vwOvOwv/DeepEISNN

Training Deep Normalization-Free Spiking Neural Networks with Lateral Inhibition

This repository contains the pretrained checkpoints for the paper Training Deep Normalization-Free Spiking Neural Networks with Lateral Inhibition.

GitHub Repository | Paper (OpenReview)

Introduction

Spiking Neural Networks (SNNs) are energy-efficient and biologically plausible models. This work proposes a normalization-free learning framework that incorporates lateral inhibition inspired by cortical circuits. By replacing traditional feedforward SNN layers with distinct excitatory (E) and inhibitory (I) neuronal populations, the framework achieves stable training of deep SNNs without relying on explicit normalization schemes.

Model Performance

Dataset	Arch	T	Top-1 Acc(%)	Model Name
CIFAR-10	ResNet-18	4	92.06	CIFAR10-ResNet18
CIFAR-10	VGG-8	4	87.03	CIFAR10-VGG8
CIFAR-10	VGG-11	4	88.43	CIFAR10-VGG11
CIFAR-10	VGG-16	4	91.01	CIFAR10-VGG16
CIFAR-10	VGG-19	4	91.36	CIFAR10-VGG19
CIFAR-100	VGG-16	4	65.90	CIFAR100-VGG16
CIFAR-100	VGG-19	4	64.06	CIFAR100-VGG19
CIFAR10-DVS	VGG-8	10	78.40	CIFAR10DVS-VGG8
CIFAR10-DVS	VGG-11	10	78.40	CIFAR10DVS-VGG11
DVS-Gesture	VGG-8	16	95.83	DVSGesture-VGG8
TinyImageNet200	ResNet-18	4	50.29	TinyImageNet200-ResNet18

Usage

Evaluation

Use the following command from the official repository to quickly evaluate the pretrained checkpoints on the evaluation set. The script will automatically download the corresponding weights from Hugging Face.

python eval_pretrained.py --model [Dataset]-[Arch] --data_path [Dataset-Path]

Citation

@inproceedings{
  liu2026training,
  title={Training Deep Normalization-Free Spiking Neural Networks with Lateral Inhibition},
  author={Peiyu Liu and Jianhao Ding and Zhaofei Yu},
  booktitle={The Fourteenth International Conference on Learning Representations},
  year={2026},
  url={https://openreview.net/forum?id=U8preGvn5G}
}