EFRI BRAID: Principles of sleep-dependent memory consolidation for adaptive and continual learning in artificial intelligence

University Of California-San Diego, La Jolla CA

Investigators

Abstract

Artificial Neural Networks (ANNs) are a form of Artificial Intelligence (AI) used in applications from self-driving cars to medicine to robotic systems. Although they can match and even exceed human performance on some learning tasks, they fail to reproduce important characteristics of the human mind, such as quick and continual learning, transfer of knowledge to the new tasks, and energy efficiency. Indeed, ANNs commonly forget what they knew when new information is learned, and so they need to be taught from scratch to re-learn. In real-life applications in changing and unpredictable environments, ANNs can only reach near human-level performance if they are trained on all possible scenarios that could happen in life. This level of training is inefficient and unrealistic. In natural brains, sleep is thought to be important for intelligence. During sleep, the brain repeats and replays what was learned during the day, and this helps to prevent forgetting, to generalize to new situations, and to create new emerging knowledge. In this project, principles learned from the biology of sleep will be used to develop powerful new algorithms for AI systems that can learn continuously and from few examples, transfer knowledge learned from old tasks to new tasks, and be robust and efficient. Because AI and ANNs are so fundamental to the modern world, from healthcare to electronics to national defense, this project has the potential to make a significant societal impact. The project takes a multi-disciplinary approach and supports broader participation of underrepresented groups in STEM research through a range of educational activities focused on high school, undergraduate, including community college, and graduate students. This project aims to translate insights from the study of sleep to improvements in deep-learning systems necessary for continual learning, generalization, and transfer of knowledge in artificial intelligence (AI). Taking advantage of the architectural similarities between information processing in ANNs and the honeybee brain, the main goals of this project are: (a) to characterize multi-phasic sleep in the honeybee brain in vivo and in biophysical in silico models in fine spatio-temporal detail to reveal the critical principles of the role of sleep in memory consolidation, and (b) to apply these results to support the development of novel machine-learning algorithms for adaptive and continual learning in complex and dynamic environments. The study will develop an empirically grounded theory of multi-phasic sleep that will be then applied to artificial neural networks, and the process of developing “sleep for AI” will help to strengthen connections between engineering, computational neuroscience, and neuroethology for researchers at a range of career stages. To accomplish this goal, the project team also plans a four-tiered educational approach targeting students in high schools, community colleges, bachelor’s degree programs, and graduate-level programs to introduce a wider range of students to the topics in AI, sleep biology, and computational neuroscience. This project is funded jointly by the Emerging Frontiers in Research and Innovation Brain-Inspired Dynamics for Engineering Energy-Efficient Circuits and Artificial Intelligence Program of the Engineering Directorate and the Neural Systems/Modulation Program of the Biological Sciences Directorate. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.