1 colspan=1>31.74</td><td rowspan=1 colspan=1>33.28</td><td rowspan=1 colspan=1>30.00</td><td rowspan=1 colspan=1>28.26</td><td rowspan=1 colspan=1>32.08</td><td rowspan=1 colspan=1>28.96</td><td rowspan=1 colspan=1>27.40</td><td rowspan=1 colspan=1>31.31</td><td rowspan=1 colspan=1>27.56</td><td rowspan=1 colspan=1>25.50</td></tr><tr><td rowspan=1 colspan=1>SR291</td><td rowspan=1 colspan=1>RNAN</td><td rowspan=1 colspan=1>37.95</td><td rowspan=1 colspan=1>34.30</td><td rowspan=1 colspan=1>31.93</td><td rowspan=1 colspan=1>33.41</td><td rowspan=1 colspan=1>30.22</td><td rowspan=1 colspan=1>28.43</td><td rowspan=1 colspan=1>32.23</td><td rowspan=1 colspan=1>29.09</td><td rowspan=1 colspan=1>27.51</td><td rowspan=1 colspan=1>31.64</td><td rowspan=1 colspan=1>27.84</td><td rowspan=1 colspan=1>25.71</td></tr></table>
333
+
334
+ # A.2 VISUAL RESULTS
335
+
336
+ Color and Gray Image Denoising. We show color and gray-scale image denoising comparisons in Figures 9 and 10 respectively. We can see that our RNAN recovers shaper edges. Unlike most of other methods, which over-smooth some details (e.g., tiny lines), RNAN can reduce noise and maintain more details. With the learned non-local mixed attention, RNAN treats different image parts distinctively, alleviating over-smoothing artifacts obviously.
337
+
338
+ Image Compression Artifacts Reduction (CAR). In Figure 11, we provide comparisons under very low image quality $\scriptstyle ( q = 1 0 )$ . The blocking artifacts can be removed to some degree, but ARCNN, TNRD, and DnCNN would also over-smooth some structures. In contrast, RNAN obatins more details with consistent structures by considering non-local mixed attention.
339
+
340
+ Image Super-Resolution (SR). In Figure 12, we conduct image SR $( \times 4 )$ with several state-of-theart methods, such as SRCNN (Dong et al., 2016), VDSR (Kim et al., 2016), LapSRN (Lai et al., 2017), MemNet (Tai et al., 2017), EDSR (Lim et al., 2017), SRMDNF (Zhang et al., 2018a), and DDBPN (Haris et al., 2018). We can see that most of compared methods would suffer from distortion or output totally wrong structures. For tiny line, edge structures, or some textures, they fail to recover and introduce blurring artifacts. Instead, our RNAN obtains better visually pleasing results with finer structures. These comparisons further demonstrate the effectiveness of our proposed RNAN with the usage of non-local mixed attention.
341
+
342
+ ![](images/36a51adf4a6c51bfefd1250608b0083189f507220fe46d1ed208bf9f0c732ba8.jpg)
343
+ Figure 9: Color image denoising results with noise level $\sigma = 5 0$ .
344
+
345
+ ![](images/34f2c0b0e7fbd8011e94fae14ee90672913c86747a06b7d5e2bffb2d6b167f31.jpg)
346
+ Figure 10: Gray-scale image denoising results with noise level $\sigma = 5 0$ .
347
+
348
+ ![](images/2bd28317c017c3513f74d59889f615fce7f6397ba91f47e04844ae75d3d2bdaa.jpg)
349
+ Figure 11: Image compression artifacts reduction results with JPEG quality $q = 1 0$
350
+
351
+ ![](images/30d7bb8dd30b6119879dfe418e24c3a99e36688db7c015ac2fb757abd1a8f90b.jpg)
352
+ Figure 12: Image super-resolution results with scaling factor $s = 4$
parse/train/HkeGhoA5FX/HkeGhoA5FX_content_list.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/HkeGhoA5FX/HkeGhoA5FX_middle.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/HkeGhoA5FX/HkeGhoA5FX_model.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/train/QYjO70ACDK/QYjO70ACDK_content_list.json ADDED
@@ -0,0 +1,1515 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "type": "text",
4
+ "text": "DISTRIBUTIONAL SLICED-WASSERSTEIN AND APPLICATIONS TO GENERATIVE MODELING ",
5
+ "text_level": 1,
6
+ "bbox": [
7
+ 174,
8
+ 99,
9
+ 725,
10
+ 146
11
+ ],
12
+ "page_idx": 0
13
+ },
14
+ {
15
+ "type": "text",
16
+ "text": "Khai Nguyen VinAI Research, Vietnam v.khainb@vinai.io ",
17
+ "bbox": [
18
+ 184,
19
+ 170,
20
+ 354,
21
+ 212
22
+ ],
23
+ "page_idx": 0
24
+ },
25
+ {
26
+ "type": "text",
27
+ "text": "Nhat $\\mathbf { H o } ^ { * }$ University of Texas, Austin VinAI Research, Vietnam minhnhat@utexas.edu ",
28
+ "bbox": [
29
+ 400,
30
+ 170,
31
+ 589,
32
+ 226
33
+ ],
34
+ "page_idx": 0
35
+ },
36
+ {
37
+ "type": "text",
38
+ "text": "Tung Pham VinAI Research, Vietnam v.tungph4@vinai.io ",
39
+ "bbox": [
40
+ 635,
41
+ 170,
42
+ 813,
43
+ 212
44
+ ],
45
+ "page_idx": 0
46
+ },
47
+ {
48
+ "type": "text",
49
+ "text": "Hung Bui ",
50
+ "text_level": 1,
51
+ "bbox": [
52
+ 184,
53
+ 247,
54
+ 253,
55
+ 261
56
+ ],
57
+ "page_idx": 0
58
+ },
59
+ {
60
+ "type": "text",
61
+ "text": "VinAI Research, Vietnam v.hungbh1@vinai.io ",
62
+ "bbox": [
63
+ 184,
64
+ 262,
65
+ 361,
66
+ 289
67
+ ],
68
+ "page_idx": 0
69
+ },
70
+ {
71
+ "type": "text",
72
+ "text": "ABSTRACT ",
73
+ "text_level": 1,
74
+ "bbox": [
75
+ 454,
76
+ 325,
77
+ 544,
78
+ 340
79
+ ],
80
+ "page_idx": 0
81
+ },
82
+ {
83
+ "type": "text",
84
+ "text": "Sliced-Wasserstein distance (SW) and its variant, Max Sliced-Wasserstein distance (Max-SW), have been used widely in the recent years due to their fast computation and scalability even when the probability measures lie in a very high dimensional space. However, SW requires many unnecessary projection samples to approximate its value while Max-SW only uses the most important projection, which ignores the information of other useful directions. In order to account for these weaknesses, we propose a novel distance, named Distributional Sliced-Wasserstein distance (DSW), that finds an optimal distribution over projections that can balance between exploring distinctive projecting directions and the informativeness of projections themselves. We show that the DSW is a generalization of Max-SW, and it can be computed efficiently by searching for the optimal push-forward measure over a set of probability measures over the unit sphere satisfying certain regularizing constraints that favor distinct directions. Finally, we conduct extensive experiments with large-scale datasets to demonstrate the favorable performances of the proposed distances over the previous sliced-based distances in generative modeling applications. ",
85
+ "bbox": [
86
+ 232,
87
+ 359,
88
+ 766,
89
+ 582
90
+ ],
91
+ "page_idx": 0
92
+ },
93
+ {
94
+ "type": "text",
95
+ "text": "1 INTRODUCTION ",
96
+ "text_level": 1,
97
+ "bbox": [
98
+ 176,
99
+ 616,
100
+ 336,
101
+ 632
102
+ ],
103
+ "page_idx": 0
104
+ },
105
+ {
106
+ "type": "text",
107
+ "text": "Optimal transport (OT) is a classical problem in mathematics and operation research. Due to its appealing theoretical properties and flexibility in practical applications, it has recently become an important tool in the machine learning and statistics community; see for example, (Courty et al., 2017; Arjovsky et al., 2017; Tolstikhin et al., 2018; Gulrajani et al., 2017) and references therein. The main usage of OT is to provide a distance named Wasserstein distance, to measure the discrepancy between two probability distributions. However, that distance suffers from expensive computational complexity, which is the main obstacle to using OT in practical applications. ",
108
+ "bbox": [
109
+ 174,
110
+ 650,
111
+ 825,
112
+ 747
113
+ ],
114
+ "page_idx": 0
115
+ },
116
+ {
117
+ "type": "text",
118
+ "text": "There have been two main approaches to overcome the high computational complexity problem: either approximate the value of OT or apply the OT adaptively to specific situations. The first approach was initiated by (Cuturi, 2013) using an entropic regularizer to speed up the computation of the OT (Sinkhorn, 1967; Knight, 2008). The entropic regularization approach has demonstrated its usefulness in several application domains (Courty et al., 2014; Genevay et al., 2018; Bunne et al., 2019). Along this direction, several works proposed efficient algorithms for solving the entropic OT (Altschuler et al., 2017; Lin et al., 2019b;a) as well as methods to stabilize these algorithms (Chizat et al., 2018; Peyré & Cuturi, 2019; Chizat et al., 2018; Schmitzer, 2019). However, these algorithms have complexities of the order $\\mathcal { O } ( k ^ { 2 } )$ , where $k$ is the number of supports. It is expensive when we need to compute the OT repeatedly, especially in learning the data distribution. ",
119
+ "bbox": [
120
+ 174,
121
+ 755,
122
+ 825,
123
+ 893
124
+ ],
125
+ "page_idx": 0
126
+ },
127
+ {
128
+ "type": "text",
129
