NanoMoE: Scaling Mixture of Experts to Individual Layers for Parameter-Efficient Deep Learning
Keywords: Mixture of Experts, Parameter Efficiency, Expressivity, Low-Rank Factorization
Abstract: Large language models (LLMs) have achieved remarkable success, but their growing size leads to significant challenges in efficiency and cost. This work explores parameter-efficient deep learning, aiming to achieve comparable performance with fewer parameters and floating-point operations (FLOPs). We introduce NanoMoE, a novel family of parameter-efficient building blocks inspired by the Mixture of Experts (MoE) framework. NanoMoE offers a modular and efficient replacement for fully connected layers within traditional neural networks. We instantiate NanoMoE with three variants of increasing complexity and theoretically demonstrate its superior expressivity compared to low-rank factorization with minimal parameter increase. Empirical results validate that NanoMoE achieves superior model quality compared to low-rank factorization under the same parameter or FLOP budget, confirming its enhanced efficiency.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 10906
Loading