Selective classification using a robust meta-learning approach
This work addresses the problem of robust uncertainty estimation for building reliable models in applications like selective classification and domain adaptation, offering incremental improvements across various real-world datasets.
The paper tackles predictive uncertainty in machine learning models by proposing a novel instance-conditioned reweighting approach using a meta-learning framework, resulting in significant gains such as up to 3.4% accuracy and 3.3% AUC improvements over state-of-the-art methods in selective classification for diabetic retinopathy.
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.