Interpretable Machine Learning: Visualization, Sparse Models, and Neural Networks - Online Course

ABSTRACT

This three-part lecture covers three pillars of interpretable machine learning: dimension reduction for data visualization, sparse models for tabular data, and interpretable neural networks for computer vision. These are essential topics for any researcher working in high-stakes machine learning applications—and genuinely useful ones.

Part 1: Dimension Reduction for Data Visualization

Dimension reduction (DR) for data visualization provides unique insights into the structure of high-dimensional data. DR offers a bird’s-eye view of a dataset, revealing clusters and their relationships, manifolds, branching patterns, and even potential errors in the data. It is extremely effective for scientific discovery and hypothesis generation. We will discuss key elements of DR algorithms leading to the derivation of the PaCMAP algorithm, with applications in bioinformatics, name-ethnicity classification, finance, and neurology.

Part 2: Sparse Models and Rashomon Sets for Tabular Data

While the trend in machine learning has moved toward increasingly complex black-box models, such models have shown no performance advantage for many real-world tabular datasets. For these datasets, simpler models—sometimes small enough to fit on an index card—can be just as accurate and far easier to use. The challenge is that designing interpretable models is difficult due to the “interaction bottleneck,” which arises when domain experts must work closely with machine learning algorithms. We’ll review two families of interpretable models—optimal sparse decision trees and sparse generalized additive models—and introduce the Rashomon set framework as a principled approach to managing the interaction bottleneck, with examples from finance and criminal justice.

Part 3: Interpretable Neural Networks for Computer Vision

Prototype neural networks are among the most popular inherently interpretable architectures for computer vision and signal processing. These models make predictions by comparing parts of an input image to parts of prototypical images, assigning a score to each comparison and summing those scores to form the final prediction. We will discuss the ProtoPNet algorithm and its extension, ProtoConcept, in which a cluster of images defines a “concept prototype,” making comparisons richer and more informative. An application to ICU neurology is also included.

The lecture concludes with an application to computer-aided mammography, in which an interpretable neural network led to a scientific discovery: subtle left-right asymmetries in mammograms can predict breast cancer up to five years in advance. Our AsymMirai algorithm was the fifth most-viewed paper in Radiology in 2024.

See also:

Cynthia Rudin, Chaofan Chen, Zhi Chen, Haiyang Huang, Lesia Semenova, and Chudi Zhong. “Interpretable Machine Learning: Fundamental Principles and 10 Grand Challenges.” Statistics Surveys, 2022. https://arxiv.org/abs/2103.11251

This Distinguished Speaker Series seminar will consist of three hours of lecture and Q&A, held live* via the free video-conferencing software Zoom.

*The video recording of the seminar will be made available to registrants within 24 hours and will be accessible for four weeks thereafter. That means that you can watch all of the class content and discussion even if you cannot participate synchronously.

Closed captioning is available for all live and recorded sessions. Captions can be translated to a variety of languages including Spanish, Korean, and Italian. For more information, click here.

If you’re looking for a comprehensive introduction to interpretable machine learning, check out our livestream seminar Interpretable Machine Learning on June 9-12.