Selective classification using a robust meta-learning approach

TL;DR

This paper introduces a meta-learning approach for selective classification that effectively captures predictive uncertainty, leading to improved performance across various real-world tasks and datasets.

Contribution

It proposes a novel instance-conditioned reweighting method with a meta-objective to better model uncertainty, unifying training and test-time applications.

Findings

Effective uncertainty capture across diverse notions of uncertainty.

Significant accuracy and AUC improvements in real-world datasets.

Enhanced performance over state-of-the-art methods and pretrained models.

Abstract

Predictive uncertainty-a model's self awareness regarding its accuracy on an input-is key for both building robust models via training interventions and for test-time applications such as selective classification. We propose a novel instance-conditioned reweighting approach that captures predictive uncertainty using an auxiliary network and unifies these train- and test-time applications. The auxiliary network is trained using a meta-objective in a bilevel optimization framework. A key contribution of our proposal is the meta-objective of minimizing the dropout variance, an approximation of Bayesian Predictive uncertainty. We show in controlled experiments that we effectively capture the diverse specific notions of uncertainty through this meta-objective, while previous approaches only capture certain aspects. These results translate to significant gains in real-world settings-selective classification, label noise, domain adaptation, calibration-and across datasets-Imagenet, Cifar100, diabetic retinopathy, Camelyon, WILDs, Imagenet-C,-A,-R, Clothing1M, etc. For Diabetic Retinopathy, we see upto 3.4%/3.3% accuracy and AUC gains over SOTA in selective classification. We also improve upon large-scale pretrained models such as PLEX.