Advancing NeuroAI through Biologically-Plausible Self-supervised Learning

The field of neuroscience-informed artificial intelligence (NeuroAI) seeks to bring together neuroscience and artificial intelligence (AI) around a shared set of principles that both explain how the brain works and guide the design of smarter, more capable AI systems. Today's leading AI systems, while impressive, are different from the brain in important ways: they have more challenges with unexpected situations and they require enormous resources to train and implement. This project investigates the gap between artificial and natural intelligence with two key goals in mind. First, it draws on recent advances in machine learning to develop an updated scientific theory of how the brain's sensory systems learn to make sense of the world. Second, this project uses those insights to build AI models that are more efficient, more robust, and more consistent with how the brain works. A central focus of this project is a form of machine learning called self-supervised representation learning (SSRL). Representation learning is the process by which a brain or AI system learns to convert raw sensory input, such as light hitting the retina, into a compact internal format useful for thinking and acting. Self-supervision means learning to do this autonomously without an expert telling the brain or model what to look for. One key difference between current AI systems and the brain is that AI tends to process information in a feedforward direction only; in contrast, the brain is full of feedback loops, also known as recurrent connections. This project's first aim is to build NeuroAI models for SSRL that incorporate such feedback loops; the goal is to produce systems to empower AI that are both smaller and more capable. The project’s second aim is to test whether these new models exhibit phenomena that are known to depend on feedback in the brain; the goal is to provide an updated mechanistic understanding of how the brain processes its sensory inputs, in particular how our brains make sense of degraded, incomplete, and otherwise ambiguous inputs. The project’s third aim is to develop detailed models of how individual brain connections change and strengthen during learning, derived from these same SSRL principles; the goal is to advance our understanding of the biological machinery of learning itself as well as foster important progress in NeuroAI and computational neuroscience research. 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.