ZHANGYUXUAN-zR commited on
Commit
5069737
·
verified ·
1 Parent(s): 35a0194

Add files using upload-large-folder tool

Browse files
md/train/B1gJ1L2aW/B1gJ1L2aW.md CHANGED
@@ -110,7 +110,7 @@ $X$ : a dataset of normal examples $H ( x )$ : a pre-trained DNN with $L$ transf
110
 
111
  # Output:
112
 
113
- Detector(LID) . a detector 1: $\mathrm { L I D } _ { n e g } { = } [ ]$ , $\mathrm { L I D } _ { p o s } { = } [ ]$ 2: for $B _ { n o r m }$ in $X$ do $\textsf { \textsf { D } } B _ { n o r m }$ : a minibatch of normal examples 3: $B _ { a d v } : =$ adversarial attack $B _ { n o r m }$ . $B _ { a d v }$ : a minibatch of adversarial examples 4: $B _ { n o i s y }$ : $: =$ add random noise to $B _ { n o r m }$ . $B _ { n o i s y }$ : a minibatch of noisy examples 5: N = |Bnorm| $\triangleright$ number of examples in $B _ { n o r m }$ 6: LIDnorm, LIDnoisy, $\mathrm { L I D } _ { n o i s y } = \mathrm { z e r o s } [ N , L ]$ 7: for $i$ in $[ 1 , L ]$ do 8: Anorm = Hi(Bnorm) $\triangleright i$ -th layer activations of $B _ { n o r m }$ 9: Aadv = Hi(Badv) $\triangleright i$ -th layer activations of $B _ { a d v }$ 10: $A _ { n o i s y } = H ^ { i } ( B _ { n o i s y } )$ ${ \triangleright } i$ -th layer activations of $B _ { n o i s y }$ 11: for $j$ in $[ 1 , N ]$ do 12: $\begin{array} { r } { \dot { \mathrm { L I D } _ { n o r m } } \dot { [ j , i ] } = - \Big ( \frac { 1 } { k } \sum _ { i = 1 } ^ { k } \log { \frac { r _ { i } ( A _ { n o r m } [ j ] , A _ { n o r m } ) } { r _ { k } ( A _ { n o r m } [ j ] , A _ { n o r m } ) } } \Big ) ^ { - 1 } } \end{array}$ 13: $\begin{array} { r } { \mathbf { L I D } _ { a d v } [ j , i ] = - \Big ( \frac { 1 } { k } \sum _ { i = 1 } ^ { k } \log \frac { r _ { i } ( A _ { a d v } [ j ] , A _ { n o r m } ) } { r _ { k } ( A _ { a d v } [ j ] , A _ { n o r m } ) } \Big ) ^ { - \frac { 1 } { \gamma _ { k } } } } \end{array}$ 1 14: LIDnoisy[j, i] = − 1k Pki=1 log ri(Anoisy[j],Anorm)rk(Anoisy[j],Anorm)  15: $\triangleright r _ { i } ( A [ j ] , A _ { n o r m } )$ : the $L _ { 2 }$ distance of $A _ { - } [ j ]$ to its $i$ -th nearest neighbor in $A _ { n o r m }$ 16: end for 17: end for 18: ${ \mathrm { L I D } } _ { n e g }$ .append $( \mathrm { L I D } _ { n o r m } )$ , ${ \mathrm { L I D } } _ { n e g }$ .append $( \mathrm { L I D } _ { n o i s y } )$ 19: $\mathrm { L I D } _ { p o s }$ .append $\mathrm { L I D } _ { a d v , }$ ) 20: end for 21: Detector $\left( \mathrm { L I D } \right) =$ train a classifier on $( \mathrm { L I D } _ { n e g } , \mathrm { L I D } _ { p o s } )$
114
 
