Variational Information Bottleneck for Effective Low-Resource Fine-Tuning

TL;DR

This paper introduces a Variational Information Bottleneck approach to improve low-resource fine-tuning of large language models by reducing overfitting and enhancing out-of-domain generalization.

Contribution

The paper proposes a novel VIB-based method for low-resource fine-tuning that suppresses irrelevant features and improves robustness and transferability.

Findings

Significantly improves transfer learning in low-resource settings

Enhances out-of-domain generalization on NLI benchmarks

Reduces overfitting in low-resource fine-tuning scenarios

Abstract

While large-scale pretrained language models have obtained impressive results when fine-tuned on a wide variety of tasks, they still often suffer from overfitting in low-resource scenarios. Since such models are general-purpose feature extractors, many of these features are inevitably irrelevant for a given target task. We propose to use Variational Information Bottleneck (VIB) to suppress irrelevant features when fine-tuning on low-resource target tasks, and show that our method successfully reduces overfitting. Moreover, we show that our VIB model finds sentence representations that are more robust to biases in natural language inference datasets, and thereby obtains better generalization to out-of-domain datasets. Evaluation on seven low-resource datasets in different tasks shows that our method significantly improves transfer learning in low-resource scenarios, surpassing prior work. Moreover, it improves generalization on 13 out of 15 out-of-domain natural language inference benchmarks. Our code is publicly available in https://github.com/rabeehk/vibert.