Exploring Visual Representation Using Hierarchical Temporal VAEs

Our group has developed hierarchical VAE models that can capture top-down effects of biological vision[1]. We want to extend these models to handle temporal data and generate dynamic scenes. Our main goal is to separate the content and motion factors in these models, which is not only relevant for machine learning but also essential for understanding how the primate brain processes visual information in different pathways, such as the ventral stream (e.g., V1 and V2) and the dorsal stream (MT).