Conference on Cognitive Computational Neuroscience

Artificial neural networks have helped investigate questions that were previously beyond the scope of neuroscience. However, the most commonly used model architecture, convolutional neural networks (CNNs), are limited in the questions they can answer. One such question concerns the emergence of cortical topography and its relation with visual behaviour. CNNs as a model class cannot address this question, as their units share features invariantly of their locations. Here, we overcome this limitation by introducing end-to-end Topographic Neural Networks (All-TNNs). Similar to the brain, units of All-TNNs have varying selectivity and are all arranged in a 2D cortical sheet. Trained on an object classification task, the neural network develops several topographical features reminiscent of the human visual system: the first model layer displays smooth orientation selectivity maps, while category-selective areas emerge in later layers. We test the alignment between ANN models (All-TNNs and CNNs) and the brain by conducting a psychophysical experiment in which we characterise humans' positional dependence during object classification. We show that All-TNNs display a more human-like dependence on spatial location for object recognition. Together, these results pave the way towards biologically more faithful modelling of the biological visual system via All-TNNs.