[Paper Review] Event-Driven Learning for Spiking Neural Networks

Wenjie Wei, Malu Zhang+8 more|arXiv (Cornell University)|Mar 1, 2024

Advanced Memory and Neural Computing14 citations

TL;DR

The paper introduces two novel event-driven learning algorithms, STD-ED and MPD-ED, for deep spiking neural networks, addressing over-sparsity and gradient reversal, and demonstrates state-of-the-art performance and substantial energy savings on neuromorphic hardware.

ABSTRACT

Brain-inspired spiking neural networks (SNNs) have gained prominence in the field of neuromorphic computing owing to their low energy consumption during feedforward inference on neuromorphic hardware. However, it remains an open challenge how to effectively benefit from the sparse event-driven property of SNNs to minimize backpropagation learning costs. In this paper, we conduct a comprehensive examination of the existing event-driven learning algorithms, reveal their limitations, and propose novel solutions to overcome them. Specifically, we introduce two novel event-driven learning methods: the spike-timing-dependent event-driven (STD-ED) and membrane-potential-dependent event-driven (MPD-ED) algorithms. These proposed algorithms leverage precise neuronal spike timing and membrane potential, respectively, for effective learning. The two methods are extensively evaluated on static and neuromorphic datasets to confirm their superior performance. They outperform existing event-driven counterparts by up to 2.51% for STD-ED and 6.79% for MPD-ED on the CIFAR-100 dataset. In addition, we theoretically and experimentally validate the energy efficiency of our methods on neuromorphic hardware. On-chip learning experiments achieved a remarkable 30-fold reduction in energy consumption over time-step-based surrogate gradient methods. The demonstrated efficiency and efficacy of the proposed event-driven learning methods emphasize their potential to significantly advance the fields of neuromorphic computing, offering promising avenues for energy-efficiency applications.

Motivation & Objective

• Assess limitations of existing event-driven learning algorithms for deep SNNs.
• Develop two new event-driven learning methods that leverage spike timing or membrane potential.
• Address over-sparsity and gradient reversal to enable stable training.
• Demonstrate performance improvements on static and neuromorphic datasets.
• Validate energy efficiency of the proposed methods on neuromorphic hardware.

Proposed method

• Propose Adaptive Firing Threshold-based Integrate-and-Fire (AFT-IF) neurons to mitigate over-sparsity and gradient reversal.
• Develop STD-ED where learning occurs at spike times with derivatives computed via membrane potential and spike timing relationships.
• Introduce MPD-ED using AFT-LIF neurons and a masked surrogate gradient (MSG) that enables event-driven backpropagation through membrane potential signals.
• Use linear PSP kernel (for STD-ED) and membrane-potential-based learning signals (for MPD-ED) to avoid gradient reversal.
• Derive backpropagation equations for both methods to update weights and spike timings.
• Evaluate on static (F-MNIST, CIFAR-10/100) and neuromorphic (N-MNIST, DVS-Gesture, DVS-CIFAR10) datasets.

Experimental results

Research questions

• RQ1 Can fully event-driven learning approaches (STD-ED and MPD-ED) outperform existing event-driven methods on challenging datasets?
• RQ2 Do adaptive firing threshold mechanisms and membrane-potential signals enable stable, efficient training of deep SNNs?
• RQ3 What are the trade-offs between spike-timing and membrane-potential-based learning in terms of accuracy and efficiency?
• RQ4 How much energy efficiency can be achieved on neuromorphic hardware with on-chip event-driven learning?

Key findings

• STD-ED and MPD-ED achieve superior performance over existing event-driven methods on benchmark datasets.
• On CIFAR-100, STD-ED and MPD-ED outperform counterparts by up to 2.51% and 6.79%, respectively.
• On-chip learning experiments show a 30-fold energy reduction compared to time-step-based surrogate gradient methods.
• The proposed methods are validated on both static and neuromorphic datasets, demonstrating their effectiveness for energy-efficient neuromorphic computing.
• Energy efficiency and practicality are supported by theoretical analysis and hardware implementation results.