Understanding new tasks through the lens of training data via exponential tilting

TL;DR

This paper introduces a method to reweight training data using exponential tilting to better understand and adapt to new target tasks, improving model deployment in different contexts.

Contribution

It proposes a novel exponential tilt-based reweighting approach to estimate target task distribution from training data, aiding in performance evaluation and model adaptation.

Findings

Effective reweighting on Waterbirds benchmark

Improved target performance estimation

Facilitates model fine-tuning and selection

Abstract

Deploying machine learning models to new tasks is a major challenge despite the large size of the modern training datasets. However, it is conceivable that the training data can be reweighted to be more representative of the new (target) task. We consider the problem of reweighing the training samples to gain insights into the distribution of the target task. Specifically, we formulate a distribution shift model based on the exponential tilt assumption and learn train data importance weights minimizing the KL divergence between labeled train and unlabeled target datasets. The learned train data weights can then be used for downstream tasks such as target performance evaluation, fine-tuning, and model selection. We demonstrate the efficacy of our method on Waterbirds and Breeds benchmarks.