The massive neural activity data collected through transformative new technologies present an incredible opportunity in facilitating the mission of the BRAIN Initiative and the wider neuroscience community: to understand brain function. However, the analysis and interpretation of these massive datasets represent a bottleneck for computational models and theories, which traditionally simulate brain areas as modules in isolation, rather than as interacting, intercommunicating neural circuit ensembles. We propose to eliminate this bottleneck through the development of powerful new, scalable multi-region neural network models and analysis tools that will enable discovery of the connectivity and circuit mechanisms within and between brain regions. We will build off of the advantages of, and our prior successes with, recurrent neural networks (RNNs), the state-of-the-art for modeling time-varying behaviors. Specifically, we propose to develop and test a new class of multi-region ?network of networks? RNNs to model the nested, hierarchical computations occurring across multiple interacting brain regions during complex behavior. These models will provide another advantage over existing models in being trained to match experimental data from the outset, and using the dynamics of learning to reveal effective connectivity, comprising both network structures and communication principles. The entire pipeline?training and testing the multi-region models and analysis steps for real and simulated data?is scalable and modular, enabling our community of collaborators and end-users to reproduce, validate, and even customize it. Finally, the models, methods, and theories developed are aimed as extensible deliverables. They will lead to widely disseminable new extensions for probing the circuit mechanisms of inter-area communication during both functional and dysfunctional states, and bridging together several experimental observations and model systems.
The specific aims of the project include a number of critical subprojects focused on: first, developing multi- region RNN models of adaptive learning, initially focusing on the use case of adversity evasion by zebrafish; and second, modeling circuit dysfunctions in maladaptive learning, the use case being learned immobility in the face of inescapable, persistent adversity. Third, in parallel, this project will scale up and expand the range of the multi-region models and tools for inferring effective connectivity to larger, yet less-sampled datasets, and disseminate them to a community of collaborative end-users. The proposed multi-region RNN models and analytical tools have the potential to be widely used in the neuroscience community to dissect features of brain function in both health and disease. Moreover, our resultant theoretical frameworks will have a strong influence on fundamental approaches to understanding neuroscience data and will inform experimental paradigms and drive future data collection.
Transformative new technologies now capture neural activity from thousands of single neurons across multiple areas and even whole brains with sub-second resolution. These methods are critical facilitators for the BRAIN initiative and the wider neuroscience community. However, they produce large-scale datasets whose analysis and interpretation represent a bottleneck for computational models and theories, which traditionally simulate brain areas as modules in isolation, rather than as interacting, intercommunicating neural circuits. We will develop powerful new, scalable multi-region neural network models and analysis tools for eliminating this bottleneck, opening up new avenues for probing circuit mechanisms within and between brain regions in both health and disease.
