2022 Program for Women and Mathematics: The Mathematics of Machine Learning

Jessica Gronsbell, University of Toronto
11:45 am - 12:05 pm 
Title: Statistical learning with high volume, high noise health data 
Abstract: In this talk, I will give a brief overview of my background and my path to statistics and machine learning.  I will also discuss my current research on the development of statistical learning methods for large observational health data sets such as electronic health records.  I will emphasize the importance of knowing your data and embracing its complex structure to make analysis more trustworthy and efficient.

Rachel Freedman, University of California, Berkeley

12:05 pm - 12:25 pm 
Title: Objective Misspecification and Value Alignment
Abstract: As machine learning (ML) systems become increasingly capable and general, it becomes increasingly important to ensure that their behavior aligns with human values. However, specifying reward functions for ML agents that operate in environments without a natural reward signal can be challenging, and incorrectly specified rewards can incentivise degenerate or dangerous behavior.  A promising alternative to manually specifying reward functions is to enable ML systems to infer them from human interaction and feedback, but this is sensitive to incorrectly specified models of human decision-making. In this talk, I'll discuss the objective misspecification problem and my work to improve reward inference in the face of human model misspecification.

Harlin Lee, University of California, Los Angeles
12:25 pm - 12:45 pm
Title: Identifying future connections in machine learning topics

Abstract: The field of artificial intelligence (AI) and machine learning (ML) is exponentially growing with no sign of waning popularity. As the discipline gets larger, it is impossible for an individual researcher to be familiar with the entire body of literature. This forces them to specialize in a sub-field. Such insulation can hinder the birth of ideas that arise from new connections, eventually slowing down scientific progress. As such, discovering fruitful interdisciplinary connections by analyzing scientific publications is an important problem in the science of science, or metascience. We present dynamic-embedding-based methods for link prediction in a ML semantic network, where the nodes are concepts in machine learning, and the time-stamped edges indicate co-occurrence in scientific papers.