COMPUTING WITH CONTROLLABLE NEURO-GLIAL NETWORKS
Abstract
The aim of the project is to quantify the role that neural, astrocytic, and actin dynamics play in the processing and representation of stimuli presented to brain networks, and to apply these insights to advanced artificial intelligence (AI) / machine learning (ML) algorithms. We propose to test whether this multiscale and multimodal activity contributes to the representational capacity of brain networks when considered as reservoir computers. To do this we will quantify the amount of information in neural and glial, and cytoskeletal activity separately by examining the degree to which various decoding schemes capture distinct information in the two neural populations during a predictive coding task, and how the dynamics of the biomechanical system affects behavior. To do this we will drive neural populations containing varying concentrations of astrocytes optogenetically with input generated from an evolving system of dynamical equations. We will capture neural and glial activity with calcium and/or voltage-based imaging, and in some studies cytoskeletal activity with actin-GFP. The effect of astrocytes on information encoding will be determined by varying the size of astrocytic populations in the preparation. The role of biomechanics will be determined by arresting the biomechanical actin network. The predictive power of both linear models and biologically inspired neural network models will be assessed based on the experimental dynamical system. As a first step, we will use mutual information metrics between neural, glial, and predicted output activity to measure the representational capacity of both neural and glial activity (alone and in combination), and the impact of biomechanical activity on the performance of the system. Overall Significance: Demonstrating that astrocytes, and actin dynamics are intimately involved in representing information about dynamical stimuli and that they play a role in shaping the performance of the global brain network during learning will revolutionize our understanding of neural information processing. It may suggest new targets for memory performance enhancement. Finally it may lead to the production of novel neuro-inspired machine learning algorithms with improved generalization performance and fewer learning epochs.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 20, 2023
- Source ID
- FA95502210405
Entities
People
- Wolfgang Losert
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of Maryland