Adversarial Filters of Dataset Biases

TLDR

Large neural models excel on standard benchmarks yet falter on adversarial or out‑of‑distribution samples, suggesting they may overfit to dataset biases rather than learning the underlying task. This study investigates AFLite, an adversarial filtering method designed to reduce such biases and thereby curb the overestimation of machine performance. We provide a theoretical analysis that situates AFLite within a generalized framework for optimum bias reduction. AFLite demonstrably lowers measurable dataset biases, improves out‑of‑distribution generalization, causes a marked drop in model accuracy (e.g., from 92 % to 62 % on SNLI) while human performance remains high, and creates upgraded benchmarks that pose new challenges for robust generalization.

Abstract

Large neural models have demonstrated human-level performance on language and vision benchmarks, while their performance degrades considerably on adversarial or out-of-distribution samples. This raises the question of whether these models have learned to solve a dataset rather than the underlying task by overfitting to spurious dataset biases. We investigate one recently proposed approach, AFLite, which adversarially filters such dataset biases, as a means to mitigate the prevalent overestimation of machine performance. We provide a theoretical understanding for AFLite, by situating it in the generalized framework for optimum bias reduction. We present extensive supporting evidence that AFLite is broadly applicable for reduction of measurable dataset biases, and that models trained on the filtered datasets yield better generalization to out-of-distribution tasks. Finally, filtering results in a large drop in model performance (e.g., from 92% to 62% for SNLI), while human performance still remains high. Our work thus shows that such filtered datasets can pose new research challenges for robust generalization by serving as upgraded benchmarks.