S-TLLR: STDP-inspired Temporal Local Learning Rule for Spiking Neural Networks

TL;DR

S-TLLR introduces a biologically inspired, three-factor temporal local learning rule for spiking neural networks that uses an instantaneous STDP-like eligibility trace modulated by a learning signal. By dropping the recurrent state from the eligibility trace, it achieves memory complexity $O(n)$ and time-local updates, while incorporating both causal and non-causal timing relations via a secondary activation function. Across vision, audio, and optical-flow tasks on event-based datasets, S-TLLR delivers competitive accuracy compared with BPTT, with large reductions in memory ($5$–$50\times$) and MACs (up to $6.6\times$), and benefits are enhanced by including non-causal terms. This makes online, edge-friendly learning feasible for deep SNNs without sacrificing significant performance.

Abstract

Spiking Neural Networks (SNNs) are biologically plausible models that have been identified as potentially apt for deploying energy-efficient intelligence at the edge, particularly for sequential learning tasks. However, training of SNNs poses significant challenges due to the necessity for precise temporal and spatial credit assignment. Back-propagation through time (BPTT) algorithm, whilst the most widely used method for addressing these issues, incurs high computational cost due to its temporal dependency. In this work, we propose S-TLLR, a novel three-factor temporal local learning rule inspired by the Spike-Timing Dependent Plasticity (STDP) mechanism, aimed at training deep SNNs on event-based learning tasks. Furthermore, S-TLLR is designed to have low memory and time complexities, which are independent of the number of time steps, rendering it suitable for online learning on low-power edge devices. To demonstrate the scalability of our proposed method, we have conducted extensive evaluations on event-based datasets spanning a wide range of applications, such as image and gesture recognition, audio classification, and optical flow estimation. In all the experiments, S-TLLR achieved high accuracy, comparable to BPTT, with a reduction in memory between $5-50\times$ and multiply-accumulate (MAC) operations between $1.3-6.6\times$.

Figures (5)

• Figure 1: Comparison of weight update computation of a feed-forward spiking layer for BPTT, STDP, and our proposed learning rule S-TLLR. The spiking layer is unrolled over time for the three algorithms while showing the signals involved in the weight updates. The top-down learning signal is shown in green, while the signals locally available to the layer are represented as red for the causal term and blue for non-causal terms. Also, note that the learning signal in BPTT relies on future time steps, whereas in S-TLLR, this signal is computed locally in time.
• Figure 2: Comparison of weight update computation of three feed-forward spiking layers for BPTT and S-TLLR. The spiking layers are unrolled over time for both algorithms while showing the signals involved in the weight updates. The top-down learning signal for BPTT ($\frac{\partial \mathcal{L}}{\partial y [t]}$) is shown in green, note that at time step $t$ this learning signal depends on future time steps. In contrast, for S-TLLR the learning signal ($[\frac{\partial \mathcal{L}}{\partial y [t]}]_{local}$) and all the variables involved are computed with information available only the time step $t$.
• Figure 3: GPU memory usage for BPTT and S-TLLR for a five-layer fully connected SNN models with a different number of time steps ($T$).
• Figure 4: Evaluating the effect of $\alpha_{\mathrm{post}}$ on DVS Gesture and SHD datasets using DFA for learning signal generation. Constant STDP parameters for DVS Gesture are $(\lambda_{post}, \lambda_{pre}, \alpha_{pre}) =(0.2, 0.75, 1)$, and for SHD, they are $(\lambda_{post}, \lambda_{pre}, \alpha_{pre})=(0.5, 1, 1)$. The solid purple line represents the median value, and the dashed black line represents the mean value averaged over five trials.
• Figure 5: Effects of using a decaying factor, $\lambda_{\mathrm{pre}}$, different from the leak spiking parameter ($\gamma=0.5$) to compute the causal term on the eligibility trace, with constant parameters $(\lambda_{post}, \alpha_{post}, \alpha_{pre}) =(0.2, -1, 1)$ when the learning signal is generated using BP. Plots are based on five trials. The solid purple line represents the median value, and the dashed black line represents the mean value.