TESS: A Scalable Temporally and Spatially Local Learning Rule for Spiking Neural Networks

TL;DR

TESS introduces a fully local learning rule for spiking neural networks that addresses both temporal and spatial credit assignment using eligibility traces and locally generated learning signals. By enforcing locality, memory scales as $O(Ln)$ and compute as $O(LCn)$, independent of the sequence length $T$, while achieving accuracy close to BPTT on edge-relevant vision tasks. Empirical results show TESS matching or surpassing non-local methods on IBM DVS Gesture, CIFAR10, and CIFAR100, with dramatic reductions in MACs ($205$–$661\times$) and memory ($3$–$10\times$). These findings highlight TESS as a practical, scalable solution for on-device learning in resource-constrained edge environments.

Abstract

The demand for low-power inference and training of deep neural networks (DNNs) on edge devices has intensified the need for algorithms that are both scalable and energy-efficient. While spiking neural networks (SNNs) allow for efficient inference by processing complex spatio-temporal dynamics in an event-driven fashion, training them on resource-constrained devices remains challenging due to the high computational and memory demands of conventional error backpropagation (BP)-based approaches. In this work, we draw inspiration from biological mechanisms such as eligibility traces, spike-timing-dependent plasticity, and neural activity synchronization to introduce TESS, a temporally and spatially local learning rule for training SNNs. Our approach addresses both temporal and spatial credit assignments by relying solely on locally available signals within each neuron, thereby allowing computational and memory overheads to scale linearly with the number of neurons, independently of the number of time steps. Despite relying on local mechanisms, we demonstrate performance comparable to the backpropagation through time (BPTT) algorithm, within $\sim1.4$ accuracy points on challenging computer vision scenarios relevant at the edge, such as the IBM DVS Gesture dataset, CIFAR10-DVS, and temporal versions of CIFAR10, and CIFAR100. Being able to produce comparable performance to BPTT while keeping low time and memory complexity, TESS enables efficient and scalable on-device learning at the edge.

Figures (2)

• Figure 1: Comparison of three learning rule strategies for training a recurrent model with state variable ${\bm{s}}[t]$. (a) Non-local learning method (e.g., BPTT): Both spatial and temporal credit assignment problems are solved by propagating errors through time and space (layers). (b) Temporal local method: Temporal credit assignment is addressed using eligibility traces (${\bm{e}}[t]$), which are auxiliary variables that track the history of neural activity. These traces are modulated by a learning signal ($[\frac{\partial\mathcal{L}}{\partial{\bm{s}}^{(l)}[t]}]_{\text{local}}$), which propagates errors across layers but not through time. (c) Fully local method (e.g., our proposed method TESS): In addition to eligibility traces, the learning signal (${\bm{m}}[t]$) is generated locally, addressing both spatial and temporal credit assignment entirely within the local context.
• Figure 2: Overview of TESS. The diagram illustrates an SNN model unrolled in time, where ${\bm{u}}^{(l)}[t]$ denotes the membrane potential of neurons in the $l$-th layer at time step $t$, and ${\bm{o}}^{(l)}[t]$ represents the corresponding output spikes. Signals involved in weight update computation are highlighted: red represents the eligibility trace based on causal relationships between inputs and outputs, blue represents the eligibility trace for non-causal relationships, and green represents the local learning signal ${\bm{m}}^{(l)}[t]$ used to modulate the eligibility traces. The local learning signal is generated independently for each layer through a learning signal generation (LSG) process. The fixed binary matrix ${\bm{B}}^{(l)}$ used in the LSG process features columns corresponding to square wave functions. While, $f(\cdot)$ is a softmax function, and ${\bm{t}}^*[t]$ represent the labels.