Sparse Representations in Artificial and Biological Neural Networks

Abstract

This thesis explores how sparsity, the idea that only a small fraction of neurons are active at any time, is a common thread connecting biological brains and artificial intelligence. By combining theory, experiments, and real-world applications, we show how sparsity is a key ingredient underlying core cognitive abilities like attention, memory, and learning. We start by uncovering a surprising link between the "attention" mechanism powering recent artificial intelligence (AI) breakthroughs and a classic theory of human memory called Sparse Distributed Memory (SDM). This suggests that brains and AI may leverage similar computational tricks. Taking inspiration from the brain's cerebellum, we then use SDM to improve an AI's ability to learn continuously without forgetting previous knowledge. This showcases sparsity's ability to enable more flexible learning. We also find that simply adding noise during training pushes AI to use sparse representations, causing it to develop more brain-like properties. This provides clues about why sparsity emerges in the brain while offering an easy way to encourage it in AI. Finally, we use sparsity to peek inside the black box of large language models like ChatGPT and Claude. By pulling apart the tangled web of information these models use to think, we make progress towards more transparent and controllable AI. Together, these findings paint sparsity as a unifying principle for intelligent systems, be they made of biological neurons or silicon chips. By connecting the dots between neuroscience and AI, this thesis advances our understanding of intelligence while charting a course towards more capable and interpretable AI systems.