SHARCS: Efficient Transformers Through Routing with Dynamic Width Sub-networks

Mohammadreza Salehi, Sachin Mehta, Aditya Kusupati, Ali Farhadi, Hannaneh Hajishirzi

Submission Type: Regular Short Paper

Submission Track: Efficient Methods for NLP

Submission Track 2: Machine Learning for NLP

Keywords: Efficiency, Routing, hardness

TL;DR: Our method makes inference more efficient by learning a router on any transformer network, enabling the network to direct different samples to sub-networks with varying widths.

Abstract: We introduce SHARCS for adaptive inference that takes into account the hardness of input samples. SHARCS can train a router on any transformer network, enabling the model to direct different samples to sub-networks with varying widths. Our experiments demonstrate that: (1) SHARCS outperforms or complements existing per-sample adaptive inference methods across various classification tasks in terms of accuracy vs. FLOPs; (2) SHARCS generalizes across different architectures and can be even applied to compressed and efficient transformer encoders to further improve their efficiency; (3) SHARCS can provide a 2 times inference speed up at an insignificant drop in accuracy.

Submission Number: 3729