SSTDP

SSTDP integrates supervised spike-timing-dependent plasticity with backpropagation to train spiking neural networks for accurate, low-latency classification using Integrate-and-Fire neurons and temporal-based coding.

Key Features:

• BP–STDP integration: Integrates global optimization via backpropagation (BP) with local weight updates from spike-timing-dependent plasticity (STDP).
• Integrate-and-Fire neuron model: Employs an Integrate-and-Fire (IF) neuron model to provide computational advantages.
• Temporal-based coding: Uses temporal-based coding to leverage spatio-temporal information processing in SNNs.
• Non-differentiability mitigation: Mitigates the non-differentiability of discrete binary spike events to enable effective training of SNNs.
• Local feature extraction and stability: Enhances local feature extraction while reducing network latency and minimizing vanishing spikes during training.
• Classification performance: Achieves 99.3% on Caltech 101, 98.1% on MNIST, and 91.3% on CIFAR-10, surpassing other SNN training methods.
• Latency reduction: Delivers inference latency reductions of 25–32× compared with directly trained SNNs.
• Event-based efficiency: Reduces the number of addition operations per inference by factors ranging from 1.3× to 37.7×, improving event-based computational efficiency.
• Neuromorphic targeting: Suited for low-power, event-driven neuromorphic hardware by leveraging spiking, event-based computations and biological plausibility.

Scientific Applications:

• Image classification: Enables high-accuracy classification on datasets including Caltech 101, MNIST, and CIFAR-10.
• Low-latency neuromorphic inference: Supports inference with 25–32× lower latency for deployment on event-driven hardware.
• Event-based computation efficiency: Reduces arithmetic operations per inference (1.3–37.7× reduction) to improve energy and computational efficiency.
• SNN training research: Provides a hybrid training approach to narrow the performance gap between spiking neural networks (SNNs) and artificial neural networks (ANNs).

Methodology:

SSTDP computationally combines supervised spike-timing-dependent plasticity (STDP) with backpropagation (BP), employs temporal-based coding, and implements an Integrate-and-Fire (IF) neuron model.