Continuous Statistical Models for Modern Computational Neuroscience
Presented by William Consagra, Harvard Medical School
Monday, January 29 2024
3:30 PM-4:30 PM ET
AUST 105
Webex Meeting Link
Coffee will be served at 3:00 pm in the Noether Lounge (AUST 326)
Recent advances in neuroimaging technologies hold great promise to bolster our understanding of the structure and connectivity of the human brain, its heterogeneity across individuals, and its relationship to various traits and diseases. However, the scale and complexity of modern neuroimaging datasets pose challenges for traditional modeling approaches, which often work with discretized representations of the imaging data, resulting in poor computational and statistical performance for high-resolution images. In this talk, I will discuss two novel methodologies that employ fully continuous models of the neuroimaging data to address fundamental tasks in computational neuroscience.
In the first part of the talk, I present a deep-learning-based framework for inferring brain structure and connectivity from sparse and noisy MRI images. This approach proposes a novel, partially stochastic neural network to form a continuous parameterization of the brain's spatially varying diffusion field, enabling fast inference at any spatial location. Using both simulated data and state-of-the-art high-resolution in-vivo data, the proposed method is demonstrated to outperform standard discrete approaches for a variety of reconstruction and image uncertainty quantification tasks.
In the second part of the talk, I introduce a novel continuous model for brain connectivity using latent functions defined on the symmetric product manifold of the brain's surface. A new algorithm is developed to estimate the model parameters and is designed to be scalable to arbitrarily fine surface discretizations. Using real data from the Human Connectome Project, I demonstrate the superiority of the proposed method over state-of-the-art discrete network-based approaches on important neuroscience tasks in trait prediction, hypothesis testing, and subnetwork discovery.
Speaker Bio:
Will is a postdoctoral research fellow at Harvard Medical School. He received his PhD in statistics from the University of Rochester in 2022. His research interests include functional data analysis, computational statistics, and deep learning, with applications to neuroimaging inverse problems and connectomics.