Home

Awesome

SNN-Neural-Similarity-Movie

Official implementation of "Long-Range Feedback Spiking Network Captures Dynamic and Static Representations of the Visual Cortex under Movie Stimuli" (NeurIPS2024). See preprint.

By Liwei Huang, Zhengyu Ma, Liutao Yu, Huihui Zhou, Yonghong Tian.

We propose the long-range feedback spiking network (LoRaFB-SNet), which mimics top-down connections between cortical regions and incorporates spike information processing mechanisms inherent to biological neurons. We present Time-Series Representational Similarity Analysis (TSRSA) to measure the similarity between model representations and visual cortical representations of mice. Through evaluation, LoRaFB-SNet is highly competent in capturing both dynamic and static representations of the mouse visual cortex and contributes to the understanding of movie processing mechanisms of the visual system.

overview

Requirements

In order to run this project you will need:

pip install spikingjelly==0.0.0.0.14

Model Training

The code is stored in the file folder train. It supports single GPU or multiple GPUs.

Train on the ImageNet:

python train_imagenet.py --epochs 320 --batch-size 32 --opt sgd --lr 0.1 --lr-scheduler cosa --lr-warmup-epochs 5 --lr-warmup-decay 0.01 --amp --model-name lorafb_snet18 --T 4 --output-path logs/

Train on the UCF101:

python train_video.py --dataset UCF101 --frames 16 --f-steps 16 --epochs 100 --batch-size 32 --opt sgd --lr 0.1 --lr-scheduler cosa --lr-warmup-epochs 10 --lr-warmup-decay 0.001 --amp --model-name lorafb_snet18 --output-path logs/

The checkpoints of all pretrained models in the paper are available at https://drive.google.com/drive/folders/1o9PtoLEYdXkZytMprWsSITYjEDGMSZaT?usp=sharing.

Time-Series Representational Similarity

The code is stored in the file folder similarity.

The experiments are performed on a subset of the Allen Brain Observatory Visual Coding dataset, which includes two visual stimuli (Movie1 and Movie2). For data preprocessing, please see the official implementation of the public dataset and the introduction of the paper.

The preprocessed neural data are available at https://drive.google.com/drive/folders/1zJcN-kFxoPwyVT5RPEhKDpo60r7PBPem?usp=sharing and the preprocessed visual stimuli are available at https://drive.google.com/drive/folders/1MNCWc5FIvcibx6BmFPicVk_0pJEJbFQT?usp=sharing.

Test on the original movie:

python similarity.py --model lorafb_snet18 --train-dataset ucf101 --checkpoint-path model_checkpoint/ucf101/lorafb_snet18.pth --neural-dataset allen_natural_movie_one --neural-dataset-dir neural_dataset/ --metric TSRSA --stimulus-dir stimulus/ --output-dir results/

Test on the original movie with different movie clip lengths:

python similarity.py --model lorafb_snet18 --train-dataset ucf101 --checkpoint-path model_checkpoint/ucf101/lorafb_snet18.pth --neural-dataset allen_natural_movie_one --neural-dataset-dir neural_dataset/ --metric TSRSA --stimulus-dir stimulus/ --output-dir results/ --trial-for-clip 10 --clip-len 300

Test on the shuffled movie:

python similarity.py --model lorafb_snet18 --train-dataset ucf101 --checkpoint-path model_checkpoint/ucf101/lorafb_snet18.pth --neural-dataset allen_natural_movie_one --neural-dataset-dir neural_dataset/ --metric TSRSA --stimulus-dir stimulus/ --output-dir results/ --trial-for-ablation 10 --shuffle --window 300 --best-layer

Test on the movie with replaced frames:

python similarity.py --model lorafb_snet18 --train-dataset ucf101 --checkpoint-path model_checkpoint/ucf101/lorafb_snet18.pth --neural-dataset allen_natural_movie_one --neural-dataset-dir neural_dataset/ --metric TSRSA --stimulus-dir stimulus/ --output-dir results/ --trial-for-ablation 10 --replace --replace-type gaussian --window 5 --best-layer

Citation

If you find our work is useful for your research, please kindly cite our paper:

@misc{huang2024longrangefeedbackspikingnetwork,
      title={Long-Range Feedback Spiking Network Captures Dynamic and Static Representations of the Visual Cortex under Movie Stimuli}, 
      author={Liwei Huang and Zhengyu Ma and Liutao Yu and Huihui Zhou and Yonghong Tian},
      year={2024},
      eprint={2306.01354},
      archivePrefix={arXiv},
      primaryClass={cs.NE},
      url={https://arxiv.org/abs/2306.01354}, 
}