Core Concepts
Equilibrium Propagation (EP) can be used to efficiently train spiking convolutional neural networks, achieving performance comparable to state-of-the-art methods while being more memory-efficient.
Abstract
The paper explores how spiking convolutional neural networks can be trained using Equilibrium Propagation (EP), a biologically plausible local learning algorithm. The key contributions are:
Formulation of spiking convolutional layers in the context of energy-based systems trained by EP. This bridges the gap between spiking and non-spiking convergent recurrent neural networks (RNNs) in the regime of EP.
Theoretical and experimental analysis on the issues with maximum pooling and unpooling operations in spiking convergent RNNs trained by EP. The information mismatch between forward and backward connections significantly degrades the performance.
Proposal to replace maximum pooling and unpooling with average pooling and nearest neighbor upsampling, which solves the aforementioned problem. This enables spiking convergent RNNs trained by EP to achieve state-of-the-art performance on MNIST (0.97% test error) and FashionMNIST (8.89% test error) datasets, comparable to networks trained by backpropagation.
Demonstration of the memory efficiency of EP compared to backpropagation through time (BPTT) for training spiking neural networks.
The findings highlight EP as an optimal choice for on-chip training and a biologically plausible method for computing error gradients in spiking neural networks.
Stats
The average activation summation of the middle 3 convolutional layers in the forward route is 0.9x, 0.8x, and 0.8x of the backward route in convergent RNNs, while it is 0.9x, 1.0x, and 0.9x in spiking neural networks.
The standard deviation of activations in the forward route is 0.9x, 0.7x, and 0.7x of the backward route in convergent RNNs, while it is 3.1x, 2.6x, and 2.3x in spiking neural networks.
Quotes
"Equilibrium Propagation (EP) leverages unified circuitry and inherent dynamics of neural systems to adjust synaptic weights, which is analogous to how learning might occur in biological systems."
"Spiking convergent RNNs trained by EP with the proposed modifications achieve performance at par with SNNs trained by BPTT and convergent RNNs trained by BPTT and EP, while they consume notable reduction in memory footprint compared to those trained by BPTT."