Sparse Mixers: Combining MoE and Mixing to build a more efficient BERT

TL;DR

This paper introduces Sparse Mixer, a novel encoder combining MoE and mixing transformations, achieving faster training and inference while maintaining competitive performance on NLP benchmarks.

Contribution

The paper presents Sparse Mixer, a new model that integrates MoE with mixing transformations, improving efficiency and stability over traditional MoE models.

Findings

Sparse Mixer outperforms BERT on GLUE and SuperGLUE by less than 1%.

Sparse Mixer trains 65% faster and runs inference 61% faster than BERT.

Fast Sparse Mixer trains and runs nearly twice as fast as BERT, with slight performance trade-offs.

Abstract

We combine the capacity of sparsely gated Mixture-of-Experts (MoE) with the speed and stability of linear, mixing transformations to design the Sparse Mixer encoder model. Sparse Mixer slightly outperforms (<1%) BERT on GLUE and SuperGLUE, but more importantly trains 65% faster and runs inference 61% faster. We also present a faster variant, prosaically named Fast Sparse Mixer, that marginally underperforms BERT on SuperGLUE, but trains and runs nearly twice as fast. We justify the design of these two models by carefully ablating through various mixing mechanisms, MoE configurations and hyperparameters. Sparse Mixer overcomes many of the latency and stability concerns of MoE models and offers the prospect of serving sparse student models, without resorting to distilling them to dense variants.