Distribution Shift and Geometric Sensitivity Decomposition

Motivation

In conventional machine learning, the test distribution is often assumed to be the same as the training distribution. However, this assumption has become less justified in the era of deep learning. Deep learning models have been increasingly deployed to the real world and training paradigms have evolved beyond supervised learning. In the self-driving domain, whether a vision system can detect distribution shift is critical to safety; in semi-supervised learning, correct categorization of unlabeled data, not necessarily from the same training distribution, is important for effective leverage of large-scale data. Understanding the causes of distribution shift and mechanisms by which it affects deep learning models is a fundamental research topic.

About this project

We present ideas from two papers: Exploring Covariate and Concept Shift for Out-of-Distribution Detection and Calibration (tian21explore) and A Geometric Perspective towards Neural Calibration via Sensitivity Decomposition (tian21gsd). The former is accepted to the NeurIPS 2021 workshop on Distribution Shift and the latter is accepted as a spotlight paper in Neurips2021.

These two works establish a potential new direction to study the effects of distribution shifts on the feature space learned by deep learning models. Specifically, we approach it from two levels: representation level and modeling level. At the representation level, we derive score functions that output scalar score functions capturing the severity of distribution shifts (tian21explore). At the modeling level, we propose a theory (geometric sensitivity decomposition) to improve the sensitivity of the score functions to distribution shifts (tian21gsd). Last, we incorporate the proposed theory in training (parametrized training) and demonstrate improved performance on calibration and out-of-distribution (OOD) detection.

Resources

1. Link to Github

2. Posters (coming soon)

3. Video Presentations (coming soon)

4. Contact: Junjiao Tian (jtian73@gatech.edu)

Acknowledgment

The work is supported primarily by ONR grant N00014-18-1-2829. Exploring Covariate and Concept Shift for Out-of-Distribution Detection and Calibration was partially developed during Junjiao's internship at Samsung Research America.