Zoom Meeting ID: 997 6555 9213
Speaker: Cynthia Rudin, PhD
Professor of Computer Science, Electrical and Computer Engineering, and Statistical Science, Duke University
Interpretability vs. Explainability in Machine Learning
With widespread use of machine learning, there have been serious societal consequences from using black box models for high-stakes decisions, including flawed bail and parole decisions in criminal justice. Explanations for black box models are not reliable, and can be misleading. If we use interpretable machine learning models, they come with their own explanations, which are faithful to what the model actually computes.
In this talk, I will discuss some of the reasons that black boxes with explanations can go wrong, whereas using inherently interpretable models would not have these same problems. I will give an example of where an explanation of a black box model went wrong, namely, I will discuss ProPublica's analysis of the COMPAS model used in the criminal justice system: ProPublica’s explanation of the black box model COMPAS was flawed because it relied on wrong assumptions to identify the race variable as being important. Luckily in recidivism prediction applications, black box models are not needed because inherently interpretable models exist that are just as accurate as COMPAS.
I will also give examples of interpretable models in healthcare. One of these models, the 2HELPS2B score, is actually used in intensive care units in hospitals; most machine learning models cannot be used when the stakes are so high.
Finally, I will discuss two long-term projects my lab is working on, namely optimal sparse decision trees and interpretable neural networks.