Efficient Online Learning for Networks of Two-Compartment Spiking Neurons

TL;DR

This work tackles training efficiency for temporally rich neural circuits by extending the e-prop online learning framework to two-compartment TC-LIF neurons and introducing Adaptive TC-LIF with time-varying memory decays. It derives TC-LIF eligibility traces, redesigns the neuron’s parameter space to support online updates, and extends e-prop to multi-layer SNNs, enabling online, memory-efficient training. Across sequential benchmarks, the approach achieves competitive accuracies with online learning, closely matching offline BPTT while significantly reducing memory requirements. The results demonstrate the practical viability of high-capacity multi-compartment SNNs on neuromorphic hardware for processing temporal signals.

Abstract

The brain-inspired Spiking Neural Networks (SNNs) have garnered considerable research interest due to their superior performance and energy efficiency in processing temporal signals. Recently, a novel multi-compartment spiking neuron model, namely the Two-Compartment LIF (TC-LIF) model, has been proposed and exhibited a remarkable capacity for sequential modelling. However, training the TC-LIF model presents challenges stemming from the large memory consumption and the issue of gradient vanishing associated with the Backpropagation Through Time (BPTT) algorithm. To address these challenges, online learning methodologies emerge as a promising solution. Yet, to date, the application of online learning methods in SNNs has been predominantly confined to simplified Leaky Integrate-and-Fire (LIF) neuron models. In this paper, we present a novel online learning method specifically tailored for networks of TC-LIF neurons. Additionally, we propose a refined TC-LIF neuron model called Adaptive TC-LIF, which is carefully designed to enhance temporal information integration in online learning scenarios. Extensive experiments, conducted on various sequential benchmarks, demonstrate that our approach successfully preserves the superior sequential modeling capabilities of the TC-LIF neuron while incorporating the training efficiency and hardware friendliness of online learning. As a result, it offers a multitude of opportunities to leverage neuromorphic solutions for processing temporal signals.

Figures (9)

• Figure 1: Comparison of the proposed neuron models and learning algorithms. (a) The vanilla TC-LIF neuron captures the interaction between dendritic and somatic compartments in biological neurons to facilitate temporal signal processing. (b) The proposed Adaptive TC-LIF neuron further introduces time-varying membrane potentials decaying constants to facilitate temporal information integration during online learning. (c) The vanilla TC-LIF model is trained using BPTT algorithm, where the gradients are propagated from the last time step to all preceding time steps for parameter update. (d) Our proposed Adaptive TC-LIF neuron facilitates efficient online learning by computing parameter updates at each time step based on the local loss and eligibility trace derived at that specific time step. Importantly, the computation of eligibility traces for both neuronal compartments occurs in a forward manner, eliminating the need to store intermediate network states as required in BPTT training.
• Figure 2: The comparison of gradient update principle of BPTT and e-prop algorithms at time $t$. BPTT algorithm updates the network parameters based on the global loss, which can only be obtained at the last time step $T$. In contrast, the e-prop algorithm accumulates eligibility traces that propagate forward in time and multiplies them with the online error term. This allows for the online update of network parameters.
• Figure 4: The impact of hyperparameter setting on the test accuracy of the S-MNIST dataset. In this figure, the dark green color indicates high accuracy, while light red indicates low accuracy. The location where both $\alpha_1$ and $\alpha_2$ are equals to 1.0 corresponds to the vanilla TC-LIF model.
• Figure 5: Comparison of the performance with and without time-varying membrane decaying constants on the SHD dataset.
• Figure 6: Actual GPU memory usage with varying time sequence lengths on the S-MNIST dataset.