Illuminating protein space with generative models @SimonsInstituteTOC

Simons Institute | Illuminating protein space with generative models @SimonsInstituteTOC | Uploaded 3 months ago | Updated 20 hours ago
John Ingraham (Generate Biomedicines)
https://simons.berkeley.edu/talks/john-ingraham-generate-biomedicines-2024-06-11
AI≡Science: Strengthening the Bond Between the Sciences and Artificial Intelligence

Proteins are the dominant functional molecules on earth, and yet our ability to leverage them to perform new functions that would be useful to people has largely relied on copying and paraphrasing nature. What does it take to build learning systems that can generalize to new parts of protein space? Amidst the flurry of activity in applying generative modeling to protein design in recent years, I will share some of our own experiences with building learning systems that can generalize, scale, and be programmed to build fit-for-purpose protein complexes on demand.

Sparsification for communication-efficient distributed symmetry-breaking

Understanding Generalization from Pre-training Loss to Downstream Tasks

Circuit Minimization with QBF and SAT-Based Exact Synthesis

Sublinear time algorithms for better than 1/2 approximation algorithms for max-cut on expanders

Counterexample-Guided Inference of Modular Specifications

Machine learning models of differential gene expression

Computing a fixed point of contraction maps in polynomial queries

Bioacoustics and Machine Learning as Key Tools in Conservation

$Hypothesis selection with computational constraints Maryam Aliakbarpour (Rice University) https://simons.berkeley.edu/talks/maryam-aliakbarpour-rice-university-2024-06-21 Extroverted Sublinear Algorithms Hypothesis selection is a fundamental problem in learning theory and statistics. Given a dataset and a finite set of candidate distributions, the goal is to select a distribution that matches the data as well as possible. More specifically, suppose we have sample access to an unknown distribution $P$ over a domain $mathcal{X}$. We assume that $P$ is well-approximated by one of $n$ distributions in a class (called hypotheses), denoted $mathcal{H} coloneqq {H_1, H_2, ldots, H_n}$. The goal is to design an algorithm that outputs a distribution $hat{H} in mathcal{H}$ whose total variation distance from $P$ is nearly minimal. In this work, we study the hypothesis selection problem under memory and time constraints. In particular, we discuss how to achieve a nearly optimal tradeoff between memory usage and sample complexity. Additionally, we explore methods to obtain optimal accuracy via algorithms with nearly optimal time complexity.$