Analysis & Modeling
Jonathan Pillow, Mark Goldman, Carlos Brody,
Mikio Aoi, Brian DePasquale
Working memory, the ability to temporarily hold multiple pieces of information in mind for manipulation, is central to virtually all cognitive abilities. This multi-component research project aims to comprehensively dissect the neural circuit mechanisms of this ability across multiple brain areas. Large population recordings, such as those that will be obtained in other components of this proposal, open the door to assessing the dynamics of brain states on a single-trial, moment-by-moment basis. Yet their size and complexity present a challenge, as does the variety of data that will be collected, incorporating anatomy, behavior, neural activity, and perturbations. This project will develop and apply novel statistical analyses and modeling approaches to meet these challenges. The lion’s share of the variance in neural population activity is often dominated by variations in a small number of variables, which are called “latent variables.” This project will leverage the very large data sets, collected in other components of the project, of many simultaneously recorded neurons to develop advanced linear and nonlinear methods to identify the most informative latent variables. To analyze these datasets, the researchers will develop new latent variable discovery methods. First, they will combine advanced quantitative behavioral analysis with advanced statistical neural analysis. Second, they will combine latent space discovery with fitting of generalized linear models to neural data. The resulting nonlinear methods will provide an unprecedentedly complete statistical description of the data: these methods aim to simultaneously discover and capture the dynamics of the most important latent variables, and to produce a full statistical characterization of the responses of each individual recorded neuron. In biophysical modeling work, critical to creating a mechanistic understanding at the neural circuit level, this project will develop and test models of both local and multi-brain-region activity during working memory and decision-making. These models will build upon rigorous sensitivity-analysis techniques for identifying the critical network interactions underlying observed behavior. The models will be used both to interpret existing data and to design maximally informative experiments about inter-regional network interactions, and they will provide a principled platform from which to design future experiments that test specific hypotheses about function and further constrain the models.