+ "text": "The second approach, known as \"slicing\", takes a rather different perspective. It leverages two key ideas: the OT closed-form expression for two distributions in one-dimensional space, and the transformation of a distribution into a set of projected one-dimensional distributions by the Radon transform (RT) (Helgason, 2010). The popular proposal along this direction is Sliced-Wasserstein (SW) distance (Bonneel et al., 2015), which samples the projecting directions uniformly over a unit sphere in the data ambient space and takes the expectation of the resulting one-dimensional OT distance. The SW distance hence requires a significantly lower computation cost than the original Wasserstein distance and is more scalable than the first approach. Due to its solid statistical guarantees and efficient computation, the SW distance has been successfully applied to a variety of practical tasks (Deshpande et al., 2018; Liutkus et al., 2019; Kolouri et al., 2018; Wu et al., 2019; Deshpande et al., 2019) where it has been shown to have comparative performances to other distances and divergences between probability distributions. However, there is an inevitable bottleneck of computing the SW distance. Specifically, the expectation with respect to the uniform distribution of projections in SW is intractable to compute; therefore, the Monte Carlo method is employed to approximate it. Nevertheless, drawing from a uniform distribution of directions in high-dimension can result in an overwhelming number of irrelevant directions, especially when the actual data lies in a low-dimensional manifold. Hence, SW typically needs to have a large number of samples to yield an accurate estimation of the discrepancy. Alternatively, in the other extreme, Max Sliced-Wasserstein (Max-SW) distance (Deshpande et al., 2019) uses only one important direction to distinguish the probability distributions. However, other potentially relevant directions are ignored in Max-SW. Therefore, Max-SW can miss some important differences between the two distributions in high dimension. We note that the linear projections in the Radon transform can be replaced by non-linear projections resulting in the generalized sliced-Wasserstein distance and its variants (Beylkin, 1984; Kolouri et al., 2019). ",
130
+ "bbox": [
131
+ 174,
132
+ 103,
133
+ 825,
134
+ 435
135
+ ],
136
+ "page_idx": 1
137
+ },
138
+ {
139
+ "type": "text",
140
+ "text": "Apart from these main directions, there are also few proposals that try either to modify them or to combine the advantages of the above-mentioned approaches. In particular, Paty & Cuturi (2019) extended the idea of the max-sliced distance to the max-subspace distance by considering finding an optimal orthogonal subspace. However, this approach is computationally expensive, since it could not exploit the closed-form of the one-dimensional Wasserstein distance. Another approach named the Projected Wasserstein distance (PWD), which was proposed in (Rowland et al., 2019), uses sliced decomposition to find multiple one-dimension optimal transport maps. Then, it computes the average cost of those maps equally in the original dimension. ",
141
+ "bbox": [
142
+ 174,
143
+ 443,
144
+ 825,
145
+ 554
146
+ ],
147
+ "page_idx": 1
148
+ },
149
+ {
150
+ "type": "text",
151
+ "text": "Our contributions. Our paper also follows the slicing approach. However, we address key friction in this general line of work: how to obtain a relatively small number of slices simultaneously to maintain the computational efficiency, but at the same time, cover the major differences between two high-dimensional distributions. We take a probabilistic view of slicing by using a probability measure on the unit sphere to represent how important each direction is. From this viewpoint, SW uses the uniform distribution while Max-SW searches for the best delta-Dirac distribution over the projections, both can be considered as special cases. In this paper, we propose to search for an optimal distribution of important directions. We regularize this distribution such that it prefers directions that are far away from one another, hence encouraging an efficient exploration of the space of directions. In the case of no regularization, our proposed method recovers max-(generalized) SW as a special case. In summary, our main contributions are two-fold: ",
152
+ "bbox": [
153
+ 174,
154
+ 561,
155
+ 825,
156
+ 714
157
+ ],
158
+ "page_idx": 1
159
+ },
160
+ {
161
+ "type": "text",
162
+ "text": "1. First, we introduce a novel distance, named Distributional Sliced-Wasserstein distance (DSW), to account for the issues of previous sliced distances. Our main idea is to search for not just a single most important projection, but an optimal distribution over projections that could balance between an expansion of the area around important projections and the informativeness of projections themselves, i.e., how well they can distinguish the two target probability measures. We show that DSW is a proper metric in the probability space and possesses appealing statistical and computational properties as the previous sliced distances. ",
163
+ "bbox": [
164
+ 214,
165
+ 731,
166
+ 825,
167
+ 829
168
+ ],
169
+ "page_idx": 1
170
+ },
171
+ {
172
+ "type": "text",
173
+ "text": "2. Second, we apply the DSW distance to generative modeling tasks based on the generative adversarial framework. The extensive experiments on real and large-scale datasets show that DSW distance significantly outperforms the SW and Max-SW distances under similar computational time on these tasks. Furthermore, the DSW distance helps model distribution converge to the data distribution faster and provides more realistic generated images than the SW and Max-SW distances. ",
174
+ "bbox": [
175
+ 214,
176
+ 840,
177
+ 823,
178
+ 922
179
+ ],
180
+ "page_idx": 1
181
+ },
182
+ {
183
+ "type": "text",
184
+ "text": "Organization. The remainder of the paper is organized as follows. In Section 2, we provide backgrounds for Wasserstein distance and its slice-based versions. In Section 3, we propose distributional (generalized) sliced-Wasserstein distance and analyze some of its theoretical properties. Section 4 includes extensive experiment results followed by discussions in Section 5. Finally, we defer the proofs of key results and extra materials in the Appendices. ",
185
+ "bbox": [
186
+ 173,
187
+ 103,
188
+ 825,
189
+ 174
190
+ ],
191
+ "page_idx": 2
192
+ },
193
+ {
194
+ "type": "text",
195
+ "text": "Notation. For any $\\theta , \\theta ^ { \\prime } \\in \\mathbb { R } ^ { d }$ , $\\begin{array} { r } { \\cos ( \\theta , \\theta ^ { \\prime } ) = \\frac { \\theta ^ { \\dagger } \\overset { \\star } { \\theta ^ { \\prime } } } { \\| \\theta \\| \\| \\theta ^ { \\prime } \\| } } \\end{array}$ θ>θ0kθkkθ0k , where k.k is \\`2 norm. For any d ≥ 2, Sd−1 denotes the unit sphere in $d$ dimension in $\\ell _ { 2 }$ norm . Furthermore, $\\delta$ denotes the Dirac delta function, and $\\langle \\cdot , \\cdot \\rangle$ is the Euclidean inner-product. For any $p \\geq 1$ , $\\mathbb { L } ^ { p } ( \\mathbb { R } ^ { d } )$ is the set of real-valued functions on $\\mathbb { R } ^ { d }$ with finite $p$ -th moment. ",
196
+ "bbox": [
197
+ 174,
198
+ 174,
199
+ 825,
200
+ 234
201
+ ],
202
+ "page_idx": 2
203
+ },
204
+ {
205
+ "type": "text",
206
+ "text": "2 BACKGROUND ",
207
+ "text_level": 1,
208
+ "bbox": [
209
+ 174,
210
+ 255,
211
+ 326,
212
+ 272
213
+ ],
214
+ "page_idx": 2
215
+ },
216
+ {
217
+ "type": "text",
218
+ "text": "In this section, we provide necessary backgrounds for the (generalized) Radon transform, the Wasserstein, and sliced-Wasserstein distances. ",
219
+ "bbox": [
220
+ 174,
221
+ 287,
222
+ 823,
223
+ 315
224
+ ],
225
+ "page_idx": 2
226
+ },
227
+ {
228
+ "type": "text",
229
+ "text": "2.1 WASSERSTEIN DISTANCE ",
230
+ "text_level": 1,
231
+ "bbox": [
232
+ 174,
233
+ 333,
234
+ 388,
235
+ 348
236
+ ],
237
+ "page_idx": 2
238
+ },
239
+ {
240
+ "type": "text",
241
+ "text": "We start with a formal definition of Wasserstein distance. For any $p \\geq 1$ , we define $\\mathcal { P } _ { p } ( \\mathbb { R } ^ { d } )$ as the set of Borel probability measures with finite $p$ -th moment defined on a given metric space $( \\mathbb { R } ^ { d } , \\lVert . \\rVert )$ For any probability measures $\\mu , \\nu$ defined on $\\boldsymbol { \\mathcal { X } } , \\boldsymbol { \\mathcal { Y } } \\subseteq \\mathbb { R } ^ { d }$ , we denote their corresponding probability density functions as $I _ { \\mu }$ and $I _ { \\nu }$ . The Wasserstein distance of order $p$ between $\\mu$ and $\\nu$ is given by (Villani, 2008; Peyré & Cuturi, 2019): ",
242
+ "bbox": [
243
+ 173,
244
+ 359,
245
+ 825,
246
+ 431
247
+ ],
248
+ "page_idx": 2
249
+ },
250
+ {
251
+ "type": "equation",
252
+ "img_path": "images/a15babc2c5f1ea6f242d0af798d69ae8726b94fc0939ae988a198fc66dd7f745.jpg",
253
+ "text": "$$\nW _ { p } ( \\mu , \\nu ) : = \\Big ( \\operatorname* { i n f } _ { \\pi \\in \\Pi ( \\mu , \\nu ) } \\int _ { \\mathcal { X } \\times \\mathcal { Y } } \\| x - y \\| ^ { p } d \\pi ( x , y ) \\Big ) ^ { \\frac { 1 } { p } } ,\n$$",
254
+ "text_format": "latex",
255
+ "bbox": [
256
+ 325,
257
+ 439,
258
+ 669,
259
+ 473
260
+ ],
261
+ "page_idx": 2
262
+ },
263
+ {
264
+ "type": "text",
265
+ "text": "where $\\Pi ( \\mu , \\nu )$ is a set of all transportation plans $\\pi$ such that the marginal distributions of $\\pi$ are $\\mu$ and $\\nu$ , respectively. In order to simplify the presentation, we abuse the notation by using both $W _ { p } ( \\mu , \\nu )$ and $W _ { p } ( I _ { \\mu } , I _ { \\nu } )$ interchangeably for the Wasserstein distance between $\\mu$ and $\\nu$ . When $\\mu$ and $\\nu$ are one-dimension measures, the Wasserstein distance between $\\mu$ and $\\nu$ has a closed-form expression $\\begin{array} { l } { { W _ { p } ( \\mu , \\nu ) ~ = ~ ( \\int _ { 0 } ^ { 1 } | F _ { \\mu } ^ { - 1 } ( z ) - F _ { \\nu } ^ { - 1 } ( z ) | ^ { p } d z ) ^ { 1 / p } } } \\end{array}$ where $F _ { \\mu }$ and $F _ { \\nu }$ are the cumulative distribution function (CDF) of $I _ { \\mu }$ and $I _ { \\nu }$ , respectively. ",
266
+ "bbox": [
267
+ 173,
268
+ 479,
269
+ 825,
270
+ 568
271
+ ],
272
+ "page_idx": 2
273
+ },
274
+ {
275
+ "type": "text",
276
+ "text": "2.2 (GENERALIZED) RADON TRANSFORMS ",
277
+ "text_level": 1,
