Improving Models of the Ventral Visual Pathway

This lecture explores computational models of the ventral visual system, focusing on optimizing networks for performance on real-world challenging tasks, comparing to brain data, and discussing optimization strategies. It delves into the use of convolutional neural networks, hierarchical and retinotopic structures, and the importance of filter parameters. The lecture also covers the prediction of neural responses, the evaluation of learning rules, and the challenges of unsupervised learning. Additionally, it discusses the emergence of grid-like representations in artificial agents and the application of deep contrastive embeddings in neural predictivity. The lecture concludes with a discussion on the evaluation of learning rules from neural network observables and the implications of deep contrastive embeddings in improving ImageNet performance.