Advances in neural recording technology now provide access to neural activity at high temporal resolutions, from many brain areas, and during complex and naturalistic behavior. Interpreting these types of high-dimensional and unconstrained neural recordings is still a major challenge in neuroscience.
The aim of this project is to develop innovative methods for distilling high-dimensional neural activity patterns into simpler low-dimensional formats that can be effectively compared across time, conditions, or even across species. Our team is uniquely positioned to not only develop these novel methods, but also apply them to characterize changes in neural systems across a wide range of clinically-relevant perturbations, including addiction, sensory manipulation, and disease. In this project, theory, methods, and models will be developed for: 1) learning low-dimensional latent space models that align many neural datasets onto a common reference frame for comparison, 2) comparing datasets and testing the impact of a variety of perturbations (e.g. monocular deprivation, addiction and withdrawal) on the shape or geometry of neural activity from its baseline state, and 3) investigating the role of specific cell types and microcircuits on shaping population activity over time, during sleep, and in response to certain classes of perturbations. This project will provide new tools and frameworks for comparing neural datasets, leading to robust measures of disease, signatures of addiction, and other network-level reflections of environment and behavior. Significance: As neural datasets continue to grow in size, new methods for analysis are becoming of utmost importance in driving scientific understanding of the brain. The methods developed in this proposal will identify new ways to learn network-level signatures that allow us to link and compare different neural activity patterns. A robust ability to compare activity across time and animals will have wide reaching impacts, and provide new tools to advance network-level understanding of disease. Innovation: This project will leverage state-of-the-art approaches in high-dimensional statistics and geometry, which are simultaneously advancing in the context of deep learning (DL) architectures, and to tackle challenges in neural coding. This project represents a truly innovative combination of tools in machine learning and computational neuroscience which will likely transfer knowledge in both directions: from machine learning to neuroscience and back. The unique application of advanced mathematical tools in geometry and optimization to population-level analysis of perturbations will be transformative, not only for neuroscience but also in the study of DL architectures.
Interpreting patterns of activity across large networks of neurons is an incredibly challenging problem that remains a key objective of the BRAIN initiative. This project will develop transformative methods for distilling neural activity patterns into compact, informative representations that facilitate comparison across datasets from different brain areas and multiple time-scales, and at different states of neurological disease.
