ModelDiff: A Framework for Comparing Learning Algorithms

TL;DR

ModelDiff is a framework that compares learning algorithms by identifying feature transformations that reveal differences in model predictions, helping understand how algorithms utilize training data.

Contribution

The paper introduces ModelDiff, a novel method leveraging the datamodels framework to compare learning algorithms based on their data usage and prediction differences.

Findings

Effective in distinguishing models trained with different data augmentation

Able to compare models with/without pre-training

Identifies differences due to hyperparameter variations

Abstract

We study the problem of (learning) algorithm comparison, where the goal is to find differences between models trained with two different learning algorithms. We begin by formalizing this goal as one of finding distinguishing feature transformations, i.e., input transformations that change the predictions of models trained with one learning algorithm but not the other. We then present ModelDiff, a method that leverages the datamodels framework (Ilyas et al., 2022) to compare learning algorithms based on how they use their training data. We demonstrate ModelDiff through three case studies, comparing models trained with/without data augmentation, with/without pre-training, and with different SGD hyperparameters. Our code is available at https://github.com/MadryLab/modeldiff .