115
  # 5.1 EXPERIMENTAL SETUP
116
 
 
110
 
111
  # Output:
112
 
113
+ Detector(LID) . a detector 1: $\mathrm { L I D } _ { n e g } { = } [ ]$ , $\mathrm { L I D } _ { p o s } { = } [ ]$ 2: for $B _ { n o r m }$ in $X$ do $\textsf { \textsf { D } } B _ { n o r m }$ : a minibatch of normal examples 3: $B _ { a d v } : =$ adversarial attack $B _ { n o r m }$ . $B _ { a d v }$ : a minibatch of adversarial examples 4: $B _ { n o i s y }$ : $: =$ add random noise to $B _ { n o r m }$ . $B _ { n o i s y }$ : a minibatch of noisy examples 5: N = |Bnorm| $\triangleright$ number of examples in $B _ { n o r m }$ 6: LIDnorm, LIDnoisy, $\mathrm { L I D } _ { n o i s y } = \mathrm { z e r o s } [ N , L ]$ 7: for $i$ in $[ 1 , L ]$ do 8: Anorm = Hi(Bnorm) $\triangleright i$ -th layer activations of $B _ { n o r m }$ 9: Aadv = Hi(Badv) $\triangleright i$ -th layer activations of $B _ { a d v }$ 10: $A _ { n o i s y } = H ^ { i } ( B _ { n o i s y } )$ ${ \triangleright } i$ -th layer activations of $B _ { n o i s y }$ 11: for $j$ in $[ 1 , N ]$ do 12: $\begin{array} { r } { \dot { \mathrm { L I D } _ { n o r m } } \dot { [ j , i ] } = - \Big ( \frac { 1 } { k } \sum _ { i = 1 } ^ { k } \log { \frac { r _ { i } ( A _ { n o r m } [ j ] , A _ { n o r m } ) } { r _ { k } ( A _ { n o r m } [ j ] , A _ { n o r m } ) } } \Big ) ^ { - 1 } } \end{array}$ 13: $\begin{array} { r } { \mathbf { L I D } _ { a d v } [ j , i ] = - \Big ( \frac { 1 } { k } \sum _ { i = 1 } ^ { k } \log \frac { r _ { i } ( A _ { a d v } [ j ] , A _ { n o r m } ) } { r _ { k } ( A _ { a d v } [ j ] , A _ { n o r m } ) } \Big ) ^ { - \frac { 1 } { \gamma _ { k } } } } \end{array}$ 1 14: LIDnoisy[j, i] = − 1k Pki=1 log ri(Anoisy[j],Anorm)rk(Anoisy[j],Anorm)  15: $\triangleright r _ { i } ( A [ j ] , A _ { n o r m } )$ : the $L _ { 2 }$ distance of $A _ { - } [ j ]$ to its $i$ -th nearest neighbor in $A _ { n o r m }$ 16: end for 17: end for 18: ${ \mathrm { L I D } } _ { n e g }$ .append $( \mathrm { L I D } _ { n o r m } )$ , ${ \mathrm { L I D } } _ { n e g }$ .append $( \mathrm { L I D } _ { n o i s y } )$ 19: $\mathrm { L I D } _ { p o s }$ .append $\mathrm { L I D } _ { a d v , }$ ) 20: end for 21: Detector $\left( \mathrm { L I D } \right) =$ train a classifier on $( \mathrm { L I D } _ { n e g } , \mathrm { L I D } _ { p o s } )$
114
 
115
  # 5.1 EXPERIMENTAL SETUP
116
 
md/train/HJxiMAVtPH/HJxiMAVtPH.md CHANGED
@@ -83,7 +83,7 @@ Interpretation: Assuming $\mathcal { G }$ is ergodic: $\begin{array} { r } { p (
83
 
84
  # 4.1 MULTI-SCALE CASE (MUSAE)
85
 
86
- We know that the SGNS aspect of MUSAE (Algorithm 2, Line 17) is minimized when the learned embeddings grv , hrf satisfy grw>hrf ≈ log #(w,f )r|Dr|#(w)r#(f )r  − log b ∀w ∈ V, f ∈ F. Our aim is to express this factorization in terms of known properties of the graph $\mathcal { G }$ and its features.
87
 
88
  Lemma 1. The empirical statistics of node-feature pairs obtained from random walks give unbiased estimates of joint probabilities of observing feature $f \in \mathbb { F } r$ steps $( i )$ after; or $( i i )$ before node $v \in \mathbb { V }$ , as given by:
89
 
 
83
 
84
  # 4.1 MULTI-SCALE CASE (MUSAE)
85
 
86
+ We know that the SGNS aspect of MUSAE (Algorithm 2, Line 17) is minimized when the learned embeddings grv , hrf satisfy grw>hrf ≈ log #(w,f )r|Dr|#(w)r#(f )r  − log b ∀w ∈ V, f ∈ F. Our aim is to express this factorization in terms of known properties of the graph $\mathcal { G }$ and its features.
87
 
88
  Lemma 1. The empirical statistics of node-feature pairs obtained from random walks give unbiased estimates of joint probabilities of observing feature $f \in \mathbb { F } r$ steps $( i )$ after; or $( i i )$ before node $v \in \mathbb { V }$ , as given by:
89
 
md/train/HkCsm6lRb/HkCsm6lRb.md CHANGED
@@ -357,7 +357,7 @@ As explained in the main paper, we fit the joint graphical model $p ( x , y , z
357
 
