Multi-Bit Mechanism: A Novel Information Transmission Paradigm for Spiking Neural Networks

TL;DR

The paper addresses information loss and limited expressiveness in spiking neural networks caused by single-bit spikes and layer-by-layer information flow. It introduces a multi-bit spike mechanism that extends spikes to fixed-point representations with $m$ integer bits and $n$ fractional bits, paired with interlaminar connections enhanced by Efficient Channel Attention to re-stimulate neurons. The approach preserves the MAC-to-AC computation advantage of SNNs and yields consistent improvements on both static (CIFAR-10/100) and neuromorphic (DVS-Gesture) datasets, including strong performance at ultra-low time steps. This work proposes a new information transmission paradigm for SNNs with modest parameter overhead and potential for energy-efficient, high-precision neuromorphic computing.

Abstract

Since proposed, spiking neural networks (SNNs) gain recognition for their high performance, low power consumption and enhanced biological interpretability. However, while bringing these advantages, the binary nature of spikes also leads to considerable information loss in SNNs, ultimately causing performance degradation. We claim that the limited expressiveness of current binary spikes, resulting in substantial information loss, is the fundamental issue behind these challenges. To alleviate this, our research introduces a multi-bit information transmission mechanism for SNNs. This mechanism expands the output of spiking neurons from the original single bit to multiple bits, enhancing the expressiveness of the spikes and reducing information loss during the forward process, while still maintaining the low energy consumption advantage of SNNs. For SNNs, this represents a new paradigm of information transmission. Moreover, to further utilize the limited spikes, we extract effective signals from the previous layer to re-stimulate the neurons, thus encouraging full spikes emission across various bit levels. We conducted extensive experiments with our proposed method using both direct training method and ANN-SNN conversion method, and the results show consistent performance improvements.

Figures (7)

• Figure 1: Neurons that emit single-bit spikes and neurons that emit multi-bit spikes.
• Figure 2: Overview of the proposed method. Our approach is grounded in the Spiking ResNet-20 architecture, where we enhance LIF neurons with a multi-bit mechanism to build an MBLIF model, thereby increasing the information density of neural spikes. Additionally, we incorporate interlaminar connections within the ResNet's basic block, aiming to preserve information maximally and ensure MBLIF's efficient spike generation across all bit positions.
• Figure 3: (a)The range of values for the original spikes is $\{0,1\}$; (b)Expanding the spike by one integer bit upwards, the range of spike values is expanded to $\{0,1,2,3\}$; (c)Expanding the spike by one decimal place downward, the range of spike values expands to $\{0,0.5,1,1.5\}$.
• Figure 4: The diagram represents the case of $m=2, n=1$, where the green matrix denotes the activation values corresponding to three bit positions with weights $2^1, 2^0,$ and $2^{-1}$. These weights are absorbed by the blue convolutional layer parameter matrix and remain unchanged during the inference process. Thus, the final computation between the multi-bit spikes and the convolutional layer parameters is still addition.
• Figure 5: Heatmaps of the membrane potentials in a certain layer, with colors ranging from blue to yellow indicating neurons transitioning from silent to active.