Towards Robust Low-Resource Fine-Tuning with Multi-View Compressed Representations
This addresses overfitting issues for researchers and practitioners working with low-resource NLP fine-tuning, though it is incremental as it builds on existing fine-tuning methods.
The paper tackles overfitting in low-resource fine-tuning of pretrained language models by inserting random autoencoders between hidden layers to create multi-view compressed representations, resulting in promising performance improvements across various NLP tasks without added inference costs.
Due to the huge amount of parameters, fine-tuning of pretrained language models (PLMs) is prone to overfitting in the low resource scenarios. In this work, we present a novel method that operates on the hidden representations of a PLM to reduce overfitting. During fine-tuning, our method inserts random autoencoders between the hidden layers of a PLM, which transform activations from the previous layers into multi-view compressed representations before feeding them into the upper layers. The autoencoders are plugged out after fine-tuning, so our method does not add extra parameters or increase computation cost during inference. Our method demonstrates promising performance improvement across a wide range of sequence- and token-level low-resource NLP tasks.