358
  • Image and Label encoder, $q ( z | x , y )$ : Our architecture (Figure 9) for the image-label encoder first separately processes the images and the labels, and then concatenates them downstream in the network and then passes the concatenated features through a multi-layered perceptron. More specifically, we have convolutional layers which process image into 32, 64, 128, 16 feature maps with strides $1 , 2 , 2 , 2$ in the corresponding layers. We use batch normalization in the convolutional layers before applying the ELU non-linearity. On the label encoder side, we first encode the each attribute label into a 32d continuous vector and then pass each individual attribute vector through a 2-layered MLP with 512 hidden dimensions each. For example, for MNIST-A we have 4 attributes, which gives us 4 vectors of 512d. We then concatenate these vectors and pass it through a two layer MLP. Finally we concatenate this label feature with the image feature after the convolutional layers (after flattening the conv-features) and then pass the result through a 2 layer MLP to predict the mean $( \mu )$ and standard deviation $( \sigma )$ for the latent space gaussian. Following standard practice, we predict $\log \sigma$ for the standard deviation in order to get values which are positive.
359
 
360
- # µ
361
 
362
  ![](images/3c8666b4b928e365d467fb0e5f9ee430477aaa538630478bba8cd6275380d78a.jpg)
363
  Figure 9: Architecture for the $q ( z | x , y )$ network in our JVAE models for MNIST-A. Images are ( $6 4 \times 6 4 \times 1 )$ , class has 10 possible values, scale has 2 possible values, orientation has 3 possible values, and location has 4 possible values.
 
357
 
358
  • Image and Label encoder, $q ( z | x , y )$ : Our architecture (Figure 9) for the image-label encoder first separately processes the images and the labels, and then concatenates them downstream in the network and then passes the concatenated features through a multi-layered perceptron. More specifically, we have convolutional layers which process image into 32, 64, 128, 16 feature maps with strides $1 , 2 , 2 , 2$ in the corresponding layers. We use batch normalization in the convolutional layers before applying the ELU non-linearity. On the label encoder side, we first encode the each attribute label into a 32d continuous vector and then pass each individual attribute vector through a 2-layered MLP with 512 hidden dimensions each. For example, for MNIST-A we have 4 attributes, which gives us 4 vectors of 512d. We then concatenate these vectors and pass it through a two layer MLP. Finally we concatenate this label feature with the image feature after the convolutional layers (after flattening the conv-features) and then pass the result through a 2 layer MLP to predict the mean $( \mu )$ and standard deviation $( \sigma )$ for the latent space gaussian. Following standard practice, we predict $\log \sigma$ for the standard deviation in order to get values which are positive.
359
 
360
+ # µ
361
 
362
  ![](images/3c8666b4b928e365d467fb0e5f9ee430477aaa538630478bba8cd6275380d78a.jpg)
363
  Figure 9: Architecture for the $q ( z | x , y )$ network in our JVAE models for MNIST-A. Images are ( $6 4 \times 6 4 \times 1 )$ , class has 10 possible values, scale has 2 possible values, orientation has 3 possible values, and location has 4 possible values.
md/train/SPrVNsXnGd/SPrVNsXnGd.md CHANGED
@@ -163,7 +163,7 @@ where $\epsilon \left( \lambda \right)$ is the RDP of the subsampled shuffle mec
163
 
164
  The proof outline of Theorem $3$ is as follows: Note that $\mathcal { A } _ { \mathrm { c l d p } }$ is an iterative algorithm, where in each iteration we use the subsampled shuffle mechanism as defined in $\textcircled{3}$ , for which we have computed the RDP guarantees in Theorem $\bigtriangledown$ Now, for the privacy analysis of $\mathcal { A } _ { \mathrm { c l d p } }$ , we use the adaptive composition theorem from $\textcircled { 1 3 7 }$ Proposition 1] and then use the RDP to DP conversion given in Lemma $^ { 1 . }$ For the convergence analysis, we use a standard non-private SGD convergence result and compute the required parameters for that. See Appendix F for a complete proof of Theorem 3.
165
 
