Milo Lin (UT Southwestern): Schmidt AI in Science Speaker Series
Organized by the University of Chicago’s Eric and Wendy Schmidt AI in Science Fellowship Program.
Agenda
4:00pm – 4:45pm: Presentation
4:45pm – 5:00pm: Q&A
5:00pm – 5:30pm: Reception
Meeting location
William Eckhardt Research Center. Room 401
5640 S Ellis Avenue, Chicago, IL 60637
Map It
Title: Distilling rules from data using Essence Neural Networks
Abstract: Despite their increasing impact on science, society, and medicine, existing neural networks are “black boxes” because it is often difficult to understand how the network computes a given answer. This is because such networks are trained in an end-to-end fashion using backpropagation, resulting in neurons having no individual role aside from contributing to the collective task of minimizing the network error. To address this limitation, we developed an alternative hierarchical method to train neural networks called essence neural networks (ENNs). Using limited training data, ENNs can discover algorithms, or the “rules” to explain observed input-output patterns in the data, which can be converted in an automated fashion directly from the neural network into succinct computer code so humans can understand the discovered rules. We also find that ENNs naturally recover connectivity and firing properties of biological neural networks which are not intrinsic to backpropagation-trained networks, allowing direct comparison between the firing statistics of single-neuron recordings of human patients and those of ENNs trained on the same visual task.
Bio: Milo Lin is an Associate Professor in the Lyda Hill Department of Bioinformatics, UT Southwestern, since September 2024. He joined the department as assistant professor in 2021. Since 2015, he is also a member of the Cecil H. and Ida Green Center for Systems Biology. He is the Cecil H. and Ida M. Green Endowed Scholar in Biomedical Computational Science. Milo Lin received his B.S. and Ph.D. in physics from Caltech in 2006 and 2012, respectively. His thesis work, under the mentorship of Ahmed Zewail, was on physical principles of protein folding. He was a Miller Independent Research Fellow, hosted by David Chandler, at UC Berkeley from 2012-2015.
