An Empirical Study on the Transferability of Transformer Modules in Parameter-Efficient Fine-Tuning

TL;DR

This study empirically examines the transferability of individual transformer modules in BERT during parameter-efficient fine-tuning, revealing that certain modules like LayerNorms can effectively transfer knowledge with minimal trainable parameters.

Contribution

It demonstrates that each transformer module in BERT can serve as a winning ticket for fine-tuning, with LayerNorms showing exceptional transferability using only 0.003% of parameters.

Findings

LayerNorm modules transfer knowledge effectively with minimal parameters.

Fine-tuning individual modules can achieve performance comparable to full model fine-tuning.

High-magnitude weights in LayerNorms contribute to their transferability.

Abstract

Parameter-efficient fine-tuning approaches have recently garnered a lot of attention. Having considerably lower number of trainable weights, these methods can bring about scalability and computational effectiveness. In this paper, we look for optimal sub-networks and investigate the capability of different transformer modules in transferring knowledge from a pre-trained model to a downstream task. Our empirical results suggest that every transformer module in BERT can act as a winning ticket: fine-tuning each specific module while keeping the rest of the network frozen can lead to comparable performance to the full fine-tuning. Among different modules, LayerNorms exhibit the best capacity for knowledge transfer with limited trainable weights, to the extent that, with only 0.003% of all parameters in the layer-wise analysis, they show acceptable performance on various target tasks. On the reasons behind their effectiveness, we argue that their notable performance could be attributed to their high-magnitude weights compared to that of the other modules in the pre-trained BERT.