166
- Remark 1. Note that our convergence bound is affected by the variance of the $\epsilon _ { \mathrm { 0 } }$ -LDP mechanism $\mathcal { R } _ { p }$ . For example, when $f$ is $L$ -Lipschitz continuous w.r.t. the $\ell _ { 2 }$ -norm, we can use the LDP mechanism $\mathcal { R } _ { 2 }$ proposed in $\boxed { 1 1 }$ that has variance $\begin{array} { r } { G _ { 2 } ^ { 2 } ( L ) = 1 4 L ^ { 2 } d \big ( \frac { e ^ { \epsilon _ { 0 } } + 1 } { e ^ { \epsilon _ { 0 } } - 1 } \big ) ^ { 2 } } \end{array}$ ; and when $f$ is $L$ -Lipschitz continuous w.r.t. the $\ell _ { 1 }$ -norm or $\ell _ { \infty }$ -norm, we can use the LDP mechanisms $\mathcal { R } _ { \infty }$ or $\mathcal { R } _ { 1 }$ , respectively, proposed in $\lVert 2 7 \rVert$ 1 that have variances $\begin{array} { r } { G _ { \infty } ^ { 2 } ( L ) = L ^ { 2 } d ^ { 2 } \big ( \frac { e ^ { \epsilon _ { 0 } } + 1 } { e ^ { \epsilon _ { 0 } } - 1 } \big ) ^ { 2 } } \end{array}$ 2 and G21(L) = L2d e✏0 +1e✏0 1 2 , respectively. By plugging these variances $G _ { p } ^ { 2 } ( L )$ (for $p = 1 , 2 , \infty )$ into Theorem $3 .$ we get the convergence rate of the $L$ -Lipschitz continuous loss function w.r.t. the $\ell _ { p }$ -norm (for $p \equiv \infty , 2 , 1 $ ).
167
 
168
  Remark 2. The privacy parameter in $( 5 )$ is not in a closed form expression and could be obtained by solving an optimization problem. However, we numerically compute it for several interesting regimes of parameters in our numerical experiments; see Section 4 for more details.
169
 
@@ -226,7 +226,7 @@ $$
226
  \begin{array} { r } { \mathcal { D } _ { \mathrm { s a m e } } ^ { m } = \{ ( \mathcal { D } _ { m } , \mathcal { D } _ { m } ^ { \prime } , \mathcal { D } _ { m } ^ { \prime \prime } ) : \mathcal { D } _ { m } = ( d , \ldots , d , d ) \in \mathcal { X } ^ { m } , \mathcal { D } _ { m } ^ { \prime } = ( d , \ldots , d , d ^ { \prime } ) \in \mathcal { X } ^ { m } , \mathrm { ~ a ~ n ~ d ~ } \mathcal { X } = 1 , \ldots , d , 1 \} , } \\ { \mathcal { D } _ { m } ^ { \prime \prime } = ( d , \ldots , d , d ^ { \prime \prime } ) \in \mathcal { X } ^ { m } , \mathrm { ~ w h e r e ~ } d , d ^ { \prime } , d ^ { \prime \prime } \in \mathcal { X } \} } \end{array}
227
  $$
228
 
229
- Consider arbitrary neighboring datasets $\mathcal { D } = \left( d _ { 1 } , \ldots , d _ { k - 1 } , d _ { k } \right)$ , $\mathcal { D } ^ { \prime } = ( d _ { 1 } , \ldots , d _ { k - 1 } , d _ { k } ^ { \prime } )$ , and $\mathcal { D } ^ { \prime \prime } = ( d _ { 1 } , \dots , d _ { k - 1 } ^ { - } , d _ { k } ^ { \overline { { \prime } } \prime } )$ , each having $k$ elements. For any $m \in \{ 0 , \ldots , k - 1 \}$ , we define new neighboring dataeach having = (d00k , . . . , d00k , dk), D0(k)m+1 = (d00k , . . . , d00k , d0k), and D00(k)m+1 $\mathcal { D } _ { m + 1 } ^ { \prime \prime ( k ) } = ( d _ { k } ^ { \prime \prime } , \dots , d _ { k } ^ { \prime \prime } )$ $m + 1$ $( \mathcal { D } _ { m + 1 } ^ { \prime \prime ( k ) } , \mathcal { D } _ { m + 1 } ^ { \prime ( k ) } , \mathcal { D } _ { m + 1 } ^ { ( k ) } ) \in \mathcal { D } _ { \mathrm { s a m e } } ^ { m }$
230
 
231
  The first step of the proof is given in the following theorem.
232
 
 
163
 
164
  The proof outline of Theorem $3$ is as follows: Note that $\mathcal { A } _ { \mathrm { c l d p } }$ is an iterative algorithm, where in each iteration we use the subsampled shuffle mechanism as defined in $\textcircled{3}$ , for which we have computed the RDP guarantees in Theorem $\bigtriangledown$ Now, for the privacy analysis of $\mathcal { A } _ { \mathrm { c l d p } }$ , we use the adaptive composition theorem from $\textcircled { 1 3 7 }$ Proposition 1] and then use the RDP to DP conversion given in Lemma $^ { 1 . }$ For the convergence analysis, we use a standard non-private SGD convergence result and compute the required parameters for that. See Appendix F for a complete proof of Theorem 3.
165
 
