Learning object representations in deep neural networks
Abstract
Humans have the ability to learn visual representations of the surrounding environment with limited supervision. A major challenge in cognitive neuroscience is to understand the neural computations that give rise to this ability. Recent work has started modelling the neural computations implemented by the ventral visual system using deep convolutional neural networks (DCNNs). Despite their successes, DCNNs leave substantial amounts of variance in brain representations unexplained. We hypothesize that this may in part be due to the DNNs' sole reliance on supervision during representation learning. In this thesis, we investigate the role of training algorithms (supervised versus unsupervised) on the representational similarity between the computational models and brain data from human inferior temporal cortex. We show that one implementation of unsupervised contrastive learning yields more brain-like representations than the selected supervised learning method. Our findings suggest that human visual representations may in part arise from unsupervised learning during development.