SimSMoE: Solving Representational Collapse via Similarity Measure

TL;DR

SimSMoE introduces a novel similarity-based algorithm to prevent representation collapse in sparse mixture of experts models, improving training stability and performance in large language models.

Contribution

The paper proposes SimSMoE, a new similarity measure that guarantees solutions to representation collapse in SMoE models, enhancing their training and scalability.

Findings

SimSMoE outperforms existing SMoE training methods in various tasks.

It effectively addresses the representation collapse problem.

The method demonstrates robustness and scalability across large language models.

Abstract

Sparse mixture of experts (SMoE) have emerged as an effective approach for scaling large language models while keeping a constant computational cost. Regardless of several notable successes of SMoE, effective training such architecture remains elusive due to the representation collapse problem, which in turn harms model performance and causes parameter redundancy. In this work, we present Similarity-based Sparse Mixture of Experts (SimSMoE), a novel similarity of neural network algorithm, that guarantees a solution to address the representation collapse issue between experts given a fixed FLOPs budget. We conduct extensive empirical evaluations on three large language models for both Pre-training and Fine-tuning tasks to illustrate the efficacy, robustness, and scalability of our method. The results demonstrate that SimSMoE significantly enhances existing routing policy and outperforms other SMoE training methods in performance for the tasks.