When
Where
Speaker: Maksim Bazhenov, Health Sciences, UCSD
Title: Do Neural Networks Dream of Electric Sheep?
Abstract: Although artificial neural networks (ANNs) have demonstrated impressive performance on many tasks, they fall short in critical areas of human intelligence, such as rapid and continual learning, abstraction, and calibrated judgment - the ability to align confidence with actual accuracy. For example, ANNs are prone to catastrophic forgetting, where their performance on previously learned tasks deteriorates when learning new ones sequentially. They also tend to exhibit overconfidence in their predictions, assigning probabilities that are systematically higher than the true likelihood of being correct.
In contrast, human and animal brains possess the remarkable ability to learn continuously, generalize episodic knowledge, and effectively balance certainty and uncertainty. Empirical evidence suggests that sleep plays a crucial role in consolidating and calibrating memories, safeguarding against catastrophic forgetting and enabling the generalization of episodic knowledge to semantic categories.
In my talk, I will first elucidate the key features and mechanisms of sleep in the brain. I will then present our recent findings on the application of sleep-related concepts in computational models of brain networks and artificial intelligence.