LISNN: Improving Spiking Neural Networks with Lateral Interactions for Robust Object Recognition

Xiang Cheng, Yunzhe Hao, Jiaming Xu, Bo XuIJCAI 2020 · DOI

Introduction

Spiking Neural Networks (SNNs) represent the "third generation" of neural network models, designed to more closely emulate the biological brain's energy efficiency and temporal processing capabilities. While traditional Artificial Neural Networks (ANNs) have achieved remarkable success, they often rely on purely feedforward architectures. In biological systems, however, intra-layer lateral interactions play a critical role in signal processing, such as edge enhancement and noise filtering in the retina.

This paper introduces LISNN (Lateral Interaction Spiking Neural Network), a model that incorporates these biological lateral connections into SNNs to enhance feature extraction and provide inherent robustness against environmental noise.

Architecture and Methodology

The core innovation of LISNN is the integration of lateral interactions into the standard Leaky Integrate-and-Fire (LIF) neuron model. In a typical SNN, a neuron's state is influenced by its previous membrane potential and the current input from the preceding layer. LISNN introduces a third component: the spiking states of neighboring neurons within the same layer from the previous time step.

The LISNN architecture featuring convolutional layers with lateral interactions between LIF neurons and mean pooling.

The Lateral Interaction Term

These lateral interactions are modeled via a trainable interaction kernel ($\omega$). By convolving the local spiking activity with this kernel, the network learns to either excite or inhibit neighboring neurons. This mechanism allows the network to dynamically "sharpen" relevant features, effectively acting as a spatial filter that distinguishes objects from background noise.

Training and Implementation

To maintain high information density in spike trains, LISNN utilizes mean pooling rather than max pooling, which can lead to significant information loss in discrete temporal signals. Training is conducted using backpropagation with a pseudo-derivative method (a rectangular function) to handle the non-differentiable nature of discrete spikes. For static image input, a Poisson-like encoding scheme is employed, where pixel intensities are translated into firing frequencies.

Experimental Results

LISNN was evaluated on three primary benchmarks: MNIST, Fashion-MNIST, and N-MNIST (a neuromorphic dataset). The model achieved state-of-the-art (SOTA) performance on Fashion-MNIST with an accuracy of 92.07%, outperforming other spiking models like ST-RSBP and HM-2BP. It also reached competitive accuracies of 99.50% on MNIST and 99.45% on N-MNIST.

Robustness to Noise

A defining characteristic of LISNN is its resilience to image degradation. When subjected to Gaussian and Impulse noise, LISNN maintains significantly higher classification accuracy compared to standard feedforward SNNs.

Classification accuracy of LISNN versus standard SNN models across increasing levels of noise.

As shown in the performance analysis, while standard SNNs experience sharp declines in accuracy as noise levels increase, LISNN's performance decays much more gracefully. Notably, this robustness is inherent to the architecture; the model demonstrates these denoising capabilities even when it has not been specifically trained on noisy datasets.

Conclusion

LISNN demonstrates that incorporating biologically inspired lateral interactions can significantly improve both the performance and the robustness of Spiking Neural Networks. By learning to manage intra-layer dependencies, LISNN provides a structural solution to noise interference that feedforward models lack. Future research will focus on scaling these lateral interaction mechanisms to more complex, high-resolution datasets like ImageNet and exploring their implementation on specialized neuromorphic hardware to maximize energy efficiency.

Summary updated

BibTeX
@inproceedings{cheng2020lisnn,
  author = {Xiang Cheng and Yunzhe Hao and Jiaming Xu and Bo Xu},
  title = {{LISNN: Improving Spiking Neural Networks with Lateral Interactions for Robust Object Recognition}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.24963/ijcai.2020/211},
}
Figure