278
+ "bbox": [
279
+ 176,
280
+ 584,
281
+ 485,
282
+ 598
283
+ ],
284
+ "page_idx": 2
285
+ },
286
+ {
287
+ "type": "text",
288
+ "text": "Now, we review (generalized) Radon transform maps, which are key to the notion of (generalized) sliced-Wasserstein distance and its variants. The Radon transform (RT) maps a function $\\mathbf { \\bar { \\chi } } _ { I } \\in \\mathbb { L } ^ { 1 } ( \\mathbb { R } ^ { d } )$ to the space of functions defined over space of lines in $\\mathbb { R } ^ { d }$ . In particular, for any $t \\in \\mathbb { R }$ and direction $\\theta \\in \\mathbb { S } ^ { d - 1 }$ , the RT is defined as follows (Helgason, $\\begin{array} { r } { 2 0 1 0 ) : \\mathcal { R } \\bar { I } ( t , \\theta ) : = \\int _ { \\mathbb { R } ^ { d } } I ( \\bar { x } ) \\delta ( t - \\langle x , \\theta \\rangle ) d x } \\end{array}$ . ",
289
+ "bbox": [
290
+ 174,
291
+ 609,
292
+ 826,
293
+ 667
294
+ ],
295
+ "page_idx": 2
296
+ },
297
+ {
298
+ "type": "text",
299
+ "text": "The generalized Radon transform (GRT) (Beylkin, 1984) extends the original one from integration over hyperplanes of $\\mathbb { R } ^ { d }$ to integration over hypersurfaces. In particular, it is defined as: $\\mathcal { G } I ( t , { \\boldsymbol { \\theta } } ) : =$ $\\textstyle \\int _ { \\mathbb { R } ^ { d } } I ( { \\bar { x } } ) \\delta ( { \\bar { t } } - g ( x , \\theta ) ) d x$ , where $t \\in \\mathbb R$ and $\\theta \\in \\Omega _ { \\theta }$ . Here, $\\Omega _ { \\theta }$ is a compact subset of $\\mathbb { R } ^ { d }$ and $\\boldsymbol { g } : \\mathbb { R } ^ { d } \\times \\mathbb { S } ^ { d - 1 } \\mapsto \\mathbb { R }$ is a defining function (cf. Assumptions H1-H4 in (Kolouri et al., 2019) for the definition of defining function) inducing the hypersurfaces. When $g ( x , \\theta ) = \\langle x , \\theta \\rangle$ and $\\Omega _ { \\theta } = \\mathbb { S } ^ { d - 1 }$ , the generalized Radon transform becomes the standard Radon transform. ",
300
+ "bbox": [
301
+ 173,
302
+ 672,
303
+ 825,
304
+ 760
305
+ ],
306
+ "page_idx": 2
307
+ },
308
+ {
309
+ "type": "text",
310
+ "text": "2.3 (GENERALIZED) SLICED-WASSERSTEIN DISTANCES ",
311
+ "text_level": 1,
312
+ "bbox": [
313
+ 173,
314
+ 775,
315
+ 573,
316
+ 790
317
+ ],
318
+ "page_idx": 2
319
+ },
320
+ {
321
+ "type": "text",
322
+ "text": "The sliced-Wasserstein distance (SW) between two probability measures $\\mu$ and $\\nu$ is defined as (Bonneel et al., 2015): $\\begin{array} { r } { S W _ { p } ( \\mu , \\nu ) : = ( \\int _ { \\mathbb { S } ^ { d - 1 } } W _ { p } ^ { p } \\big ( \\mathcal { R } I _ { \\mu } ( \\cdot , \\theta ) , \\mathcal { R } I _ { \\nu } ( \\cdot , \\theta ) \\big ) d \\theta ) ^ { 1 / p } } \\end{array}$ . Similarly, the generalized sliced-Wasserstein distance (Kolouri et al., 2019) (GSW) is given by $\\mathrm { G S W } _ { p } ( \\mu , \\nu ) : =$ $\\begin{array} { r } { ( \\int _ { \\Omega _ { \\theta } } W _ { p } ^ { p } \\bigl ( \\mathcal { G } I _ { \\mu } ( \\cdot , \\theta ) , \\mathcal { G } I _ { \\nu } ( \\cdot , \\theta ) \\bigr ) d \\theta ) ^ { 1 / p } } \\end{array}$ , where $\\Omega _ { \\theta }$ is the compact set of feasible parameter. However, these integrals are usually intractable. Thus, they are often approximated by using Monte Carlo scheme to draw uniform samples $\\{ \\theta _ { i } \\} _ { i = 1 } ^ { N }$ from $\\bar { \\mathbb { S } } ^ { d - 1 }$ and $\\Omega _ { \\theta }$ . In particular, $S W _ { p } ^ { p } ( \\mu , \\nu ) \\approx$ $\\begin{array} { r } { \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } W _ { p _ { \\cdot } } ^ { p } \\big ( \\mathscr { R } I _ { \\mu } ( \\cdot , \\theta _ { i } ) , \\mathscr { R } I _ { \\nu } ( \\cdot , \\theta _ { i } ) \\big ) } \\end{array}$ and $\\begin{array} { r } { \\mathbf { G S W } _ { p } ^ { p } ( \\mu , \\nu ) \\approx \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } W _ { p } ^ { p } \\big ( \\mathcal { G } I _ { \\mu } ( \\cdot , \\theta _ { i } ) , \\mathcal { G } I _ { \\nu } ( \\cdot , \\theta _ { i } ) \\big ) } \\end{array}$ . In order to obtain a good approximation of (generalized) SW distances, $N$ needs to be sufficiently large. ",
323
+ "bbox": [
324
+ 173,
325
+ 800,
326
+ 826,
327
+ 924
328
+ ],
329
+ "page_idx": 2
330
+ },
331
+ {
332
+ "type": "text",
333
+ "text": "However, important directions are not distributed uniformly over the sphere. Thus, this approach will draw potentially many unimportant projections that are not only expensive but also greatly reduce the effect of the SW distance. ",
334
+ "bbox": [
335
+ 174,
336
+ 103,
337
+ 823,
338
+ 146
339
+ ],
340
+ "page_idx": 3
341
+ },
342
+ {
343
+ "type": "text",
344
+ "text": "2.4 MAX (GENERALIZED) SLICED-WASSERSTEIN DISTANCES ",
345
+ "text_level": 1,
346
+ "bbox": [
347
+ 174,
348
+ 162,
349
+ 609,
350
+ 178
351
+ ],
352
+ "page_idx": 3
353
+ },
354
+ {
355
+ "type": "text",
356
+ "text": "An approach to using only informative directions is to simply take the best slice in discriminating two given probability distributions. That distance is max sliced-Wasserstein distance (Max-SW) (Deshpande et al., 2019), which is given by $\\begin{array} { r l } & { \\operatorname* { m a x } S W _ { p } ( \\mu , \\nu ) : = \\operatorname* { m a x } _ { \\theta \\in \\mathbb { S } ^ { d - 1 } } W _ { p } ( \\mathcal { R } I _ { \\mu } ( \\cdot , \\theta ) , \\mathcal { R } I _ { \\nu } ( \\cdot , \\theta ) ) } \\end{array}$ . By combining this idea with non-linear projections from generalized Radon transform, we obtain max generalized sliced-Wasserstein distance (Max-GSW) (Kolouri et al., 2019). The formal definition of that distance is: $\\begin{array} { r } { \\operatorname* { m a x } G S W _ { p } ( \\mu , \\nu ) : = \\operatorname* { m a x } _ { \\theta \\in \\Omega _ { \\theta } } W _ { p } ( \\mathcal { G } I _ { \\mu } ( \\cdot , \\theta ) , \\mathcal { G } I _ { \\nu } ( \\cdot , \\theta ) ) } \\end{array}$ . The (generalized) MaxSW distances focus on finding only the most important direction. Meanwhile, other informative directions play no role in the distance. Therefore, (generalized) Max-SW distances can ignore useful information about the structure of high dimensional probability measures. ",
357
+ "bbox": [
358
+ 174,
359
+ 188,
360
+ 825,
361
+ 314
362
+ ],
363
+ "page_idx": 3
364
+ },
365
+ {
366
+ "type": "text",
367
+ "text": "3 DISTRIBUTIONAL SLICED-WASSERSTEIN DISTANCE ",
368
+ "text_level": 1,
369
+ "bbox": [
370
+ 176,
371
+ 333,
372
+ 638,
373
+ 349
374
+ ],
375
+ "page_idx": 3
376
+ },
377
+ {
378
+ "type": "text",
379
+ "text": "With the aim of improving the limitations of the previous sliced distances, we propose a novel distance, named Distributional Sliced-Wasserstein distance (DSW), that can search for not just a single but a distribution of important directions on the unit sphere. We prove that it is a well-defined metric and discuss its connection to the existing sliced-based distances in Section 3.1. Then, we provide a procedure to approximate DSW based on its dual form in Section 3.2. ",
380
+ "bbox": [
381
+ 174,
382
+ 364,
383
+ 825,
384
+ 435
385
+ ],
386
+ "page_idx": 3
387
+ },
388
+ {
389
+ "type": "text",
390
+ "text": "3.1 DEFINITION AND METRICITY ",
391
+ "text_level": 1,
392
+ "bbox": [
393
+ 176,
394
+ 452,
395
+ 415,
396
+ 465
397
+ ],
398
+ "page_idx": 3
399
+ },
400
+ {
401
+ "type": "text",
402
+ "text": "We first start with a definition of distributional sliced-Wasserstein distance. We say $C > 0$ admissible if the set $\\mathbb { M } _ { C }$ of probability measures $\\sigma$ on $\\mathbb { S } ^ { d - 1 }$ satisfying $\\begin{array} { r } { \\mathbb { E } _ { \\boldsymbol { \\theta } , \\boldsymbol { \\theta } ^ { \\prime } \\sim \\sigma } \\left[ \\vert \\boldsymbol { \\theta } ^ { \\top } \\boldsymbol { \\theta } ^ { \\prime } \\vert \\right] \\le \\dot { C } } \\end{array}$ is not empty. ",
403
+ "bbox": [
404
+ 171,
405
+ 477,
406
+ 823,
407
+ 507
408
+ ],
409
+ "page_idx": 3
410
+ },
411
+ {
412
+ "type": "text",
413
+ "text": "Definition 1. Given two probability measures $\\mu$ and $\\nu$ on $\\mathbb { R } ^ { d }$ with finite $p$ -th moments where $p \\geq 1$ and an admissible regularizing constant $C > 0$ . The distributional sliced-Wasserstein distance (DSW) of order $p$ between $\\mu$ and $\\nu$ is given by: ",
414
+ "bbox": [
415
+ 174,
416
+ 512,
417
+ 825,
418
+ 555
419
+ ],
420
+ "page_idx": 3
421
+ },
422
+ {
423
+ "type": "equation",
424
+ "img_path": "images/a975bee6b11279c449387bfdd5cb5dcf8347b9f30c8685d990f18ffc14917479.jpg",
425
+ "text": "$$\n{ D S W } _ { p } ( \\mu , \\nu ; C ) : = \\operatorname* { s u p } _ { \\sigma \\in \\mathbb { M } _ { C } } \\bigg ( \\mathbb { E } _ { \\theta \\sim \\sigma } \\bigg [ W _ { p } ^ { p } ( \\mathcal { R } I _ { \\mu } ( \\cdot , \\theta ) , \\mathcal { R } I _ { \\nu } ( \\cdot , \\theta ) ) \\bigg ] \\bigg ) ^ { \\frac { 1 } { p } } ,\n$$",
426
+ "text_format": "latex",
427
+ "bbox": [
428
+ 282,
429
+ 560,
430
+ 712,
431
+ 599
432
+ ],
433
+ "page_idx": 3
434
+ },
435
+ {
436
+ "type": "text",
437
+ "text": "where $\\mathcal { R }$ is the Radon transform operator. ",
438
+ "bbox": [
439
+ 176,
440
+ 604,
441
+ 447,
442
+ 619
443
+ ],
444
+ "page_idx": 3
445
+ },
446
+ {
447
+ "type": "text",
448
