CartesianMoE: Boosting Knowledge Sharing among Experts via Cartesian Product Routing in Mixture-of-Experts

TL;DR

CartesianMoE introduces a novel routing mechanism for Mixture-of-Experts models that enhances knowledge sharing through a multiplicative approach, leading to improved performance and robustness in large language models.

Contribution

It proposes CartesianMoE, a new knowledge sharing method in MoE models inspired by collective matrix factorization, improving routing robustness and model performance.

Findings

Outperforms previous MoE models in perplexity and downstream tasks.

Achieves better expert routing robustness.

Demonstrates effectiveness in large language model training.

Abstract

Large language models (LLM) have been attracting much attention from the community recently, due to their remarkable performance in all kinds of downstream tasks. According to the well-known scaling law, scaling up a dense LLM enhances its capabilities, but also significantly increases the computational complexity. Mixture-of-Experts (MoE) models address that by allowing the model size to grow without substantially raising training or inference costs. Yet MoE models face challenges regarding knowledge sharing among experts, making their performance somehow sensitive to routing accuracy. To tackle that, previous works introduced shared experts and combined their outputs with those of the top $K$ routed experts in an ``addition'' manner. In this paper, inspired by collective matrix factorization to learn shared knowledge among data, we propose CartesianMoE, which implements more effective knowledge sharing among experts in more like a ``multiplication'' manner. Extensive experimental results indicate that CartesianMoE outperforms previous MoE models for building LLMs, in terms of both perplexity and downstream task performance. And we also find that CartesianMoE achieves better expert routing robustness.