Parameter-Efficient Mixture-of-Experts Architecture for Pre-trained Language Models
This addresses the problem of scaling language models efficiently for researchers and practitioners, though it is incremental as it builds on existing MoE methods.
The paper tackles the high parameter cost of Mixture-of-Experts (MoE) architectures in large language models by proposing a parameter-efficient MoE that shares core parameters among experts using tensor decomposition, achieving a 27.2x reduction in parameters compared to Switch Transformers while maintaining superior performance.
Recently, Mixture-of-Experts (short as MoE) architecture has achieved remarkable success in increasing the model capacity of large-scale language models. However, MoE requires incorporating significantly more parameters than the base model being extended. In this paper, we propose building a parameter-efficient MoE architecture by sharing information among experts. We adopt the matrix product operator (MPO, a tensor decomposition from quantum many-body physics) to reconstruct the parameter matrix in the expert layer and increase model capacity for pre-trained language models by sharing parameters of the central tensor (containing the core information) among different experts while enabling the specificity through the auxiliary tensors (complementing the central tensor) of different experts. To address the unbalanced optimization issue, we further design the gradient mask strategy for the MPO-based MoE architecture. Extensive experiments based on T5 and GPT-2 show improved performance and efficiency of the pre-trained language model (27.2x reduction in total parameters for the superior model performance, compared with the Switch Transformers). Our code is publicly available at https://github.com/RUCAIBox/MPOE.