+ "text": "The DSW aims to find the optimal probability measure of slices on the unit sphere $\\mathbb { S } ^ { d - 1 }$ . Note that, the Max-SW distance is equivalent to searching for the best Dirac measure on a single point in $\\mathbb { S } ^ { d - 1 }$ , which puts all weights in only one direction. Meanwhile, the uniform measure in the formulation of SW distance distributes the same weights in all directions. Indeed, the uniform and Dirac measures are two special cases, because they view that either all directions are equally important or only one direction is important. That view is too restricted if the data actually lie on low dimensional space. Thus, we aim to find a probability measure which concentrates only on areas around important directions. Furthermore, we do not want these directions to lie in only one small area, because under the orthogonal projection of RT, their corresponding one-dimensional distributions will become similar. In order to achieve this, we search for an optimal measure $\\sigma$ that satisfies the regularization constraint $\\begin{array} { r } { \\mathbb { E } _ { \\theta , \\theta ^ { \\prime } \\sim \\sigma } [ \\left| \\theta ^ { \\top } \\theta ^ { ' } \\right| ] \\leq C } \\end{array}$ . By Cauchy-Schwarz inequality, $C$ is no greater than 1, thus $\\mathbb { M } _ { 1 }$ contains all probability measures on the unit sphere. Optimizing over $\\mathbb { M } _ { 1 }$ simply returns the best Dirac measure corresponding to the Max-SW distance. When $C$ is small, the constraint forces the measure $\\sigma$ to distribute more weights to directions that are far from each other (in terms of their angles). Thus, a small appropriate value of $C$ will help to balance between the distinctiveness and informativeness of these targeted directions. For further discussion about $C$ , see Appendix B.1. ",
449
+ "bbox": [
450
+ 173,
451
+ 631,
452
+ 826,
453
+ 856
454
+ ],
455
+ "page_idx": 3
456
+ },
457
+ {
458
+ "type": "text",
459
+ "text": "Next, we show that DSW is a well-defined metric on the probability space. ",
460
+ "bbox": [
461
+ 174,
462
+ 863,
463
+ 661,
464
+ 877
465
+ ],
466
+ "page_idx": 3
467
+ },
468
+ {
469
+ "type": "text",
470
+ "text": "Theorem 1. For any $p \\geq 1$ and admissible $C > 0$ , $D S W _ { p } ( \\cdot , \\cdot ; C )$ is a well-defined metric in the space of Borel probability measures with finite $p$ -th moment. In particular, it is non-negative, symmetric, identity, and satisfies the triangle inequality. ",
471
+ "bbox": [
472
+ 176,
473
+ 881,
474
+ 825,
475
+ 924
476
+ ],
477
+ "page_idx": 3
478
+ },
479
+ {
480
+ "type": "text",
481
+ "text": "The proof of Theorem 1 is in Appendix A.1. Our next result establishes the topological equivalence between DSW distance and (max)-sliced Wasserstein and Wasserstein distances. ",
482
+ "bbox": [
483
+ 171,
484
+ 103,
485
+ 823,
486
+ 132
487
+ ],
488
+ "page_idx": 4
489
+ },
490
+ {
491
+ "type": "text",
492
+ "text": "Theorem 2. For any $p \\geq 1$ and admissible $C > 0 ;$ , the following holds ",
493
+ "bbox": [
494
+ 173,
495
+ 136,
496
+ 640,
497
+ 151
498
+ ],
499
+ "page_idx": 4
500
+ },
501
+ {
502
+ "type": "equation",
503
+ "img_path": "images/f86ba042621b0a5de0acb5f2137ef9c46979ef6cc97d30481a2cd0278bb55d99.jpg",
504
+ "text": "$$\nD S W _ { p } ( \\mu , \\nu ; C ) \\leq m a x S W _ { p } ( \\mu , \\nu ) \\leq W _ { p } ( \\mu , \\nu ) .\n$$",
505
+ "text_format": "latex",
506
+ "bbox": [
507
+ 230,
508
+ 161,
509
+ 545,
510
+ 179
511
+ ],
512
+ "page_idx": 4
513
+ },
514
+ {
515
+ "type": "text",
516
+ "text": "(b) If $C \\geq 1 / d ,$ , we have $\\begin{array} { r } { D S W _ { p } ( \\mu , \\nu ; C ) \\geq \\left( \\frac { 1 } { d } \\right) ^ { 1 / p } m a x { S W _ { p } ( \\mu , \\nu ) } \\geq \\left( \\frac { 1 } { d } \\right) ^ { 1 / p } S W _ { p } ( \\mu , \\nu ) . } \\end{array}$ ",
517
+ "bbox": [
518
+ 205,
519
+ 185,
520
+ 789,
521
+ 209
522
+ ],
523
+ "page_idx": 4
524
+ },
525
+ {
526
+ "type": "text",
527
+ "text": "As a consequence, when $p \\geq 1$ and $C \\geq 1 / d , D S W _ { p } ( \\cdot , \\cdot ; C ) , \\cdot$ $S W _ { p }$ , max $S W _ { p }$ , and $W _ { p }$ are topologically equivalent, namely, the convergence of probability measures under $D \\bar { S } W _ { p } ( \\cdot , \\cdot ; \\mathbf { \\bar { C } } )$ implies the convergence of these measures under other metrics and vice versa. ",
528
+ "bbox": [
529
+ 176,
530
+ 218,
531
+ 823,
532
+ 262
533
+ ],
534
+ "page_idx": 4
535
+ },
536
+ {
537
+ "type": "text",
538
+ "text": "The proof of Theorem 2 is in Appendix A.2. As a consequence of Theorem 2, the statistical error of estimating the unknown distribution based on the empirical distribution of $n$ i.i.d data under DSW distance is $C _ { d } \\cdot n ^ { - 1 / 2 }$ with high probability where $C _ { d }$ is some universal constant depending on dimension $d$ (see Appendix B.3). Therefore, as other sliced-based Wasserstein distances, the DSW distance does not suffer from the curse of dimensionality. ",
539
+ "bbox": [
540
+ 173,
541
+ 272,
542
+ 825,
543
+ 344
544
+ ],
545
+ "page_idx": 4
546
+ },
547
+ {
548
+ "type": "text",
549
+ "text": "3.2 COMPUTATION OF DSW DISTANCE ",
550
+ "text_level": 1,
551
+ "bbox": [
552
+ 176,
553
+ 361,
554
+ 457,
555
+ 376
556
+ ],
557
+ "page_idx": 4
558
+ },
559
+ {
560
+ "type": "text",
561
+ "text": "Direct computation of DSW distance is challenging. Hence we consider a dual form of DSW distance and a reparametrization of $\\sigma$ as follows. ",
562
+ "bbox": [
563
+ 173,
564
+ 387,
565
+ 821,
566
+ 416
567
+ ],
568
+ "page_idx": 4
569
+ },
570
+ {
571
+ "type": "text",
572
+ "text": "Definition 2. For any $p \\geq 1$ and admissible $C > 0$ , there exists a non-negative constant $\\lambda _ { C }$ depending on $C$ such that the dual form of DSW distance takes the following form ",
573
+ "bbox": [
574
+ 174,
575
+ 419,
576
+ 820,
577
+ 449
578
+ ],
579
+ "page_idx": 4
580
+ },
581
+ {
582
+ "type": "equation",
583
+ "img_path": "images/4e5158cd418ab423ff1ca4a813026fb54b5a6a40f56f251f0e116f30ef3ed9eb.jpg",
584
+ "text": "$$\n\\begin{array} { r } { { 2 } S W _ { p } ^ { * } ( \\mu , \\nu ; C ) = \\underset { \\sigma \\in \\mathbb { H } } { \\operatorname* { s u p } } \\left. \\left( \\mathbb { E } _ { \\theta \\sim \\sigma } \\left[ W _ { p } ^ { p } ( \\mathcal { R } I _ { \\mu } ( \\cdot , \\theta ) , \\mathcal { R } I _ { \\nu } ( \\cdot , \\theta ) ) \\right] \\right) ^ { \\frac { 1 } { p } } - \\lambda _ { C } \\mathbb { E } _ { \\theta , \\theta ^ { \\prime } \\sim \\sigma } \\left[ | \\theta ^ { \\top } \\theta ^ { \\prime } | \\right] \\right. + \\lambda _ { C } C , } \\end{array}\n$$",
585
+ "text_format": "latex",
586
+ "bbox": [
587
+ 181,
588
+ 454,
589
+ 823,
590
+ 497
591
+ ],
592
+ "page_idx": 4
593
+ },
594
+ {
595
+ "type": "text",
596
+ "text": "where M denotes the space of all probability measures on the unit sphere $\\mathbb { S } ^ { d - 1 }$ ",
597
+ "bbox": [
598
+ 176,
599
+ 505,
600
+ 692,
601
+ 520
602
+ ],
603
+ "page_idx": 4
604
+ },
605
+ {
606
+ "type": "text",
607
+ "text": "By the Lagrangian duality theory, $\\mathrm { D S W } _ { p } ( \\mu , \\nu ; C ) \\geq \\mathrm { D S W } _ { p } ^ { * } ( \\mu , \\nu ; C )$ for any $p \\geq 1$ and admissible $C > 0$ . In Definition 2, the set $\\mathbb { M } _ { C }$ disappears and $\\lambda _ { C }$ plays the tuning role for the regularized term $\\mathbb { E } _ { \\theta , \\theta ^ { \\prime } \\sim \\sigma } \\big [ | \\theta ^ { \\top } \\theta ^ { \\prime } | \\big ]$ . When $\\lambda _ { C }$ is large, $\\mathbb { E } _ { \\theta , \\theta ^ { \\prime } \\sim \\sigma } ^ { \\sim } \\big [ | \\theta ^ { \\top } \\theta ^ { \\prime } | \\big ]$ needs to be small, meaning that $C$ is small. When $\\lambda _ { C }$ is small, the value of $\\mathbb { E } _ { \\theta , \\theta ^ { \\prime } \\sim \\sigma } \\big [ | \\theta ^ { \\top } \\theta ^ { \\prime } | \\big ]$ becomes less important, i.e., $C$ is large. ",
608
+ "bbox": [
609
+ 173,
610
+ 530,
611
+ 825,
612
+ 592
613
+ ],
614
+ "page_idx": 4
615
+ },
616
+ {
617
+ "type": "text",
618
+ "text": "For reparametrizing the measure $\\sigma$ , we use a pushforward of uniform measure on the unit sphere through some measurable function $f$ . In particular, let $f$ be a Borel measurable function from $\\mathbb { S } ^ { d - 1 }$ to ${ \\mathbb S } ^ { d - 1 }$ . For any Borel set $A \\subset \\mathbb { S } ^ { d - \\mathrm { \\ i } }$ , we define $\\sigma ( A ) \\stackrel { \\circ } { = } \\sigma ^ { d - 1 } ( f ^ { - 1 } ( A ) )$ , where $\\sigma ^ { d - 1 }$ is the uniform probability measure on $\\mathbb { S } ^ { d - 1 }$ . Then for any Borel measurable function $g : { \\mathbb { S } } ^ { d - 1 } \\mathbb { R }$ , we have $\\begin{array} { r } { \\int _ { \\theta \\sim \\sigma } g ( \\theta ) \\dot { d } \\sigma ( \\theta ) = \\int _ { \\theta \\sim \\sigma ^ { d - 1 } } ( g \\circ f ) ( \\theta ) d \\sigma ^ { d - \\bar { 1 } } ( \\theta ) } \\end{array}$ . Therefore, we obtain the equivalent dual form of DSW as follows: ",
619
+ "bbox": [
620
+ 173,
621
+ 597,
622
+ 826,
623
+ 680
624
+ ],
625
+ "page_idx": 4
626
+ },
627
+ {
628
+ "type": "equation",
629
+ "img_path": "images/7da9016419f851b69e972abacc776cbd9a36115619fc92416ebb81b8e1d72c50.jpg",
630
