“Harnessing machine learning to model biological systems”

Abstract:
Classically, models of biological systems follow two different approaches. In bottom-up approaches, biological data are used to constrain a phenomenological model of the system in question, and the model is the studied to identify potential functions or potential consequences of the observations that flow into the model. Top-down approaches, on the other hand, start with a presumed function and ask how this question could be implemented in a biologically inspired architecture. Both approaches have been very successful, but both suffer from their own kind of problems. Bottom-up approaches often suffer from (potentially many) parameters that cannot be sufficiently constrained from data. Top-down approaches were in the past hard to combine with the complexities of the biological system in question. Recent advances in machine learning (ML) software now offer a promising hybrid approach, because they allow to optimize not only neural nets, but basically any dynamical system by gradient descent. I will offer a few examples how we have used ML to study biological systems, ranging from behavioral level (nature vs. nurture) down to the level of neural circuits (role of feedback for invariant sensory processing, and, time permitting, the function of different cell classes).

***Want to attend one of our events? Sign up here.
To get regular updates, subscribe to our mailing list from this page.
The Zoom Link will be sent the day before the lecture.