alejandrorg00/flynet-robustness

FlyNet Robustness

This repository contains the trained decoders and selected model checkpoints associated with the paper:

Garcia-Duran, A., Rodriguez-Garcia, A., Molano-Mazón, M., Hyafil, A., Ramaswamy, S., & Lappalainen, J. K. (2026). Measuring robustness and efficiency in a connectome-constrained fly visual system model on a collision-detection task. Advances in Neural Information Processing Systems (NeurIPS 2026)

The study investigates the robustness and computational efficiency of a connectome-constrained model of the Drosophila visual system (FlyNet) compared with artificial neural network architectures on a looming-based collision-detection task. The models are evaluated under adversarial attacks and common video perturbations to examine the contributions of network architecture, recurrence, and connectome-derived connectivity.

The model implementations, collision-detection pipeline, adversarial attacks, common video perturbations, and evaluation procedures are maintained in the accompanying source-code repository.

Code Repository: flynet-robustness.

The released checkpoints are intended to support reproduction of the evaluations reported in the paper.

Model architectures

The benchmark includes 14 models organized into four families:

Family	Models
CTRNNs	FlyNet, CTRNN, SCTRNN
CNNs	vCNN-2f, vCNN-5f, vCNN-10f, gCNN-2f
hexRAFTs	hexRAFT-1, hexRAFT-6, hexRAFT-12
RAFTs	RAFT-1, RAFT-6, RAFT-12, RAFTsm-12

To avoid duplicating pretrained weights available elsewhere, this repository provides the task-specific decoders, CNN checkpoints, and Sintel-trained hexRAFT model used in our benchmark.

Other pretrained models are available from their original repositories:

• FlyVis: Lappalainen et al. (2024) — GitHub.
• RAFT: Teed and Deng (2020) — GitHub.

The Sintel-trained hexRAFT model was trained using our RAFT-Sintel fork.

Repository structure

flynet-robustness/
├── decoders/
│   ├── flynet/
│   │   ├── hit.pkl
│   │   ├── position.pth
│   │   └── velocity.pth
│   ├── ctrnn/
│   │   └── hit.pkl
│   ├── sctrnn/
│   │   └── hit.pkl
│   ├── vcnn_2frames/
│   │   └── hit.pkl
│   ├── vcnn_5frames/
│   │   └── hit.pkl
│   ├── vcnn_10frames/
│   │   └── hit.pkl
│   ├── gcnn_2frames/
│   │   └── hit.pkl
│   ├── raft_small_12/
│   │   └── hit.pkl
│   ├── raft_1/
│   │   └── hit.pkl
│   ├── raft_6/
│   │   └── hit.pkl
│   ├── raft_12/
│   │   └── hit.pkl
│   ├── hexraft_1/
│   │   └── hit.pkl
│   ├── hexraft_6/
│   │   └── hit.pkl
│   ├── hexraft_12/
│   │   └── hit.pkl
│   ├── hexraft_1_sintel/
│   │   └── hit.pkl
│   ├── hexraft_6_sintel/
│   │   └── hit.pkl
│   └── hexraft_12_sintel/
│       └── hit.pkl
│
├── models/
│   ├── cnns/
│   │   ├── vcnn_2frames/
│   │   │   └── network_000.chkpt
│   │   ├── vcnn_5frames/
│   │   │   └── network_000
│   │   ├── vcnn_10frames/
│   │   │   └── network_000
│   │   └── gcnn_2frames/
│   │       └── network_000.chkpt
│   └── hexraft_sintel/
│       └── checkpoint.pth
│
└── README.md

Downloading checkpoints

Checkpoints can be downloaded individually or by model family using the Hugging Face Hub client.

pip install huggingface_hub

Download all decoders for a model:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="alejandrorg00/flynet-robustness",
    allow_patterns="decoders/flynet/*",
    local_dir=".",
)

Download all hit/miss decoders:

snapshot_download(
    repo_id="alejandrorg00/flynet-robustness",
    allow_patterns="decoders/*/hit.pkl",
    local_dir=".",
)

Download a particular decoder:

from huggingface_hub import hf_hub_download

checkpoint = hf_hub_download(
    repo_id="alejandrorg00/flynet-robustness",
    filename="decoders/flynet/hit.pkl",ç
    local_dir=".",
)

Download the Sintel-trained hexRAFT model:

checkpoint = hf_hub_download(
    repo_id="alejandrorg00/flynet-robustness",
    filename="models/hexraft_sintel/checkpoint.pth",
    local_dir=".",
)

The checkpoints require the corresponding model implementations and preprocessing pipelines.

Citation

If you use these models or decoders, please cite:

@inproceedings{garciaduran2026flyillusions,
  title     = {Measuring Robustness and Efficiency in a Connectome-Constrained Fly Visual System Model on a Collision-Detection Task},
  author    = {Garcia-Duran, Alexandre and Rodriguez-Garcia, Alejandro and
               Molano-Mazón, Manuel and Hyafil, Alexandre and
               Ramaswamy, Srikanth and Lappalainen, Janne K.},
  booktitle = {The Fortieth Annual Conference on Neural Information Processing Systems},
  year      = {2026},
  url       = {TBA}
}