Training deep learning models to classify input to simulations of a biological neural network

Abstract

We present a computationally effective toy model of the visual system of a biological brain, that can easily be extended to add more realism. The model takes images as input – representing visual stimuli from the eye – and outputs an estimation of the cortical LFP (local field potential) that is generated as cortex processes the input. We run a large number of simulations, each stimulated by a randomized sequence of 10 images, and use the output data to train deep learning algorithms (CNN and LSTM) to classify pieces of the LFP by input image. The classifiers reach accuracies of 66 and 65%, averaged across all 10 inputs, suggesting that the LFP indeed contain information about the stimulus that a brain is processing. They are also more likely to confuse the LFPs of images that qualitatively seem visually similar. We observe that a trained CNN transfers better to test data that deviates slightly from the training set, but that the LSTM seems marginally better at handling noise.