MNIST IMAGE CLASSIFICATION USING LIF NEURONS AND STDP ALGORITHM

Abstract

Spiking neural networks are inspired by neural system principles and may offer previously new advantages in the machine learning context. Since most SNN architectures rely on rate-based network training and then convert to an SNN, there has been a growing interest in how spiking neural networks (SNN) can be used to solve pattern recognition tasks or complex computations. We present a digit recognition SNN based on biologically relevant mechanisms like spike-timing-dependent plasticity (STDP) with an adaptive spiking threshold, time-dependent weight change, and lateral inhibition. Our architecture achieves 95% reliability on the MNIST benchmark using this unsupervised learning strategy, which is better than earlier SNN implementation. Image features can be chosen to be shared by all neurons in a subset of the population, or they can evolve independently to represent different characteristics in different areas of input space. The MNIST dataset is used as a platform for experimental testing. A baseline spiking neural network's performance and convergence speed are compared. The absence of domain-specific knowledge suggests that our network design is broadly applicable. Our network's performance scales well with the number of neurons used. It achieves comparable performance when using multiple learning rules, indicating the dependability of the combined effect of methods and implying applicability in non-homogeneous biological neural networks.