Calendar

One of the hottest sub-topics of machine learning in recent times has been Self-Supervised Learning (SSL). In SSL, a learning machine captures the dependencies between input variables, some of which may be observed, denoted X, and others not always observed, denoted Y. SSL pre-training has revolutionized natural language processing and is making very fast progress in speech and image recognition. SSL may enable machines to learn predictive models of the world through observation, and to learn representations of the perceptual world, thereby reducing the number of labeled samples or rewarded trials to learn a downstream task. In the Energy-Based Model framework (EBM), both X and Y are inputs, and the model outputs a scalar energy that measures the degree of incompatibility between X and Y. EBMs are implicit functions that can represent complex and multimodal dependencies between X and Y. EBM architectures belong to two main families: joint embedding architectures and latent-variable generative architectures. There are two main families of methods to train EBMs: contrastive methods, and volume regularization methods. Much of the underlying mathematics of EBM is borrowed from statistical physics, including concepts of partition function, free energy, and variational approximations thereof.

Zoom: https://harvard.zoom.us/j/779283357?pwd=MitXVm1pYUlJVzZqT3lwV2pCT1ZUQT09

Several optimization problems in AI Machine Learning can be solved with the maximization of functions that exhibit natural diminishing returns. Examples include feature selection for Generalized Linear Models, Data Summarization, and Bayesian experimental design. By leveraging diminishing returns, it is possible to design efficient approximation algorithms for these problems.One of the simplest notions of diminishing returns is submodularity. Submodular functions are particularly interesting, because they admit simple, yet non-trivial, polynomial-time approximation algorithms. In recent years, several definitions have been proposed, to generalize the notion of submodularity. A study of these generalized functions lead to the design of efficient approximation algorithms for non-convex problems.In this talk, I will discuss the notion of submodularity, and illustrate relevant results on this topic, including new interesting combinatorial algorithms. I will also talk about generalizations of this notion to continuous domains, and how they translate into first- and second-order conditions. I will discuss how these notions pertain interesting problems in AI Machine Learning.

Zoom: https://harvard.zoom.us/j/98231541450

The continuum formal path integral over Euclidean fermions in the background of a Euclidean gauge field is replaced by the quantum mechanics of an auxiliary system of non-self-interacting fermions. No-go “theorems” are avoided.

The main features of chiral fermions arrived at by formal continuum arguments are preserved on the lattice.

Zoom: https://harvard.zoom.us/j/977347126

The continuum formal path integral over Euclidean fermions in the background of a Euclidean gauge field is replaced by the quantum mechanics of an auxiliary system of non-self-interacting fermions. No-go “theorems” are avoided.

The main features of chiral fermions arrived at by formal continuum arguments are preserved on the lattice.

Zoom: https://harvard.zoom.us/j/977347126

Proteolysis-targeting chimeras (PROTACs) are heterobifunctional small molecules consisting of two chemical moieties connected by a linker.  The simultaneous binding of a PROTAC to both a target protein and an E3 ligase facilitates ubiquitination and degradation of the target protein. Since its proof-of-concept research in 2001, PROTAC has been vigorously developed by both research community and pharma industry, to act against therapeutically significant proteins, such as BRD4, BTK, and STAT3. However, despite the enthusiasm, designing PROTACs is challenging. Till now, no case of de novo rational design of PROTACs has been reported and the successful PROTACs usually came from the functional screen from a limitedly scaled library.  As formation of a ternary complex between the protein target, the PROTAC, and the recruited E3 ligase is considered paramount for successful degradation, several computational algorithms (PRosettaC as the example), have been developed to model this ternary complex, which have got partial agreement with the experimental data and in principle inform future rational PROTAC design. Here I will introduce some of these computational methods and share how they model the ternary complexes.

Zoom: https://harvard.zoom.us/j/98248914765?pwd=Q01tRTVWTVBGT0lXek40VzdxdVVPQT09

(Password: 419419)

Zoom: https://harvard.zoom.us/j/977347126