Large Product Key Memory for Pretrained Language Models

TL;DR

This paper explores integrating large product key memory into pretrained language models to improve their capacity and performance across various NLP tasks, addressing memory utilization issues during training.

Contribution

It introduces a new memory usage metric, proposes initialization and augmentation strategies, and demonstrates improved pretraining and downstream task performance for PKM-augmented PLMs.

Findings

Memory slots often remain outdated during training.

Initialization from memory-free pretrained weights improves performance.

Augmenting rather than replacing feed-forward networks enhances memory utilization.

Abstract

Product key memory (PKM) proposed by Lample et al. (2019) enables to improve prediction accuracy by increasing model capacity efficiently with insignificant computational overhead. However, their empirical application is only limited to causal language modeling. Motivated by the recent success of pretrained language models (PLMs), we investigate how to incorporate large PKM into PLMs that can be finetuned for a wide variety of downstream NLP tasks. We define a new memory usage metric, and careful observation using this metric reveals that most memory slots remain outdated during the training of PKM-augmented models. To train better PLMs by tackling this issue, we propose simple but effective solutions: (1) initialization from the model weights pretrained without memory and (2) augmenting PKM by addition rather than replacing a feed-forward network. We verify that both of them are crucial for the pretraining of PKM-augmented PLMs, enhancing memory utilization and downstream performance. Code and pretrained weights are available at https://github.com/clovaai/pkm-transformers.