CCN talk: March 22, 2024
Ali Yousefi
Assistant Professor, Department of Computer Science & Department of Neuroscience, Worcester Polytechnic Institute
Co-sponsored by the Breaking the Neural Code Cluster at Dartmouth
Advances in latent variable modeling for decoding cognitive states from multimodal data
Abstract:
In this talk, I will present two distinct modeling frameworks built upon the latent variable model to infer underlying mental or cognitive states from high-dimensional neural recordings, behavioral data (including reaction time and decision-making), and physiological data (such as skin conductance and heartbeat). The first model characterizes the connection between neural and behavioral data through latent dynamical variables representing the underlying cognitive states. I will then showcase an application of this framework by inferring the underlying cognitive flexibility during a multi-source inference task and demonstrate how this inference can guide the control policy of deep brain stimulation. The second model integrates Gaussian Processes (GPs) with latent variables to generate a low-dimensional manifold capturing essential neural features. This model simultaneously encodes neural data, categorical stimulus, and physiological data into a shared latent space. This shared latent variable can then decode any of these information types from the neural data. I will show an application of this framework by predicting stimulus categories in a Verbal Memory task, where its prediction accuracy outperforms state-of-the-art decoder models. The promising performance of both models underscores the significance of well-crafted machine learning techniques in decoding brain function, which has wide-ranging applications in both clinical and basic neuroscience.