166
+ Remark 1. Note that our convergence bound is affected by the variance of the $\epsilon _ { \mathrm { 0 } }$ -LDP mechanism $\mathcal { R } _ { p }$ . For example, when $f$ is $L$ -Lipschitz continuous w.r.t. the $\ell _ { 2 }$ -norm, we can use the LDP mechanism $\mathcal { R } _ { 2 }$ proposed in $\boxed { 1 1 }$ that has variance $\begin{array} { r } { G _ { 2 } ^ { 2 } ( L ) = 1 4 L ^ { 2 } d \big ( \frac { e ^ { \epsilon _ { 0 } } + 1 } { e ^ { \epsilon _ { 0 } } - 1 } \big ) ^ { 2 } } \end{array}$ ; and when $f$ is $L$ -Lipschitz continuous w.r.t. the $\ell _ { 1 }$ -norm or $\ell _ { \infty }$ -norm, we can use the LDP mechanisms $\mathcal { R } _ { \infty }$ or $\mathcal { R } _ { 1 }$ , respectively, proposed in $\lVert 2 7 \rVert$ 1 that have variances $\begin{array} { r } { G _ { \infty } ^ { 2 } ( L ) = L ^ { 2 } d ^ { 2 } \big ( \frac { e ^ { \epsilon _ { 0 } } + 1 } { e ^ { \epsilon _ { 0 } } - 1 } \big ) ^ { 2 } } \end{array}$ 2 and G21(L) = L2d e✏0 +1e✏0 1 2 , respectively. By plugging these variances $G _ { p } ^ { 2 } ( L )$ (for $p = 1 , 2 , \infty )$ into Theorem $3 .$ we get the convergence rate of the $L$ -Lipschitz continuous loss function w.r.t. the $\ell _ { p }$ -norm (for $p \equiv \infty , 2 , 1 $ ).
167
 
168
  Remark 2. The privacy parameter in $( 5 )$ is not in a closed form expression and could be obtained by solving an optimization problem. However, we numerically compute it for several interesting regimes of parameters in our numerical experiments; see Section 4 for more details.
169
 
 
226
  \begin{array} { r } { \mathcal { D } _ { \mathrm { s a m e } } ^ { m } = \{ ( \mathcal { D } _ { m } , \mathcal { D } _ { m } ^ { \prime } , \mathcal { D } _ { m } ^ { \prime \prime } ) : \mathcal { D } _ { m } = ( d , \ldots , d , d ) \in \mathcal { X } ^ { m } , \mathcal { D } _ { m } ^ { \prime } = ( d , \ldots , d , d ^ { \prime } ) \in \mathcal { X } ^ { m } , \mathrm { ~ a ~ n ~ d ~ } \mathcal { X } = 1 , \ldots , d , 1 \} , } \\ { \mathcal { D } _ { m } ^ { \prime \prime } = ( d , \ldots , d , d ^ { \prime \prime } ) \in \mathcal { X } ^ { m } , \mathrm { ~ w h e r e ~ } d , d ^ { \prime } , d ^ { \prime \prime } \in \mathcal { X } \} } \end{array}
227
  $$
228
 
229
+ Consider arbitrary neighboring datasets $\mathcal { D } = \left( d _ { 1 } , \ldots , d _ { k - 1 } , d _ { k } \right)$ , $\mathcal { D } ^ { \prime } = ( d _ { 1 } , \ldots , d _ { k - 1 } , d _ { k } ^ { \prime } )$ , and $\mathcal { D } ^ { \prime \prime } = ( d _ { 1 } , \dots , d _ { k - 1 } ^ { - } , d _ { k } ^ { \overline { { \prime } } \prime } )$ , each having $k$ elements. For any $m \in \{ 0 , \ldots , k - 1 \}$ , we define new neighboring dataeach having = (d00k , . . . , d00k , dk), D0(k)m+1 = (d00k , . . . , d00k , d0k), and D00(k)m+1 $\mathcal { D } _ { m + 1 } ^ { \prime \prime ( k ) } = ( d _ { k } ^ { \prime \prime } , \dots , d _ { k } ^ { \prime \prime } )$ $m + 1$ $( \mathcal { D } _ { m + 1 } ^ { \prime \prime ( k ) } , \mathcal { D } _ { m + 1 } ^ { \prime ( k ) } , \mathcal { D } _ { m + 1 } ^ { ( k ) } ) \in \mathcal { D } _ { \mathrm { s a m e } } ^ { m }$
230
 
231
  The first step of the proof is given in the following theorem.
232
 
md/train/ZKbZ4mebI9l/ZKbZ4mebI9l.md CHANGED
@@ -123,7 +123,7 @@ $$
123
  \widehat { \mathcal { R } } _ { k } ( \mathcal { D } ^ { ( \mathrm { t r } ) } ; \mathbf { W } _ { 1 } , \mathbf { W } _ { 2 } , \mathbf { W } _ { 3 } ) = \frac { 1 } { N ^ { \mathrm { t r } } } \sum _ { i = 1 } ^ { N ^ { \mathrm { t r } } } l \big ( r _ { \mathbf { W } } ^ { ( k , \mathrm { G N N } ) } ( G _ { i } ) , y _ { i } \big ) .
124
  $$
125
 