+ "text": "$$\n\\begin{array} { r l } & { \\mathrm { D S W } _ { p } ^ { * } ( \\mu , \\nu ; C ) = \\underset { f \\in \\mathcal { F } } { \\operatorname* { s u p } } \\Bigg \\{ \\Bigg ( \\mathbb { E } _ { \\theta \\sim \\sigma ^ { d - 1 } } \\big [ W _ { p } ^ { p } \\big ( \\mathcal { R } I _ { \\mu } ( \\cdot , f ( \\theta ) ) , \\mathcal { R } I _ { \\nu } ( \\cdot , f ( \\theta ) ) \\big ) \\big ] \\Bigg ) ^ { 1 / p } } \\\\ & { \\qquad - \\lambda _ { C } \\mathbb { E } _ { \\theta , \\theta ^ { \\prime } \\sim \\sigma ^ { d - 1 } } \\Big [ \\big | f ( \\theta ) ^ { \\top } f ( \\theta ^ { \\prime } ) \\big | \\Big ] \\Bigg \\} + \\lambda _ { C } C : = \\underset { f \\in \\mathcal { F } } { \\operatorname* { s u p } } \\mathrm { D S } ( f ) , } \\end{array}\n$$",
631
+ "text_format": "latex",
632
+ "bbox": [
633
+ 199,
634
+ 684,
635
+ 799,
636
+ 770
637
+ ],
638
+ "page_idx": 4
639
+ },
640
+ {
641
+ "type": "text",
642
+ "text": "where $\\mathcal { F }$ is a class of all Borel measurable functions from $\\mathbb { S } ^ { d - 1 }$ to $\\mathbb { S } ^ { d - 1 }$ . ",
643
+ "bbox": [
644
+ 173,
645
+ 776,
646
+ 647,
647
+ 792
648
+ ],
649
+ "page_idx": 4
650
+ },
651
+ {
652
+ "type": "text",
653
+ "text": "Finding the optimal $f$ : We parameterize $f$ in the dual form (2) by using a deep neural network with parameter $\\phi$ , defined as $f _ { \\phi }$ . Then, we estimate the gradient of the objective function $\\mathrm { D S } ( f _ { \\phi } )$ in equation (2) with respect to $\\phi$ and use stochastic gradient ascent algorithm to update $\\phi$ . Since there are expectations over uniform distribution in the gradient of $\\mathrm { D S } ( f _ { \\phi } )$ , we use the Monte Carlo method to approximate these expectations. Note that, we can use the fixed point from the stochastic ascent algorithm to approximate the dual value of DSW in equation (2). A detailed argument for this point is in Appendix B.2. Finally, in generative model applications with DSW being the loss function, we only need to use the gradient of the function $\\mathrm { D S } ( . )$ to update the parameters of interest. Therefore, we can treat $\\lambda _ { C }$ as a regularized parameter and tune it to find suitable value in these applications. ",
654
+ "bbox": [
655
+ 173,
656
+ 797,
657
+ 826,
658
+ 924
659
+ ],
660
+ "page_idx": 4
661
+ },
662
+ {
663
+ "type": "image",
664
+ "img_path": "images/e1cac424355623855569b7071b820cc80a55f82017ef08e0ff626e2af202f5c0.jpg",
665
+ "image_caption": [
666
+ "Figure 1: Empirical behavior of optimal measure $\\sigma$ , approximated by 1000 samples, on a circle for different values of $\\lambda _ { C }$ (the constant in the dual form of DSW in Definition 2) when $\\mu$ and $\\nu$ are bivariate Gaussian distributions sharing the same eigenvectors. When $\\lambda _ { C } = 0$ , $C = 1$ . When $\\lambda _ { C }$ increases, $C$ becomes small. "
667
+ ],
668
+ "image_footnote": [],
669
+ "bbox": [
670
+ 192,
671
+ 101,
672
+ 807,
673
+ 223
674
+ ],
675
+ "page_idx": 5
676
+ },
677
+ {
678
+ "type": "text",
679
+ "text": "Illustration of the roles of $\\lambda _ { C }$ and $C$ : To illustrate the roles of $\\lambda _ { C }$ and $C$ in finding optimal distribution $\\sigma$ , we conduct a simple experiment on two Gaussian distributions with zero means and covariance matrices given by $\\left( \\begin{array} { l l } { 2 } & { 0 } \\\\ { 0 } & { 2 } \\end{array} \\right)$ and $\\left( \\begin{array} { l l } { 5 } & { 0 } \\\\ { 0 } & { 1 } \\end{array} \\right)$ . The experiment optimizes the empirical form of Definition 2 with different choices of $\\lambda _ { C }$ . The results are shown in Figure 1 with the reported value of $\\lambda _ { C }$ and $\\mathbb { E } _ { \\theta , \\theta ^ { \\prime } \\sim \\sigma } \\big [ | \\theta ^ { \\top } \\theta ^ { \\prime } | \\big ]$ . For $\\lambda _ { C } = 0$ , the obtained distribution concentrates only on one direction. When $\\lambda _ { C } = 5 0$ , optimal $\\sigma$ distributes more weights to other directions on the circle. When $\\lambda _ { C } = 1 0 0 0$ , optimal $\\sigma$ is close to the discrete distribution concentrated on two eigenvectors of the covariance matrices, which are the main directions differentiating the two Gaussian distributions. ",
680
+ "bbox": [
681
+ 174,
682
+ 304,
683
+ 825,
684
+ 425
685
+ ],
686
+ "page_idx": 5
687
+ },
688
+ {
689
+ "type": "text",
690
+ "text": "Extension of DSW and comparison of DSW to Max-GSW-NN: Similar to SW, we extend DSW to distributional generalized sliced Wasserstein (DGSW) by using the non-linear projecting operator via GRT. The definition of the DGSW and its properties are in Appendix C. Finally, in Appendix E.1, we show the distinction of the DSW to Max-GSW-NN (Kolouri et al., 2019) when the neural network defining function in Max-GSW-NN is $g ( x , \\theta ) = \\langle x , f ( \\theta ) \\rangle$ where $f : \\mathbb { S } ^ { d - 1 } \\to \\mathbb { S } ^ { d - 1 }$ . ",
691
+ "bbox": [
692
+ 174,
693
+ 433,
694
+ 825,
695
+ 502
696
+ ],
697
+ "page_idx": 5
698
+ },
699
+ {
700
+ "type": "text",
701
+ "text": "4 EXPERIMENTS ",
702
+ "text_level": 1,
703
+ "bbox": [
704
+ 176,
705
+ 525,
706
+ 326,
707
+ 540
708
+ ],
709
+ "page_idx": 5
710
+ },
711
+ {
712
+ "type": "text",
713
+ "text": "In this section, we conduct extensive experiments comparing the performance in both generative quality and computational speed of the proposed DSW distance with other sliced-based distances, namely the SW, Max-SW, Max-GSW-NN (Kolouri et al., 2019) and projected robust subspace Wasserstein (PRW) (Paty & Cuturi, 2019; Lin et al., 2020) using the minimum expected distance estimator (MEDE) (Bernton et al., 2019) on MNIST (LeCun et al., 1998), CIFAR10 (Krizhevsky, 2009), CelebA (Liu et al., 2015) and LSUN (Yu et al., 2015) datasets. The details of the MEDE framework are described in Appendix D. We would like to note that the wall-clock timing of different methods may be subject to the differences in the hyperparameter settings and software implementations of different methods. On MNIST dataset, we train generative models with different distances and then evaluate their performances by comparing Wasserstein-2 distances between 10000 random generated images and all images from the MNIST test set. Due to the very large size of other datasets, e.g., 3 million images in LSUN, it is expensive to compute empirical Wasserstein-2 distance as its complexity is of order ${ \\mathcal { O } } ( k ^ { 2 } \\log k )$ where $k$ is the number of support points. Therefore, after we train generative models, we use FID score (Heusel et al., 2017) to evaluate the generative quality of these generators. The FID score is calculated from 10000 random generated images and all training samples using precomputed statistics in (Heusel et al., 2017). Finally, for $\\lambda _ { C }$ in DSW (see Definition 2), it is chosen in the set $\\{ 1 , 1 0 , 1 0 0 , 1 0 0 0 \\}$ such that its Wasserstein-2 (FID score) (between 10000 random generated images and all images from corresponding validation set) is the lowest among the four values. Detailed experiment settings are in Appendix G. Finally, we also apply the DSW into color transfer task (Rabin et al., 2010; 2014; Bonneel et al., 2015; Perrot et al., 2016) in Appendix F, where we find that DSW also performs better than SW and Max-SW in this task. ",
714
+ "bbox": [
715
+ 174,
716
+ 556,
717
+ 826,
718
+ 849
719
+ ],
720
+ "page_idx": 5
721
+ },
722
+ {
723
+ "type": "text",
724
+ "text": "4.1 RESULTS ON MNIST ",
725
+ "text_level": 1,
726
+ "bbox": [
727
+ 176,
728
+ 868,
729
+ 361,
730
+ 882
731
+ ],
732
+ "page_idx": 5
733
+ },
734
+ {
735
+ "type": "text",
736
+ "text": "Generative quality and computational speed: We report the performance of the learned generative models for MNIST in Figure 2(a). To plot this figure, we vary the number of projections $N \\in$ $\\{ 1 , 1 0 , 1 0 ^ { 2 } , 5 \\times 1 0 ^ { 2 } , 1 0 ^ { 3 } , 5 \\times 1 0 ^ { 3 } , 1 0 ^ { 4 } \\}$ for the SW, and $N \\in \\{ 1 , 1 0 , 1 0 ^ { 2 } , 5 \\times 1 0 ^ { 2 } , 1 0 ^ { 3 } , 5 \\times 1 0 ^ { 3 } \\}$ for the DSW. Then we measure the computational time per minibatch and the Wasserstein-2 score of the learned generators for each $N$ . We plot the Wasserstein-2 score and computational time of Max-SW and Max-GSW-NN in their standard settings (Kolouri et al., 2019). Except for the regime with very fast but low-quality learned models, DSW is better than all the existing slice-based baselines in terms of both model quality and computational speed. Moreover, DSW can learn good models with very few projections, e.g., DSW-10 achieves better model quality than Max-GSW-NN and Max-SW and is one order-of-magnitude faster than these sliced distances. Finally, with a similar computational time, a learned generator by DSW has the Wasserstein-2 score that is roughly $1 0 \\%$ lower than the one got from SW. For the qualitative comparison between these distances, we show random generated images from their generative models in Figure 7 in Appendix E.1. We observe that generated images from DSW are sharper and easier to classify into numbers than those from other baseline distances. ",
737
+ "bbox": [
738
+ 173,
739
+ 895,
740
+ 823,
741
+ 924
742
+ ],
743
+ "page_idx": 5
744
+ },
745
+ {
746
+ "type": "image",
747
+ "img_path": "images/05313ea7c1db89d713694264a662043e8a53b0f9cae59afe10da003d67a7180a.jpg",
748
+ "image_caption": [
749
