Monday, November 14, 2022, 1pm

As machine learning models are more frequently deployed in applications with greater impact, there is a growing need to better understand and regulate their behaviors. Explainable machine learning is a research field dedicated to this need, whose primary focus initially has been developing new methods with favorable algorithmic properties that elicit important information about the model predictions. However, as critiques have emphasized the need for more critical evaluations based on concrete tasks, many recent studies highlight an end-to-end workflow that captures both method development and application. This proposal presents our contributions within the workflow for different methods and applications.

In Person and Zoom Participation. See announcement.

We first present new algorithms providing important information about the model behavior through influential training data points and the model's trade-offs with performance and fairness measures. We then discuss an evaluation framework that tests if the existing saliency methods on images are suitable for practical tasks like spurious correlation detection. Since the task relies on the ability of the methods to correctly highlight important input regions for the model, we test this correctness under various conditions. Lastly, motivated by practical issues in academic peer review, we present an ongoing work that evaluates the utility of new and existing methods in helping human users complete a document matching task. We discuss initial results and propose additional steps to address more general problem settings in document matching.

Thesis Committee:
Ameet Talwalkar (Chair)
Nihar Shah
Adam Perer
Chenhao Tan (University of Chicago)

Event Type: Thesis Proposals
Room Number: In Person and Virtual - ET
Building: Gates Hillman 8102 and Zoom
Speaker's Name: JOON SIK KIM
Speaker Website: wnstlr.github.io
Speaker's Professional Title: Ph.D. Student, Machine Learning Department, Carnegie Mellon University
Talk Title: Perspectives on Methods and Applications for Explainable Machine Learning
For More Information: stidle@andrew.cmu.edu
Affiliations: Machine Learning Department (MLD)
Organization(s): SCS