SSTDP: Supervised Spike Timing Dependent Plasticity for Efficient Spiking Neural Network Training.

Area of Science:

• Neuromorphic Engineering
• Computational Neuroscience
• Machine Learning

Background:

• Spiking Neural Networks (SNNs) offer potential for low-power, event-driven neuromorphic hardware due to their biological plausibility and spatio-temporal processing.
• Current SNNs lag behind Artificial Neural Networks (ANNs) in accuracy, partly due to challenges in training SNNs with gradient-based methods like backpropagation (BP) owing to non-differentiable spike events.

Purpose of the Study:

• To introduce a novel learning algorithm, SSTDP, that bridges the gap between BP and Spike-Time-Dependent Plasticity (STDP) for efficient SNN training.
• To improve both the accuracy and reduce the latency of SNNs, addressing key limitations compared to ANNs.

Main Methods:

• The SSTDP algorithm integrates global optimization from BP with efficient weight updates from STDP.
• It avoids non-differentiable derivations in BP and leverages STDP's local feature extraction.
• The method utilizes temporal-based coding and Integrate-and-Fire (IF) neuron models for computational efficiency.

Main Results:

• SSTDP achieved high classification accuracies: 99.3% on Caltech 101, 98.1% on MNIST, and 91.3% on CIFAR-10.
• Compared to other SNN training methods, SSTDP-trained SNNs demonstrated 25-32x lower inference latency.
• Event-based computations showed SSTDP methods reduce addition operations by 1.3-37.7x per inference.

Conclusions:

• SSTDP effectively trains SNNs, significantly improving accuracy and reducing latency.
• The algorithm addresses the vanishing spike issue in BP and reduces time steps for faster inference.
• SSTDP demonstrates the efficacy of SNNs for efficient inference operations in the spiking domain.