+ "Figure 2: (a) Comparison between DSW, SW, Max-SW, Max-GSW-NN, PRW and WD based on execution time and performance. Here, each dot of SW and DSW corresponds to the number of projections chosen in $\\{ 1 , 1 0 , 1 0 ^ { \\cdot 2 } , 5 \\times 1 0 ^ { 2 } , 1 0 ^ { 3 } , 5 \\times 1 0 ^ { 3 } , 1 0 ^ { 4 } \\}$ . Each dot of PRW corresponds to the dimension of the subspace chosen in $\\{ 2 , 5 , 1 0 , 5 0 \\}$ ; (b) Comparison between SW, DSW, Max-SW and Max-GSW-NN based on Wasserstein-2 distance between distributions of learned model and test set over iterations; (c) Computation speed of distances based on the number of minibatch’s samples (log-log scale); (d) Effect of $\\lambda _ { C }$ on the mean of absolute values of pairwise cosine similarity between 10 random directions from the found distribution $\\sigma$ of DSW. "
750
+ ],
751
+ "image_footnote": [],
752
+ "bbox": [
753
+ 186,
754
+ 101,
755
+ 808,
756
+ 228
757
+ ],
758
+ "page_idx": 6
759
+ },
760
+ {
761
+ "type": "text",
762
+ "text": "",
763
+ "bbox": [
764
+ 173,
765
+ 358,
766
+ 825,
767
+ 526
768
+ ],
769
+ "page_idx": 6
770
+ },
771
+ {
772
+ "type": "text",
773
+ "text": "Comparison with projected robust subspace Wasserstein (PRW) and Wasserstein distance: In Figure 2(a), we plot the Wasserstein-2 score and computational time of Wasserstein distance (WD) and PRW, where the subspace dimension of PRW varies in the range $\\{ 2 , 5 , 1 0 , 5 0 \\}$ . PRW is able to improve upon the model quality of slice-based methods including DSW, however at the cost of being an order of magnitude slower than DSW with 10 projections (DSW-10). We observe that DSW-10 obtains a better Wasserstein-2 score than PRW with 5-dimensional subspace, while its corresponding computational time is 30 times faster than that of PRW-5. Using 50 dimension, PRW’s Wasserstein-2 score improves about $2 9 \\%$ to that of DSW-10 but the computational cost is also around 40 times slower. The model trained by WD gives good Wasserstein-2 score; however, it is computational expensive (about 40 times slower than DSW-10). The main computational advantage of DSW comes from the exact calculation of Wasserstein distance in one-dimension. The visual comparison between PRW, WD and DSW based on their generated images is in Figure 12 in Appendix E.2. ",
774
+ "bbox": [
775
+ 173,
776
+ 534,
777
+ 825,
778
+ 700
779
+ ],
780
+ "page_idx": 6
781
+ },
782
+ {
783
+ "type": "text",
784
+ "text": "Convergence behavior: Figure 2(b) shows that DSW learns better models at a faster speed of convergence than other baseline distances with a very small number of projections, e.g., DSW-10 is the second lowest curve compared to curves from other sliced-based distances. ",
785
+ "bbox": [
786
+ 174,
787
+ 707,
788
+ 825,
789
+ 750
790
+ ],
791
+ "page_idx": 6
792
+ },
793
+ {
794
+ "type": "text",
795
+ "text": "Scalability over sample size of minibatch: Results in Figure 2(c) show that DSW has a computational complexity of the order ${ \\mathcal { O } } ( k \\log k )$ , which is similar to those of other sliced-based distances, where $k$ is the number of samples per batch. ",
796
+ "bbox": [
797
+ 174,
798
+ 756,
799
+ 825,
800
+ 797
801
+ ],
802
+ "page_idx": 6
803
+ },
804
+ {
805
+ "type": "text",
806
+ "text": "Effect of the reguloptimal distribution ization paraof DSW with $\\lambda _ { C }$ : For each value of projections, and the $\\lambda _ { C } \\in \\{ 1 , 1 0 , 1 0 0 , 1 0 0 0 \\}$ $\\sigma$ $N = 1 0$ $\\begin{array} { r } { A _ { N } = \\frac { 1 } { N ^ { 2 } } \\sum _ { i , j = 1 } ^ { N } | \\boldsymbol { \\theta } _ { i } ^ { \\top } \\boldsymbol { \\theta } _ { j } | } \\end{array}$ an approximation of the regularized term $\\mathbb { E } _ { \\theta , \\theta ^ { \\prime } \\sim \\sigma } \\left[ \\left| \\theta ^ { \\top } \\theta ^ { \\prime } \\right| \\right]$ in the dual form of DSW in equation (2), where $\\{ \\theta _ { i } \\} _ { i = 1 } ^ { N } \\sim \\sigma$ . The results are shown in Figure 2(d). We observe that when $\\lambda _ { C }$ increases, $A _ { N }$ goes down. When $\\lambda _ { C } = 0$ , i.e., no regularization, $A _ { N }$ gets close to 1, meaning that all projected directions collapse to one direction. When $\\lambda _ { C } = 1 0 0 0$ , $A _ { N }$ is close to 0.1, suggesting that all projected directions are nearly orthogonal. ",
807
+ "bbox": [
808
+ 173,
809
+ 804,
810
+ 826,
811
+ 912
812
+ ],
813
+ "page_idx": 6
814
+ },
815
+ {
816
+ "type": "image",
817
+ "img_path": "images/2f0215c39c5f57dc4c0277750b8548580cc04641cc662db46352e1c8edb4a65c.jpg",
818
+ "image_caption": [
819
+ "Figure 3: Comparison between DSW, SW, Max-SW and Max-GSW-NN in terms of execution time and performance. Here, each dot of SW and DSW corresponds to the number of projections chosen in $\\{ 1 0 ^ { 2 } , 5 \\times$ $\\mathrm { \\dot { 1 } 0 ^ { 2 } , 1 0 ^ { 3 } , 5 \\times 1 0 ^ { 3 } , 1 0 ^ { 4 } } \\}$ . We set the minibatch size be 512 on CelebA and CIFAR, and be 4096 on LSUN. "
820
+ ],
821
+ "image_footnote": [],
822
+ "bbox": [
823
+ 199,
824
+ 101,
825
+ 800,
826
+ 247
827
+ ],
828
+ "page_idx": 7
829
+ },
830
+ {
831
+ "type": "text",
832
+ "text": "Additional experiments: We also investigate how the number of gradient-steps used for updating distribution of directions $\\sigma$ , and how the size of minibatches affects the quality of DSW (see Appendix E.1). The results show that an increasing number of gradient steps to update $\\sigma$ leads to better performance of DSW but also slows down the computation speed. Furthermore, we carry out experiments with DGSW, an extension of DSW to non-linear projections, and test the new proposed distances in training encoder-generator models on MNIST using joint contrastive inference (JCI) in Appendices E.1 and E.3. The description of these models is in Appendix D. ",
833
+ "bbox": [
834
+ 174,
835
+ 332,
836
+ 825,
837
+ 429
838
+ ],
839
+ "page_idx": 7
840
+ },
841
+ {
842
+ "type": "text",
843
+ "text": "4.2 RESULTS ON LARGE-SCALE DATASETS ",
844
+ "text_level": 1,
845
+ "bbox": [
846
+ 176,
847
+ 462,
848
+ 482,
849
+ 474
850
+ ],
851
+ "page_idx": 7
852
+ },
853
+ {
854
+ "type": "text",
855
+ "text": "Next, we conduct large-scale experiments on a range of more realistic image datasets. We train generative models using CIFAR10, CelebA, and LSUN datasets (all these datasets are rescaled to $6 4 \\mathrm { x } 6 4$ resolution). When working with high dimensional distributions, Deshpande et al. (2018) proposed a trick to improve the quality of the generator by learning a feature function which maps data to a new feature space that is more manageable in size. When the feature function is fixed, the generator is trained to match the distribution of features. When the generator is fixed, the feature function tries to tease apart the data empirical features from the generated feature distribution. For the experiments in this section, we use the same technique with DSW and all other baseline distances. ",
856
+ "bbox": [
857
+ 174,
858
+ 492,
859
+ 825,
860
+ 603
861
+ ],
862
+ "page_idx": 7
863
+ },
864
+ {
865
+ "type": "text",
866
+ "text": "We compare DSW with SW, Max-SW, and Max-GSW-NN in both generative quality (FID score) and computational time in Figure 3. We could not compare DSW with PRW on the large-scale datasets since PRW is computationally expensive to train to obtain good generated images. On CelebA and CIFAR10, we let $N$ , the number of projections of both DSW and SW, vary in the set $\\{ 1 0 ^ { 2 } , 5 \\times 1 0 ^ { 2 } , 1 0 ^ { 3 } , 5 \\times 1 0 ^ { 3 } , 1 0 ^ { 4 } \\}$ . For LSUN, since it takes considerably longer time to train each model, we only vary $N$ in the set $\\{ 1 0 ^ { 2 } , 1 0 ^ { 4 } \\}$ . On all these large datasets, DSW outperforms all the other baselines in both FID score of the learned model and computational efficiency. The gap of FID scores between DSW and other methods is especially large on CIFAR10 and LSUN. For example, on CIFAR10, with the same computational time, FID scores of DSW are always lower than those of SW about 20 units. On LSUN, with 100 projections, DSW can achieve an FID score of 46 while SW with 10000 projections still has a worse FID score of over 60. It is interesting to note that on these high-dimensional datasets, Max-SW performs rather poorly: it obtains the highest FID scores among all distances while requires heavy computation. Max-GSW-NN has better FID scores than (Max)-SW; however, it is still worse than DSW and while being slower. This is consistent with the intuition that as the number of dimension of the data grows, the use of a single important slice in Max-SW becomes a less efficient approximation. DSW, on the other hand, is able to make use of more important slices, and at the same time avoids SW’s inefficiency of uniform slice-sampling. ",
867
+ "bbox": [
868
+ 174,
869
+ 611,
870
+ 825,
871
+ 847
872
+ ],
873
+ "page_idx": 7
874
+ },
875
+ {
876
+ "type": "text",
877
+ "text": "Generated images from CelebA, CIFAR10 and LSUN are deferred to Appendix E.1. Comparing to other sliced-based Wasserstein distances, generated samples obtained from the DSW’s generative model are also more visually realistic. Further experiments to compare DGSW with GSW, Max-GSW, and Max-GSW-NN are also given in the Appendix E.1. Based on these experiments, we can conclude that the distributional approach also improves the generative quality of non-linear slicing distances. ",
