CAREER: Developing Neural Network Theory for Uncovering How the Brain Learns

Harvard University, Cambridge MA

Investigators

Abstract

Despite many recent advances enabling the collection of large-scale data on the brain's activity and connectivity, our ability to extract principles from such data of how the brain learns is still limited. This shortfall arises from the absence of a thoroughly developed and predictive theory that elucidates and models learning in the brain at the neural level. To address this gap, this project will develop new theoretical frameworks and mathematical models to help formulate experimentally testable hypotheses about how the brain's neural networks learn. These frameworks will address how data are represented in the brain and how these representations are learned through synaptic plasticity. They will further probe why existing neural network models of the brain lag behind the artificial neural networks that empower AI systems in certain tasks. Results of this project will enhance our understanding of brain function and will be integrated into in education and outreach efforts at the high school, college, graduate and post-graduate levels, including in programs aimed at groups historically under-represented in STEM fields. The project will follow three research thrusts. The first thrust will develop novel theory to elucidate signatures of learning rules and inductive biases in neuronal representations. Experimental techniques allow recording activities of tens or even hundreds of thousands of neurons in the brain. This thrust will help interpret these datasets from a functional point of view. The second thrust will develop a normative theory of biologically plausible learning rules. The investigator's previous work showed that Hebbian learning, despite being local, can implement exact gradient learning on a class of similarity matching cost functions. The project will exploit this finding to design new cost functions for object recognition as manifold disentangling, build corresponding Hebbian neural networks, and compare their learned representations to publicly available neural data from the visual cortex. The last thrust will address learning temporal sequences in recurrent neural networks. It will quantify the temporal sequence learning capabilities of spike-time dependent plasticity. It will study the robustness of neural network trajectories to noise, a key feature of sequential neuronal dynamics in the brain. Finally, the investigator will look for ways of improving sequence learning capacity through nonlinear synaptic interactions. 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.