RobustDebias: Debiasing Language Models using Distributionally Robust Optimization

Pretrained language models have been shown to exhibit biases and social stereotypes. Prior work on debiasing these models has largely focused on modifying embedding spaces during pretraining, which is not scalable for large models. Fine-tuning pretrained models on task-specific datasets can both degrade model performance and amplify biases present in the fine-tuning data. We address bias amplification during fine-tuning rather than costly pretraining, focusing on BERT models due to their widespread use in language understanding tasks. While Empirical Risk Minimization effectively optimizes downstream performance, it often amplifies social biases during fine-tuning. To counter this, we propose \textit{RobustDebias}, a novel mechanism which adapts Distributionally Robust Optimization (DRO) to debias language models during fine-tuning. Our approach debiases models across multiple demographics during MLM fine-tuning and generalizes to any dataset or task. Extensive experiments on various language models show significant bias mitigation with minimal performance impact.

Document Image Cleaning using Budget-Aware Black-Box Approximation

Recent work has shown that by approximating the behaviour of a non-differentiable black-box function using a neural network, the black-box can be integrated into a differentiable training pipeline for end-to-end training. This methodology is termed "differentiable bypass,'' and a successful application of this method involves training a document preprocessor to improve the performance of a black-box OCR engine. However, a good approximation of an OCR engine requires querying it for all samples throughout the training process, which can be computationally and financially expensive. Several zeroth-order optimization (ZO) algorithms have been proposed in black-box attack literature to find adversarial examples for a black-box model by computing its gradient in a query-efficient manner. However, the query complexity and convergence rate of such algorithms makes them infeasible for our problem. In this work, we propose two sample selection algorithms to train an OCR preprocessor with less than 10% of the original system's OCR engine queries, resulting in more than 60% reduction of the total training time without significant loss of accuracy. We also show an improvement of 4% in the word-level accuracy of a commercial OCR engine with only 2.5% of the total queries and a 32x reduction in monetary cost. Further, we propose a simple ranking technique to prune 30% of the document images from the training dataset without affecting the system's performance.