A comparative analysis of Gabor filters and biologically inspired learning rules for image classification implementing Spiking Neural Networks

Abstract

Plastic changes on the synapse drive by spike-timing have been of great interestas the main learning rule for spiking neural networks. Spike-timing-based rules arebuilt to model the behavior of a region on the brain related to experimental data inneuroscience, therefore can lead to differences in the moment of implementing therule within a spiking network. This work compares the performance of a pair-wiseand a triplet STDP with different spike interactions to clear an MNIST classificationtask. A bio-inspired preprocessing stage was implemented that consisted of a Gaborfilter (as a model of the simple cells mechanism orientation selectivity) and an inputnormalization for a homogeneous brightness level of each image. The highlightsof this work are 1) The consistent improvement of the model accuracy wheneverthey added the Gabor filter to the inputs; 2) The input normalization to preventthe overfitting of the model; 3) The Gabor filter helps to correct decoding of someimages of the dataset on the evaluation test.