878
+ "bbox": [
879
+ 176,
880
+ 853,
881
+ 825,
882
+ 924
883
+ ],
884
+ "page_idx": 7
885
+ },
886
+ {
887
+ "type": "text",
888
+ "text": "5 CONCLUSION ",
889
+ "text_level": 1,
890
+ "bbox": [
891
+ 176,
892
+ 102,
893
+ 318,
894
+ 117
895
+ ],
896
+ "page_idx": 8
897
+ },
898
+ {
899
+ "type": "text",
900
+ "text": "In this paper, we have presented the novel distributional sliced-Wasserstein (DSW) distances between two probability measures. Our main idea is to search for the best distribution of important directions while regularizing towards orthogonal directions. We prove that they are well-defined metrics and provide their theoretical and computational properties. We compare our proposed distances to other sliced-based distances in a variety of generative modeling tasks, including estimating generative models and jointly estimating both generators and inference models. Extensive experiments demonstrate that our new distances yield significantly better models and convergence behaviors during training than the previous sliced-based distances. One important future direction is to investigate theoretically the optimal choice of the regularization parameter $\\lambda _ { C }$ such that the DSW distance can capture all the important directions that can distinguish two target probability measures well. ",
901
+ "bbox": [
902
+ 174,
903
+ 132,
904
+ 825,
905
+ 272
906
+ ],
907
+ "page_idx": 8
908
+ },
909
+ {
910
+ "type": "text",
911
+ "text": "REFERENCES ",
912
+ "text_level": 1,
913
+ "bbox": [
914
+ 176,
915
+ 103,
916
+ 287,
917
+ 117
918
+ ],
919
+ "page_idx": 9
920
+ },
921
+ {
922
+ "type": "text",
923
+ "text": "Jason Altschuler, Jonathan Niles-Weed, and Philippe Rigollet. Near-linear time approximation algorithms for optimal transport via Sinkhorn iteration. In Advances in neural information processing systems, pp. 1964–1974, 2017. ",
924
+ "bbox": [
925
+ 174,
926
+ 126,
927
+ 823,
928
+ 169
929
+ ],
930
+ "page_idx": 9
931
+ },
932
+ {
933
+ "type": "text",
934
+ "text": "Luca Ambrogioni, Umut Güçlü, Yagmur Güçlütürk, Max Hinne, Marcel AJ van Gerven, and Eric ˘ Maris. Wasserstein variational inference. In Advances in Neural Information Processing Systems, pp. 2473–2482, 2018. ",
935
+ "bbox": [
936
+ 174,
937
+ 178,
938
+ 823,
939
+ 219
940
+ ],
941
+ "page_idx": 9
942
+ },
943
+ {
944
+ "type": "text",
945
+ "text": "Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In International Conference on Machine Learning, pp. 214–223, 2017. ",
946
+ "bbox": [
947
+ 171,
948
+ 228,
949
+ 825,
950
+ 258
951
+ ],
952
+ "page_idx": 9
953
+ },
954
+ {
955
+ "type": "text",
956
+ "text": "Erhan Bayraktar and Gaoyue Guo. Strong equivalence between metrics of Wasserstein type. arXiv preprint arXiv:1912.08247, 2019. ",
957
+ "bbox": [
958
+ 171,
959
+ 267,
960
+ 825,
961
+ 296
962
+ ],
963
+ "page_idx": 9
964
+ },
965
+ {
966
+ "type": "text",
967
+ "text": "Espen Bernton, Pierre E Jacob, Mathieu Gerber, and Christian P Robert. On parameter estimation with the Wasserstein distance. Information and Inference: A Journal of the IMA, 8(4):657–676, 2019. ",
968
+ "bbox": [
969
+ 176,
970
+ 304,
971
+ 825,
972
+ 347
973
+ ],
974
+ "page_idx": 9
975
+ },
976
+ {
977
+ "type": "text",
978
+ "text": "Gregory Beylkin. The inversion problem and applications of the generalized Radon transform. Communications on pure and applied mathematics, 37(5):579–599, 1984. ",
979
+ "bbox": [
980
+ 176,
981
+ 357,
982
+ 823,
983
+ 386
984
+ ],
985
+ "page_idx": 9
986
+ },
987
+ {
988
+ "type": "text",
989
+ "text": "Nicolas Bonneel and David Coeurjolly. Spot: sliced partial optimal transport. ACM Transactions on Graphics (TOG), 38(4):1–13, 2019. ",
990
+ "bbox": [
991
+ 176,
992
+ 393,
993
+ 823,
994
+ 424
995
+ ],
996
+ "page_idx": 9
997
+ },
998
+ {
999
+ "type": "text",
1000
+ "text": "Nicolas Bonneel, Julien Rabin, Gabriel Peyré, and Hanspeter Pfister. Sliced and Radon Wasserstein barycenters of measures. Journal of Mathematical Imaging and Vision, 1(51):22–45, 2015. ",
1001
+ "bbox": [
1002
+ 174,
1003
+ 433,
1004
+ 823,
1005
+ 462
1006
+ ],
1007
+ "page_idx": 9
1008
+ },
1009
+ {
1010
+ "type": "text",
1011
+ "text": "Nicolas Bonnotte. Unidimensional and evolution methods for optimal transportation. PhD thesis, Paris 11, 2013. ",
1012
+ "bbox": [
1013
+ 174,
1014
+ 469,
1015
+ 825,
1016
+ 500
1017
+ ],
1018
+ "page_idx": 9
1019
+ },
1020
+ {
1021
+ "type": "text",
1022
+ "text": "Charlotte Bunne, David Alvarez-Melis, Andreas Krause, and Stefanie Jegelka. Learning generative models across incomparable spaces. In International Conference on Machine Learning, 2019. ",
1023
+ "bbox": [
1024
+ 173,
1025
+ 508,
1026
+ 823,
1027
+ 537
1028
+ ],
1029
+ "page_idx": 9
1030
+ },
1031
+ {
1032
+ "type": "text",
1033
+ "text": "Xiongjie Chen, Yongxin Yang, and Yunpeng Li. Augmented sliced Wasserstein distances. arXiv preprint arXiv:2006.08812, 2020. ",
1034
+ "bbox": [
1035
+ 173,
1036
+ 546,
1037
+ 823,
1038
+ 575
1039
+ ],
1040
+ "page_idx": 9
1041
+ },
1042
+ {
1043
+ "type": "text",
1044
+ "text": "Lenaic Chizat, Gabriel Peyré, Bernhard Schmitzer, and François-Xavier Vialard. Scaling algorithms for unbalanced optimal transport problems. Mathematics of Computation, 87(314):2563–2609, 2018. ",
1045
+ "bbox": [
1046
+ 173,
1047
+ 584,
1048
+ 826,
1049
+ 627
1050
+ ],
1051
+ "page_idx": 9
1052
+ },
1053
+ {
1054
+ "type": "text",
1055
+ "text": "Nicolas Courty, Rémi Flamary, and Devis Tuia. Domain adaptation with regularized optimal transport. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pp. 274–289. Springer, 2014. ",
1056
+ "bbox": [
1057
+ 173,
1058
+ 636,
1059
+ 826,
1060
+ 679
1061
+ ],
1062
+ "page_idx": 9
1063
+ },
1064
+ {
1065
+ "type": "text",
1066
+ "text": "Nicolas Courty, Rémi Flamary, Amaury Habrard, and Alain Rakotomamonjy. Joint distribution optimal transportation for domain adaptation. In Advances in Neural Information Processing Systems, pp. 3730–3739, 2017. ",
1067
+ "bbox": [
1068
+ 173,
1069
+ 688,
1070
+ 825,
1071
+ 731
1072
+ ],
1073
+ "page_idx": 9
1074
+ },
1075
+ {
1076
+ "type": "text",
1077
+ "text": "Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In Advances in neural information processing systems, pp. 2292–2300, 2013. ",
1078
+ "bbox": [
1079
+ 169,
1080
+ 739,
1081
+ 823,
1082
+ 770
1083
+ ],
1084
+ "page_idx": 9
1085
+ },
1086
+ {
1087
+ "type": "text",
1088
+ "text": "Ishan Deshpande, Ziyu Zhang, and Alexander G Schwing. Generative modeling using the sliced Wasserstein distance. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 3483–3491, 2018. ",
1089
+ "bbox": [
1090
+ 174,
1091
+ 777,
1092
+ 825,
1093
+ 820
1094
+ ],
1095
+ "page_idx": 9
1096
+ },
1097
+ {
1098
+ "type": "text",
1099
+ "text": "Ishan Deshpande, Yuan-Ting Hu, Ruoyu Sun, Ayis Pyrros, Nasir Siddiqui, Sanmi Koyejo, Zhizhen Zhao, David Forsyth, and Alexander G Schwing. Max-sliced Wasserstein distance and its use for GANs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 10648–10656, 2019. ",
1100
+ "bbox": [
1101
+ 173,
1102
+ 829,
1103
+ 826,
1104
+ 886
1105
+ ],
1106
+ "page_idx": 9
1107
+ },
1108
+ {
1109
+ "type": "text",
1110
+ "text": "Vincent Dumoulin, Ishmael Belghazi, Ben Poole, Olivier Mastropietro, Alex Lamb, Martin Arjovsky, and Aaron Courville. Adversarially learned inference. arXiv preprint arXiv:1606.00704, 2016. ",
1111
+ "bbox": [
1112
+ 173,
1113
+ 895,
1114
+ 823,
1115
+ 924
1116
+ ],
1117
+ "page_idx": 9
1118
+ },
1119
+ {
1120
+ "type": "text",
1121
+ "text": "R’emi Flamary and Nicolas Courty. Pot python optimal transport library, 2017. URL https: //pythonot.github.io/. ",
1122
+ "bbox": [
1123
+ 171,
1124
+ 103,
1125
+ 825,
1126
+ 132
1127
+ ],
1128
+ "page_idx": 10
1129
+ },
1130
+ {
1131
+ "type": "text",
1132
+ "text": "Aude Genevay, Gabriel Peyre, and Marco Cuturi. Learning generative models with Sinkhorn divergences. In International Conference on Artificial Intelligence and Statistics, pp. 1608–1617, 2018. ",
1133
+ "bbox": [
1134
+ 176,
1135
+ 140,
1136
+ 823,
1137
+ 183
1138
+ ],
1139
+ "page_idx": 10
1140
+ },
1141
+ {
1142
+ "type": "text",
1143
+ "text": "Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of Wasserstein Gans. In Advances in neural information processing systems, pp. 5767–5777, 2017. ",
1144
+ "bbox": [
1145
+ 173,
1146
+ 191,
1147
+ 825,
1148
+ 234
1149
+ ],
1150
+ "page_idx": 10
1151
+ },
1152
+ {
1153
+ "type": "text",
1154
+ "text": "Sigurdur Helgason. Integral geometry and Radon transforms. Springer Science & Business Media, 2010. ",
1155
+ "bbox": [
1156
+ 173,
1157
+ 242,
1158
+ 826,
1159
+ 272
1160
+ ],
1161
+ "page_idx": 10
1162
+ },
1163
+ {
1164
+ "type": "text",
1165
+ "text": "Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems, pp. 6626–6637, 2017. ",
1166
+ "bbox": [
1167
+ 174,
1168
+ 280,
1169
+ 823,
1170
+ 324
1171
+ ],
1172
+ "page_idx": 10
1173
+ },
1174
+ {
1175
+ "type": "text",
1176
+ "text": "Diederik $\\mathrm { \\bf P }$ Kingma and Max Welling. Auto-encoding variational Bayes. arXiv preprint arXiv:1312.6114, 2013. ",
1177
+ "bbox": [
1178
+ 173,
1179
+ 330,
1180
+ 823,
1181
+ 361
1182
+ ],
1183
+ "page_idx": 10
1184
+ },
1185
+ {
1186
+ "type": "text",
1187
+ "text": "Philip A Knight. The Sinkhorn–Knopp algorithm: convergence and applications. SIAM Journal on Matrix Analysis and Applications, 30(1):261–275, 2008. ",
1188
+ "bbox": [
1189
+ 171,
1190
+ 368,
