Research Topic

The overall goal of the present project is to develop a basic computational framework to create multiscale predictive models of naturalistic behavior and brain dynamics by integrating concepts and techniques from nonlinear dynamics, topological data analysis (TDA), and machine learning. In particular, the framework will be applied to video recordings of naturalistic social interaction in humans and animals (ferrets), and simultaneously recorded brain activity (electrophysiology). There is a crucial need for such a computational framework as neuroscience shifts toward the study of naturalistic behavior and how it may be predicted from the underlying brain signals. Traditionally, brain signals are measured in highly controlled experimental conditions, where animals repeatedly perform a simple, well-trained behavioral task, in which case traditional linear statistical models suffice. In naturalistic settings, however, animals frequently switch between distinct types of behavior, where complexity, nonlinearity and non-stationarity prevail, for which traditional methods are not equipped. Thus, new computational frameworks are needed to model such complex neural and behavioral dynamics and to connect them across scales. The proposed project serves as the initial steps to creating such a framework. The proposed project will also provide me the ideal opportunities to gain indepth knowledge of TDA, machine learning, and signal processing in their application to the multiscale modeling of behavior and brain dynamics. The project has three specific aims:

Aim 1: Create topological models of social behavioral dynamics in humans and animals from video recordings. I will utilize existing video recordings of human interaction during psychotherapy sessions and ferret interaction during social play. First, movements of different body segments of humans and ferrets will be extracted as high-dimensional time series data using the DeepLabCut1–3. DeepLabCut is a (deep) transfer learning tool for posture estimation and marker-less motion tracking of multiple animals from regular video recordings for a wide range of species. Second, Temporal Mapper4, a TDA tool based on Mapper5, will be used to created state transition networks from time series extracted in step 1. The transitions networks will be further used to predict future states of the system based on current state.

Aim 2: Create topological models of animal brain dynamics during naturalistic social interaction. I will utilize existing electrophysiology data simultaneously recorded from two interacting ferrets. Electrophysiology time series will be converted to frequency-time representation (wavelet amplitude and phase), a standard procedure for neural time series analysis. Temporal Mapper4 will be used to create transitions networks and state prediction from the frequency-time representations as in Aim 1.

Aim 3: Model cross-scale relations between behavioral and neural states and transitions. Behavioral and neural state transitions networks computed from Aim 1 and 2 will be mapped to each across scales using Gromov-Wasserstein matching for directed graphs6, which belongs to a family of Optimal Transport-based machine learning methods7 and a fast parallel program has been developed for matching state transition networks4. Behavioral states will be used to predict future brain states and vice versa.

Anticipated outcome: Results from each aim will be presented at national and/or international conferences in the domain of computational neuroscience and machine learning and reported in at least one journal or conference publication. The software package produced for each aim will be made publicly available no later than the time of the corresponding publication. Moreover, this project will provide me the valuable experience with the multidisciplinary applications of machine learning, topological data analysis, and signal processing, which will be crucial for my future career in AI and data science.