Failing Loudly: An Empirical Study of Methods for Detecting Dataset Shift
This work addresses the critical issue of dataset shift detection for improving ML system reliability, but it is incremental as it builds on existing methods with empirical comparisons.
The paper tackles the problem of machine learning systems failing silently under dataset shift by empirically studying methods for detection, exemplar identification, and malignancy quantification, showing that a two-sample-testing-based approach with pre-trained classifiers performs best across various perturbations.
We might hope that when faced with unexpected inputs, well-designed software systems would fire off warnings. Machine learning (ML) systems, however, which depend strongly on properties of their inputs (e.g. the i.i.d. assumption), tend to fail silently. This paper explores the problem of building ML systems that fail loudly, investigating methods for detecting dataset shift, identifying exemplars that most typify the shift, and quantifying shift malignancy. We focus on several datasets and various perturbations to both covariates and label distributions with varying magnitudes and fractions of data affected. Interestingly, we show that across the dataset shifts that we explore, a two-sample-testing-based approach, using pre-trained classifiers for dimensionality reduction, performs best. Moreover, we demonstrate that domain-discriminating approaches tend to be helpful for characterizing shifts qualitatively and determining if they are harmful.