Multitasking Recurrent Neural Networks : Interactions Across Computations in a Single Network

To advance our understanding of the flexibility of biological networks, we study how computations for multiple tasks are implemented within a single network of neurons. We employ a set of 20 tasks for artificial recurrent neural networks and human participants enrolled in a brain-computer interface clinical trial to perform. This project is designed to develop intuitions about how biologically inspired artificial networks perform memory and decision-making computations and test these intuitions in real biological networks. We focus on the neural dynamics, the patterns of neural activity that evolve through time dependent on the current state and the input to the system. We compare RNN dynamics to those recorded from populations of cortical neurons, with single neuron resolution in human participants performing all 20 tasks. Following this novel approach, we will develop models of human cognition and behavior.

Compartmentalized Computation : Modules in the Brain

The brain is composed of interconnected regions that perform specialized functions. Strokes and other brain traumas in addition to fMRI studies in humans and widefield imaging studies in mice have provided an increasingly detailed map of the distinct functions across cortical areas. However, there is a weak theoretical framework for what it means when we say two brain regions ‘interact’. Additionally, it is unclear what advantages are provided for neural networks through compartmentalized processing. In order to develop hypotheses for how modular networks perform computations, we train artificial neural networks with sparse connections between interacting layers of recurrent neural networks.