1191
+ 825,
1192
+ 397
1193
+ ],
1194
+ "page_idx": 10
1195
+ },
1196
+ {
1197
+ "type": "text",
1198
+ "text": "Max Kochurov, Rasul Karimov, and Serge Kozlukov. Geoopt: Riemannian optimization in pytorch, 2020. ",
1199
+ "bbox": [
1200
+ 173,
1201
+ 405,
1202
+ 825,
1203
+ 435
1204
+ ],
1205
+ "page_idx": 10
1206
+ },
1207
+ {
1208
+ "type": "text",
1209
+ "text": "Soheil Kolouri, Phillip E Pope, Charles E Martin, and Gustavo K Rohde. Sliced Wasserstein auto-encoders. In International Conference on Learning Representations, 2018. ",
1210
+ "bbox": [
1211
+ 173,
1212
+ 443,
1213
+ 823,
1214
+ 472
1215
+ ],
1216
+ "page_idx": 10
1217
+ },
1218
+ {
1219
+ "type": "text",
1220
+ "text": "Soheil Kolouri, Kimia Nadjahi, Umut Simsekli, Roland Badeau, and Gustavo Rohde. Generalized sliced Wasserstein distances. In Advances in Neural Information Processing Systems, pp. 261–272, 2019. ",
1221
+ "bbox": [
1222
+ 174,
1223
+ 479,
1224
+ 825,
1225
+ 522
1226
+ ],
1227
+ "page_idx": 10
1228
+ },
1229
+ {
1230
+ "type": "text",
1231
+ "text": "Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. ",
1232
+ "bbox": [
1233
+ 173,
1234
+ 531,
1235
+ 805,
1236
+ 547
1237
+ ],
1238
+ "page_idx": 10
1239
+ },
1240
+ {
1241
+ "type": "text",
1242
+ "text": "Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998. ",
1243
+ "bbox": [
1244
+ 171,
1245
+ 554,
1246
+ 823,
1247
+ 584
1248
+ ],
1249
+ "page_idx": 10
1250
+ },
1251
+ {
1252
+ "type": "text",
1253
+ "text": "Bing Li, Shaoli Wang, et al. On directional regression for dimension reduction. Journal of the American Statistical Association, 102:997–1008, 2007. ",
1254
+ "bbox": [
1255
+ 171,
1256
+ 592,
1257
+ 823,
1258
+ 622
1259
+ ],
1260
+ "page_idx": 10
1261
+ },
1262
+ {
1263
+ "type": "text",
1264
+ "text": "Ker-Chau Li. Sliced inverse regression for dimension reduction. Journal of the American Statistical Association, 86(414):316–328, 1991. ",
1265
+ "bbox": [
1266
+ 171,
1267
+ 628,
1268
+ 825,
1269
+ 659
1270
+ ],
1271
+ "page_idx": 10
1272
+ },
1273
+ {
1274
+ "type": "text",
1275
+ "text": "Tianyi Lin, Nhat Ho, and Michael Jordan. On efficient optimal transport: An analysis of greedy and accelerated mirror descent algorithms. In International Conference on Machine Learning, pp. 3982–3991, 2019a. ",
1276
+ "bbox": [
1277
+ 173,
1278
+ 666,
1279
+ 825,
1280
+ 709
1281
+ ],
1282
+ "page_idx": 10
1283
+ },
1284
+ {
1285
+ "type": "text",
1286
+ "text": "Tianyi Lin, Nhat Ho, and Michael I Jordan. On the acceleration of the Sinkhorn and Greenkhorn algorithms for optimal transport. arXiv preprint arXiv:1906.01437, 2019b. ",
1287
+ "bbox": [
1288
+ 169,
1289
+ 718,
1290
+ 823,
1291
+ 747
1292
+ ],
1293
+ "page_idx": 10
1294
+ },
1295
+ {
1296
+ "type": "text",
1297
+ "text": "Tianyi Lin, Chenyou Fan, Nhat Ho, Marco Cuturi, and Michael I. Jordan. Projection robust Wasserstein distance and Riemannian optimization. arXiv preprint arXiv:2006.07458, 2020. ",
1298
+ "bbox": [
1299
+ 171,
1300
+ 755,
1301
+ 825,
1302
+ 785
1303
+ ],
1304
+ "page_idx": 10
1305
+ },
1306
+ {
1307
+ "type": "text",
1308
+ "text": "Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), December 2015. ",
1309
+ "bbox": [
1310
+ 171,
1311
+ 792,
1312
+ 825,
1313
+ 821
1314
+ ],
1315
+ "page_idx": 10
1316
+ },
1317
+ {
1318
+ "type": "text",
1319
+ "text": "Antoine Liutkus, Umut Simsekli, Szymon Majewski, Alain Durmus, and Fabian-Robert Stöter. Sliced-Wasserstein flows: Nonparametric generative modeling via optimal transport and diffusions. In International Conference on Machine Learning, pp. 4104–4113, 2019. ",
1320
+ "bbox": [
1321
+ 176,
1322
+ 829,
1323
+ 825,
1324
+ 873
1325
+ ],
1326
+ "page_idx": 10
1327
+ },
1328
+ {
1329
+ "type": "text",
1330
+ "text": "Cheng Meng, Yuan Ke, Jingyi Zhang, Mengrui Zhang, Wenxuan Zhong, and Ping Ma. Large-scale optimal transport map estimation using projection pursuit. In Advances in Neural Information Processing Systems, pp. 8118–8129, 2019. ",
1331
+ "bbox": [
1332
+ 174,
1333
+ 881,
1334
+ 825,
1335
+ 924
1336
+ ],
1337
+ "page_idx": 10
1338
+ },
1339
+ {
1340
+ "type": "text",
1341
+ "text": "Boris Muzellec and Marco Cuturi. Subspace detours: Building transport plans that are optimal on subspace projections. In Advances in Neural Information Processing Systems, pp. 6917–6928, 2019. ",
1342
+ "bbox": [
1343
+ 176,
1344
+ 103,
1345
+ 825,
1346
+ 146
1347
+ ],
1348
+ "page_idx": 11
1349
+ },
1350
+ {
1351
+ "type": "text",
1352
+ "text": "Kimia Nadjahi, Alain Durmus, Umut Simsekli, and Roland Badeau. Asymptotic guarantees for learning generative models with the sliced-Wasserstein distance. In Advances in Neural Information Processing Systems, pp. 250–260, 2019. ",
1353
+ "bbox": [
1354
+ 174,
1355
+ 155,
1356
+ 823,
1357
+ 198
1358
+ ],
1359
+ "page_idx": 11
1360
+ },
1361
+ {
1362
+ "type": "text",
1363
+ "text": "François-Pierre Paty and Marco Cuturi. Subspace robust Wasserstein distances. In International Conference on Machine Learning, pp. 5072–5081, 2019. ",
1364
+ "bbox": [
1365
+ 173,
1366
+ 205,
1367
+ 823,
1368
+ 236
1369
+ ],
1370
+ "page_idx": 11
1371
+ },
1372
+ {
1373
+ "type": "text",
1374
+ "text": "Michaël Perrot, Nicolas Courty, Rémi Flamary, and Amaury Habrard. Mapping estimation for discrete optimal transport. Advances in Neural Information Processing Systems, 29:4197–4205, 2016. ",
1375
+ "bbox": [
1376
+ 173,
1377
+ 243,
1378
+ 825,
1379
+ 286
1380
+ ],
1381
+ "page_idx": 11
1382
+ },
1383
+ {
1384
+ "type": "text",
1385
+ "text": "Gabriel Peyré and Marco Cuturi. Computational optimal transport. Foundations and Trends® in Machine Learning, 11(5-6):355–607, 2019. ",
1386
+ "bbox": [
1387
+ 174,
1388
+ 295,
1389
+ 821,
1390
+ 325
1391
+ ],
1392
+ "page_idx": 11
1393
+ },
1394
+ {
1395
+ "type": "text",
1396
+ "text": "Julien Rabin, Julie Delon, and Yann Gousseau. Regularization of transportation maps for color and contrast transfer. In 2010 IEEE International Conference on Image Processing, pp. 1933–1936. IEEE, 2010. ",
1397
+ "bbox": [
1398
+ 173,
1399
+ 333,
1400
+ 826,
1401
+ 376
1402
+ ],
1403
+ "page_idx": 11
1404
+ },
1405
+ {
1406
+ "type": "text",
1407
+ "text": "Julien Rabin, Sira Ferradans, and Nicolas Papadakis. Adaptive color transfer with relaxed optimal transport. In 2014 IEEE International Conference on Image Processing (ICIP), pp. 4852–4856. IEEE, 2014. ",
1408
+ "bbox": [
1409
+ 173,
1410
+ 385,
1411
+ 825,
1412
+ 428
1413
+ ],
1414
+ "page_idx": 11
1415
+ },
1416
+ {
1417
+ "type": "text",
1418
+ "text": "Erik Reinhard, Michael Adhikhmin, Bruce Gooch, and Peter Shirley. Color transfer between images. IEEE Computer graphics and applications, 21(5):34–41, 2001. ",
1419
+ "bbox": [
1420
+ 173,
1421
+ 435,
1422
+ 825,
1423
+ 467
1424
+ ],
1425
+ "page_idx": 11
1426
+ },
1427
+ {
1428
+ "type": "text",
1429
+ "text": "Mark Rowland, Jiri Hron, Yunhao Tang, Krzysztof Choromanski, Tamas Sarlos, and Adrian Weller. Orthogonal estimation of Wasserstein distances. In The 22nd International Conference on Artificial Intelligence and Statistics, pp. 186–195, 2019. ",
1430
+ "bbox": [
1431
+ 174,
1432
+ 473,
1433
+ 823,
1434
+ 517
1435
+ ],
1436
+ "page_idx": 11
1437
+ },
1438
+ {
1439
+ "type": "text",
1440
+ "text": "Bernhard Schmitzer. Stabilized sparse scaling algorithms for entropy regularized transport problems. SIAM Journal on Scientific Computing, 41(3):A1443–A1481, 2019. ",
1441
+ "bbox": [
1442
+ 169,
1443
+ 526,
1444
+ 825,
1445
+ 555
1446
+ ],
1447
+ "page_idx": 11
1448
+ },
1449
+ {
1450
+ "type": "text",
1451
+ "text": "Richard Sinkhorn. Diagonal equivalence to matrices with prescribed row and column sums. The American Mathematical Monthly, 74(4):402–405, 1967. ",
1452
+ "bbox": [
1453
+ 171,
1454
+ 564,
1455
+ 825,
1456
+ 593
1457
+ ],
1458
+ "page_idx": 11
1459
+ },
1460
+ {
1461
+ "type": "text",
1462
+ "text": "Ilya Tolstikhin, Olivier Bousquet, Sylvain Gelly, and Bernhard Schoelkopf. Wasserstein auto-encoders. In International Conference on Learning Representations, 2018. ",
1463
+ "bbox": [
1464
+ 173,
1465
+ 602,
1466
+ 825,
1467
+ 631
1468
+ ],
1469
+ "page_idx": 11
1470
+ },
1471
+ {
1472
+ "type": "text",
1473
+ "text": "Cédric Villani. Optimal transport: Old and New. Springer, 2008. ",
1474
+ "bbox": [
1475
+ 173,
1476
+ 638,
1477
+ 602,
1478
+ 655
1479
+ ],
1480
+ "page_idx": 11
1481
+ },
1482
+ {
1483
+ "type": "text",
1484
+ "text": "Martin J Wainwright. High-dimensional statistics: A non-asymptotic viewpoint. Cambridge University Press, 2019. ",
1485
+ "bbox": [
1486
+ 173,
1487
+ 662,
1488
+ 823,
1489
+ 693
1490
+ ],
1491
+ "page_idx": 11
1492
+ },
1493
+ {
1494
+ "type": "text",
1495
+ "text": "Jiqing Wu, Zhiwu Huang, Dinesh Acharya, Wen Li, Janine Thoma, Danda Pani Paudel, and Luc Van Gool. Sliced Wasserstein generative models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3713–3722, 2019. ",
1496
+ "bbox": [
1497
+ 176,
1498
+ 700,
1499
+ 823,
1500
+ 744
1501
+ ],
1502
+ "page_idx": 11
1503
+ },
1504
+ {
1505
+ "type": "text",
1506
+ "text": "Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianxiong Xiao. Lsun: Construction of a largescale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. ",
1507
+ "bbox": [
1508
+ 176,
1509
+ 752,
1510
+ 825,
1511
+ 795
1512
+ ],
1513
+ "page_idx": 11
1514
+ }
1515
+ ]
parse/train/QYjO70ACDK/QYjO70ACDK_middle.json ADDED
The diff for this file is too large to render. See raw diff