Decades of scientific efforts have aimed to understand computations in the brain, and resulting complex behaviors. Now, with advancements in experimental techniques, we have high-resolution neural and behavioral datasets. Consequently, there is a growing need for tools capable of effectively analyzing such datasets, with machine learning playing a pivotal role in modeling approaches within neuroscience. Nevertheless, the application of machine learning to neuroscientific datasets faces two significant hurdles : the high cost and tedious nature of data collection compared to mainstream machine learning, and domain-specific requirements that come with analyzing neuroscientific datasets. My thesis focuses on overcoming these hurdles, enhancing the power and flexibility of machine learning tools for neuroscience.
In the first part of my talk, I'll delve into overcoming the challenge of limited and costly datasets by introducing Bayesian approaches that leverage small datasets. We developed a dimension reduction method for analyzing high-dimensional, low-sample size neural data, and an active learning approach to minimize data needed for characterizing animal behavior. Shifting to the second half, I'll discuss incorporating domain-specific requirements into statistical approaches for neuroscience. We developed a tailored inverse reinforcement learning for characterizing animal behavior in complex environments. Finally, we're actively working on incorporating biological constraints into dynamical systems, aiming to unravel the distinct roles of various brain regions and cell types in decision-making tasks.
Adviser: Jonathan Pillow
Zoom Link: https://princeton.zoom.us/j/97838043218