UC San Diego

Advancing our understanding of neuroscience and artificial intelligence, this dissertation aims to progress our understanding of memory representation, consolidation, and robustness within neural networks. While the brain serves as a remarkable inspiration for machine learning, our comprehension of its complexities remains limited. Gaining insight in how the brain operates enables mutual progress in both fields simultaneously, one potential avenue is through exploring sleep. Sleep is a significant yet only partially understood phenomena that occurs in biological brains. This critical physiological process is prevalent across species due to its pivotal role for many biologically relevant metabolic and cognitive functions; importantly sleep has been shown to be crucial for memory enhancement and consolidation. Despite the extreme importance of natural sleep, there is no true artificial counterpart in machine learning. This work elucidates the intricate mechanisms by which sleep enhances memory representation through biophysical modeling and applies these principals to a range of network architectures across the biophysical-artificial spectrum for a variety of tasks. Specifically, sleep mechanisms are conceptualized and illustrated in biophysical Hodgkin-Huxley neural networks capable of realistic wake and sleep activity. Similar sleep-like stages are then applied to map-based spiking neural networks to mitigate catastrophic forgetting in a sequential learning paradigm. Finally, fully bridging the neuroscience / artificial intelligence gap, a sleep based algorithm for artificial convolutional neural networks is proposed which bolsters the resilience of convolutional filters thereby improving model performance in distorted contexts. Collectively, this dissertation sheds light on the role of sleep in shaping memory across diverse neural systems and reimagines the relationship between artificial and biological intelligence.