Efficient Transformers with Dynamic Token Pooling

TL;DR

This paper introduces a dynamic token pooling mechanism for Transformers that predicts segment boundaries autoregressively, improving efficiency and accuracy across multiple languages and datasets.

Contribution

It proposes a novel dynamic pooling method that adaptively segments input tokens, enhancing Transformer performance and efficiency over fixed-length pooling methods.

Findings

Dynamic pooling improves speed and accuracy.

Joint segmentation and modeling outperform fixed pooling.

Effective across diverse languages and datasets.

Abstract

Transformers achieve unrivalled performance in modelling language, but remain inefficient in terms of memory and time complexity. A possible remedy is to reduce the sequence length in the intermediate layers by pooling fixed-length segments of tokens. Nevertheless, natural units of meaning, such as words or phrases, display varying sizes. To address this mismatch, we equip language models with a dynamic-pooling mechanism, which predicts segment boundaries in an autoregressive fashion. We compare several methods to infer boundaries, including end-to-end learning through stochastic re-parameterisation, supervised learning (based on segmentations from subword tokenizers or spikes in conditional entropy), as well as linguistically motivated boundaries. We perform character-level evaluation on texts from multiple datasets and morphologically diverse languages. The results demonstrate that dynamic pooling, which jointly segments and models language, is both faster and more accurate than vanilla Transformers and fixed-length pooling within the same computational budget.