Robust modeling of within- and across-area population dynamics through the development of a practical recurrent neural network toolkit

This project aims to develop new architectures and training methods for recurrent neural network models of neural population dynamics. We propose to make hierarchical models of the dynamics within and across different areas of the neuroaxis. Miller has extensive data sets that he will provide for this project. These include data collected previously from monkeys performing movement tasks, including intracortical recordings of neural activity from motor and premotor cortices, electromyographic (EMG) recordings, and behavioral and task information. In the first 18 months of the project period, Dr. Miller and his postdoc will also collect additional monkey data, including recordings from motor cortex and EMG during stereotyped lab behaviors, as well as natural behaviors collected wirelessly within the monkey's home cage. All of the costs for this data collection, except additional post-doctoral effort, are supported by existing grants (NS053603, NS104344, NS109257). Dr. Miller and his postdoctoral fellow will also collaborate on the development and testing of a suite of deep learning tools to analyze this data, with a particular focus on multi-area models of neural population dynamics, and whether the information flow estimated to occur between areas corresponds with experimentally-measured variables (e.g., if the information flow between areas is predictive of measured reaction times, changes in attentional state, changes in measured force output, etc). Dr. Miller and his postdoctoral fellow will also contribute to the preparation of manuscripts and presentations related to the work, and in the dissemination of data and tools after publication. These efforts will be coordinated across teams through bi-weekly videoconferences between the Emory and Northwestern teams.