Using Meta-Plasticity to Discover the Biophysics of Learning
Abstract
The brain is unparalleled in its ability to learn efficiently and generalize what it has learned to new tasks. A detailed understanding of how the brain learns effectively and efficiently would vastly improve the two way communication of information between neuroscience and AI, opening floodgates to rapid advances in each field. Advances in deep artificial neural networks (DNNs) have inspired efforts to use them as a computational model for biological learning. While DNNs and biological neuronal networks share some basic building blocks, they differ in some critical ways. Notably, DNNs learn using an algorithm called backpropagation, which is fundamentally different from the synaptic plasticity through which the brain learns. But models that learn effectively through synaptic plasticity are lacking. The funded researchers will use metaplasticity to derive more accurate computational models of learning in the brain. Metaplasticity is an emerging approach in which plasticity rules themselves are learned. In other words, computational models "learn to learn." Recent results show that metaplasticity can approximate any learning rule in theory, and discovers rules that outperform backpropagation at generalizing across tasks in practice. However, existing metaplasticity studies omit features that are critical to learning in the brain and produce models that are too complex to interpret and make experimentally testable predictions. To discover circuit-level principles of biological learning, the funded researchers will apply metaplasticity to models that are biologically realistic enough to capture the fundamental aspects of biological learning and simple enough to interpret the resulting plasticity rules. The researchers will use the resulting models to make predictions that can be tested in experiments. The results from this project will advance our understanding of how the brain learns and improve the effectiveness and efficiency of artificial learning algorithms.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 21, 2022
- Source ID
- FA95502110223XX0
Entities
People
- Robert Rosenbaum
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of Notre Dame