Deep-learning-enabled probabilistic modeling for knowledge discovery

Pre-FPO Presentation
Date
Mar 6, 2023, 11:00 am12:30 pm
Location
CSML 105 & Zoom see abstract

Speaker

Details

Event Description

Probabilistic models, a principled way for flexibly modeling data distributions in the real world, have enabled ML-guided scientific knowledge exploration and discovery, such as 1) sequential experiment design/optimization using a probabilistic surrogate model (usually a Gaussian process), and 2) generating new designs using a generative model. In this talk, I will talk about three works that aim to improve the scalability and interpretability of probabilistic ML models under the setting of scientific discovery. First, I will show how to scale up Gaussian process hyperparameters identification (often the dominating cost of using GPs) by training a single neural network to amortize this expensive procedure across different tasks. Next, I will present how to find sparse solutions in Bayesian optimization to enhance solution interpretability and simplicity. Lastly, I will present ongoing work about a new class of discrete generative models by learning a neural network to approximate the marginal probability density for discrete objects.

 

Co-Advisers: Peter Ramadge & Ryan Adams

 

Zooml Link: https://princeton.zoom.us/my/sulinliu?pwd=V0Nld2NSZmEwdlo3SEI1MXVjQnQ1Z…