126
- Equation (1) is impractical for all but the smallest graphs, since $r _ { \mathbf { W } } ^ { ( k , \mathrm { G N N } ) }$ s a sum overusing a samall k-vertex induced subgraphs S(k) of G. Hence, we approximate r(k,GW $r _ { \mathbf { W } } ^ { ( k , \mathrm { G N N } ) }$ S (k)B ${ \mathcal S } _ { B } ^ { ( k ) } ~ \subset ~ { \mathcal S } ^ { ( k ) }$ at every gradient step, i.e.. Due to non-linearities in $\widehat { r } _ { \mathbf { W } } ^ { ( k , \mathrm { G N N } ) } ( G ) \ =$ $\begin{array} { r } { \rho _ { \mathbf { W } _ { 1 } } \Big ( | S ^ { ( k ) } | / | S _ { B } ^ { ( k ) } | \sum _ { S \in S _ { B } ^ { ( k ) } } \phi _ { \mathbf { W } _ { 2 } } \big ( h _ { \mathbf { W } _ { 3 } } ^ { \mathrm { ( G N N ) } } ( G [ S ] ) \big ) \Big ) } \end{array}$ $\rho _ { \mathbf { W } _ { 1 } }$ $l$ ging $\widehat { r } _ { \mathbf { W } } ^ { ( k , \mathrm { G N N } ) }$ B into Equatio (1) does not provide us with an unbiased estimate of $\widehat { \mathcal { R } } _ { k }$ . However, if i. $l ( \rho _ { \mathbf { W } _ { 1 } } ( a ) , y )$ $a$ per upper bound of our loss,. In practice, many models e., 1/N tr PN tri=1 l $\begin{array} { r } { 1 / N ^ { \mathrm { t r } } \sum _ { i = 1 } ^ { N ^ { \mathrm { t r } } } l \big ( r _ { \mathbf { W } } ^ { ( k , \mathrm { G N N } ) } ( G _ { i } ) , y _ { i } \big ) \leq 1 / N ^ { \mathrm { t r } } \sum _ { i = 1 } ^ { N ^ { \mathrm { t r } } } l \big ( \widehat { r } _ { \mathbf { W } } ^ { ( k , \mathrm { G N N } ) } ( G _ { i } ) , y _ { i } \big ) . } \end{array}$ tr1 lr (k,GNN)W (Gi), yi brely on this approximation and provide scalable and reliable training procedures, cf. [48, 79, 80, 112].
127
 
128
  # 4.1 Expressive power
129
 
 
123
  \widehat { \mathcal { R } } _ { k } ( \mathcal { D } ^ { ( \mathrm { t r } ) } ; \mathbf { W } _ { 1 } , \mathbf { W } _ { 2 } , \mathbf { W } _ { 3 } ) = \frac { 1 } { N ^ { \mathrm { t r } } } \sum _ { i = 1 } ^ { N ^ { \mathrm { t r } } } l \big ( r _ { \mathbf { W } } ^ { ( k , \mathrm { G N N } ) } ( G _ { i } ) , y _ { i } \big ) .
124
  $$
125
 
126
+ Equation (1) is impractical for all but the smallest graphs, since $r _ { \mathbf { W } } ^ { ( k , \mathrm { G N N } ) }$ s a sum overusing a samall k-vertex induced subgraphs S(k) of G. Hence, we approximate r(k,GW $r _ { \mathbf { W } } ^ { ( k , \mathrm { G N N } ) }$ S (k)B ${ \mathcal S } _ { B } ^ { ( k ) } ~ \subset ~ { \mathcal S } ^ { ( k ) }$ at every gradient step, i.e.. Due to non-linearities in $\widehat { r } _ { \mathbf { W } } ^ { ( k , \mathrm { G N N } ) } ( G ) \ =$ $\begin{array} { r } { \rho _ { \mathbf { W } _ { 1 } } \Big ( | S ^ { ( k ) } | / | S _ { B } ^ { ( k ) } | \sum _ { S \in S _ { B } ^ { ( k ) } } \phi _ { \mathbf { W } _ { 2 } } \big ( h _ { \mathbf { W } _ { 3 } } ^ { \mathrm { ( G N N ) } } ( G [ S ] ) \big ) \Big ) } \end{array}$ $\rho _ { \mathbf { W } _ { 1 } }$ $l$ ging $\widehat { r } _ { \mathbf { W } } ^ { ( k , \mathrm { G N N } ) }$ B into Equatio (1) does not provide us with an unbiased estimate of $\widehat { \mathcal { R } } _ { k }$ . However, if i. $l ( \rho _ { \mathbf { W } _ { 1 } } ( a ) , y )$ $a$ per upper bound of our loss,. In practice, many models e., 1/N tr PN tri=1 l $\begin{array} { r } { 1 / N ^ { \mathrm { t r } } \sum _ { i = 1 } ^ { N ^ { \mathrm { t r } } } l \big ( r _ { \mathbf { W } } ^ { ( k , \mathrm { G N N } ) } ( G _ { i } ) , y _ { i } \big ) \leq 1 / N ^ { \mathrm { t r } } \sum _ { i = 1 } ^ { N ^ { \mathrm { t r } } } l \big ( \widehat { r } _ { \mathbf { W } } ^ { ( k , \mathrm { G N N } ) } ( G _ { i } ) , y _ { i } \big ) . } \end{array}$ tr1 lr (k,GNN)W (Gi), yi brely on this approximation and provide scalable and reliable training procedures, cf. [48, 79, 80, 112].
127
 
