A Survey on Dynamic Neural Networks for Natural Language Processing

TL;DR

<1-2> The paper surveys dynamic neural networks for NLP, focusing on skimming, mixture of experts, and early exit to achieve efficient, input-dependent computation. It provides a taxonomy, surveys major methods (learned/unlearnable MoE routing, confidence- and ensemble-based exits, cascading), and analyzes challenges in evaluation, parallelism, and runtime optimization. The authors identify practical guidance and open questions for future work, including theoretical grounding and explainability, to enable scalable, efficient NLP on very large models and on resource-constrained devices. Overall, the work clarifies how dynamic computation can complement static model compression to push toward trillion-parameter models with feasible inference costs.

Abstract

Effectively scaling large Transformer models is a main driver of recent advances in natural language processing. Dynamic neural networks, as an emerging research direction, are capable of scaling up neural networks with sub-linear increases in computation and time by dynamically adjusting their computational path based on the input. Dynamic neural networks could be a promising solution to the growing parameter numbers of pretrained language models, allowing both model pretraining with trillions of parameters and faster inference on mobile devices. In this survey, we summarize progress of three types of dynamic neural networks in NLP: skimming, mixture of experts, and early exit. We also highlight current challenges in dynamic neural networks and directions for future research.

Figures (2)

• Figure 1: The three types of dynamic neural networks summarized in this paper. They dynamically adjust computation timewise, widthwise and depthwise, respectively.
• Figure 2: Taxonomy of dynamic neural networks for NLP.