Aligning Deep Networks with Human Vision - Novel Neural Architectures, Data Diets and Training Algorithms

Explore a thought-provoking lecture from Brown University's Thomas Serre examining the challenges and opportunities in aligning deep neural networks (DNNs) with human visual perception. Delve into the current state of artificial intelligence, where despite unprecedented scaling and sophisticated architectures, DNNs still exhibit peculiar failures that deviate from human-like behavior. Learn about innovative research aimed at developing DNNs that better mirror human perception by incorporating fundamental principles of natural intelligence. Discover ongoing efforts to characterize human visual strategies in image categorization tasks and understand how they differ from modern deep nets. Examine why novel data regimens and training algorithms may be necessary for developing more human-like visual representations. Investigate how cortex-inspired recurrent neural circuits could provide advantages over transformer architectures, particularly for complex visual reasoning tasks beyond basic categorization.