128
  # 4.1 Expressive power
129
 
md/train/aDjoksTpXOP/aDjoksTpXOP.md CHANGED
@@ -88,7 +88,7 @@ where $\kappa _ { 0 }$ and $\kappa _ { 1 }$ are given in (4).
88
 
89
  In this section, we recall approximation properties of dot-product kernels on the sphere, through spectral decompositions of integral operators in the basis of spherical harmonics. Further background is provided in Appendix A.
90
 
91
- Spherical harmonics and description of the RKHS. A standard approach to study the RKHS of a kernel is through the spectral decomposition of an integral operator $T$ given by $\begin{array} { r } { T f ( x ) = \int k ( x , y ) f ( y ) d \tau ( y ) } \end{array}$ for some measure $\tau$ , leading to Mercer’s theorem (e.g., Cucker & Smale, 2002). When inputs lie on the sphere $\mathbb { S } ^ { d - 1 }$ in $d$ dimensions, dot-product kernels of the form $k ( x , x ^ { \prime } ) = \kappa ( x ^ { \prime } x ^ { \prime } )$ are rotationally-invariant, depending only on the angle between $x$ and $x ^ { \prime }$ . Similarly to how translation-invariant kernels are diagonalized in the Fourier basis, rotation-invariant kernels are diagonalized in the basis of spherical harmonics (Smola et al., 2001; Bach, 2017a), which lead to connections between eigenvalue decays and regularity as in the Fourier setting. In particular, if $\tau$ denotes the uniform measure on $\mathbb { S } ^ { d - 1 }$ , then $T Y _ { k , j } = \mu _ { k } Y _ { k , j }$ , where $Y _ { k , j }$ is the $j$ -th spherical harmonic polynomial of degree $k$ , where $k$ plays the role of a frequency as in the Fourier case, and the number of such orthogonal polynomials of degree $k$ is given by $\begin{array} { r } { N ( d , k ) = \frac { 2 k + d - 2 } { k } \binom { k + d - 3 } { d - 2 } } \end{array}$ 2 k+d−3d−2 , which grows as $k ^ { d - 2 }$ for large $k$ . The eigenvalues $\mu _ { k }$ only depend on the frequency $k$ and are given by
92
 
93
  $$
94
  \mu _ { k } = \frac { \omega _ { d - 2 } } { \omega _ { d - 1 } } \int _ { - 1 } ^ { 1 } \kappa ( t ) P _ { k } ( t ) ( 1 - t ^ { 2 } ) ^ { ( d - 3 ) / 2 } d t ,
@@ -317,7 +317,7 @@ Difan Zou, Yuan Cao, Dongruo Zhou, and Quanquan Gu. Stochastic gradient descent
317
 
318
  In this section, we provide some background on spherical harmonics needed for our study of approximation. See (Efthimiou & Frye, 2014; Atkinson & Han, 2012; Ismail, 2005) for references, as well as (Bach, 2017a, Appendix D). We consider inputs on the $d - 1$ sphere $\mathbb { S } ^ { d - 1 } = \{ x \in \mathbb { R } ^ { d } , \| x \| = 1 \}$ .
319
 
320
- We recall some properties of the spherical harmonics $Y _ { k , j }$ introduced in Section 2.2. For j = 1, . . . , N (d, k), where N (d, k) = 2k+d−2 k+d−3, the spherical harmonics $Y _ { k , j }$ are homogeneous harmonic polynomials of degree $k$ that are orthonormal with respect to the uniform distribution $\tau$ on the $d$ –1 sphere. The degree $k$ plays the role of an integer frequency, as in Fourier series, and the collection $\{ Y _ { k , j } , k \geq 0 , j = 1 , \ldots , N ( d , k ) \}$ forms an orthonormal basis of $L ^ { 2 } ( \mathbb { S } ^ { d - 1 } , d \tau )$ . As with Fourier series, there are tight connections between decay of coefficients in this basis w.r.t. $k$ , and regularity/differentiability of functions, in this case differentiability on the sphere. This follows from the fact that spherical harmonics are eigenfunctions of the Laplace-Beltrami operator on the sphere $\Delta _ { \mathbb { S } ^ { d - 1 } }$ (see Efthimiou & Frye, 2014, Proposition 4.5):
321
 
