Estimating Example Difficulty using Variance of Gradients

Chirag Agarwal, Sara Hooker

Keywords: interpretability, human in the loop learning, atypical examples

Abstract: In machine learning, a question of great interest is understanding what examples are challenging for a model to classify. Identifying atypical examples helps inform safe deployment of models, isolates examples that require further human inspection, and provides interpretability into model behavior. In this work, we propose the Variance of Gradients (VOG) as a valuable and efficient proxy metric for detecting outliers in the data distribution. We provide quantitative and qualitative support that VOG is a meaningful way to rank data by difficulty and to surface a tractable subset of the most challenging examples for human-in-the-loop auditing. Data points with high VOG scores are more difficult for the model to learn and over-index on examples that require memorization.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

One-sentence Summary: The Variance of Gradients (VoG) metric can be used to identify atypical examples from a distribution

Community Implementations: [ 3 code implementations](https://www.catalyzex.com/paper/arxiv:2008.11600/code)

Reviewed Version (pdf): https://openreview.net/references/pdf?id=mZP94FatLt