Information representation space in artificial and biological neural networks

Abstract

Convolutional neural networks are often used as a model of the primate visual system. However, it is often overlooked how a task that the network performs and statistics of the training set affect the representation of information in the latent space of the model. This study demonstrates that the properties of artificial neurons in the first two convolutional layers represent the signal statistics (correlation coefficients R=0.63 and R=0.44), whereas the similarity between the space of the problem and information encoding in hidden layers gradually increases in the final convolutional layers (R=0.35), reaching a value of 0.73 in the fully-connected layers. At the final stages of the processing, a category is encoded using a unique set of features, characterized by no or little overlapping with other categories. Thus, in order to increase similarity between the visual system and its model, it is important to maintain a training set and a problem space of the model coherent to those of a biological organism.