322
  $$
323
  \Delta _ { \mathbb { S } ^ { d - 1 } } Y _ { k , j } = - k ( k + d - 2 ) Y _ { k , j } .
 
88
 
89
  In this section, we recall approximation properties of dot-product kernels on the sphere, through spectral decompositions of integral operators in the basis of spherical harmonics. Further background is provided in Appendix A.
90
 
91
+ Spherical harmonics and description of the RKHS. A standard approach to study the RKHS of a kernel is through the spectral decomposition of an integral operator $T$ given by $\begin{array} { r } { T f ( x ) = \int k ( x , y ) f ( y ) d \tau ( y ) } \end{array}$ for some measure $\tau$ , leading to Mercer’s theorem (e.g., Cucker & Smale, 2002). When inputs lie on the sphere $\mathbb { S } ^ { d - 1 }$ in $d$ dimensions, dot-product kernels of the form $k ( x , x ^ { \prime } ) = \kappa ( x ^ { \prime } x ^ { \prime } )$ are rotationally-invariant, depending only on the angle between $x$ and $x ^ { \prime }$ . Similarly to how translation-invariant kernels are diagonalized in the Fourier basis, rotation-invariant kernels are diagonalized in the basis of spherical harmonics (Smola et al., 2001; Bach, 2017a), which lead to connections between eigenvalue decays and regularity as in the Fourier setting. In particular, if $\tau$ denotes the uniform measure on $\mathbb { S } ^ { d - 1 }$ , then $T Y _ { k , j } = \mu _ { k } Y _ { k , j }$ , where $Y _ { k , j }$ is the $j$ -th spherical harmonic polynomial of degree $k$ , where $k$ plays the role of a frequency as in the Fourier case, and the number of such orthogonal polynomials of degree $k$ is given by $\begin{array} { r } { N ( d , k ) = \frac { 2 k + d - 2 } { k } \binom { k + d - 3 } { d - 2 } } \end{array}$ 2 k+d−3d−2 , which grows as $k ^ { d - 2 }$ for large $k$ . The eigenvalues $\mu _ { k }$ only depend on the frequency $k$ and are given by
92
 
93
  $$
94
  \mu _ { k } = \frac { \omega _ { d - 2 } } { \omega _ { d - 1 } } \int _ { - 1 } ^ { 1 } \kappa ( t ) P _ { k } ( t ) ( 1 - t ^ { 2 } ) ^ { ( d - 3 ) / 2 } d t ,
 
317
 
318
  In this section, we provide some background on spherical harmonics needed for our study of approximation. See (Efthimiou & Frye, 2014; Atkinson & Han, 2012; Ismail, 2005) for references, as well as (Bach, 2017a, Appendix D). We consider inputs on the $d - 1$ sphere $\mathbb { S } ^ { d - 1 } = \{ x \in \mathbb { R } ^ { d } , \| x \| = 1 \}$ .
319
 
320
+ We recall some properties of the spherical harmonics $Y _ { k , j }$ introduced in Section 2.2. For j = 1, . . . , N (d, k), where N (d, k) = 2k+d−2 k+d−3, the spherical harmonics $Y _ { k , j }$ are homogeneous harmonic polynomials of degree $k$ that are orthonormal with respect to the uniform distribution $\tau$ on the $d$ –1 sphere. The degree $k$ plays the role of an integer frequency, as in Fourier series, and the collection $\{ Y _ { k , j } , k \geq 0 , j = 1 , \ldots , N ( d , k ) \}$ forms an orthonormal basis of $L ^ { 2 } ( \mathbb { S } ^ { d - 1 } , d \tau )$ . As with Fourier series, there are tight connections between decay of coefficients in this basis w.r.t. $k$ , and regularity/differentiability of functions, in this case differentiability on the sphere. This follows from the fact that spherical harmonics are eigenfunctions of the Laplace-Beltrami operator on the sphere $\Delta _ { \mathbb { S } ^ { d - 1 } }$ (see Efthimiou & Frye, 2014, Proposition 4.5):
321
 
322
  $$
323
  \Delta _ { \mathbb { S } ^ { d - 1 } } Y _ { k , j } = - k ( k + d - 2 ) Y _ { k , j } .