Adaptive Routing of Experts via Fuzzy Rule Interpolation for Efficient Multimodal Learning

Bowen Shi, Huachen Qi, Ruiyu Zhuo, Fei Chao, Xiang Chang, Changjing Shang, Qiang Shen

Keywords: Multimodal Learning, Mixer-of-Experts, Fuzzy Rule Interpolation

TL;DR: We use fuzzy logic to make AI routing decisions smarter—dynamically skipping computational layers based on task difficulty rather than using fixed skip rates, achieving 67% training efficiency gains while maintaining accuracy in multimodal AI models.

Abstract:

Multimodal learning integrates heterogeneous data such as text, images, and audio, but existing Mixture of Experts (MoE) frameworks still face two critical limitations: (1) fixed skip rates that fail to adapt to task difficulty, and (2) inefficient expert selection guided by static priors. These issues hinder scalability and stability in large-scale scenarios with diverse semantic patterns. To address these limitations, we propose a Fuzzy Router that incorporates fuzzy rule interpolation (FRI) into Routing of Experts (RoE) for adaptive and efficient expert selection. A sparse fuzzy rule base, derived from prior knowledge and expert experience, is expanded via interpolation to enable nuanced routing decisions based on task complexity. This design dynamically adjusts skip rates, reduces redundant computation, and maintains accuracy through adaptive refinement. Experiments across multiple multimodal benchmarks demonstrate that our method reduces routing data and time by more than 68\%, shortens overall training time by 16.7\%, and preserves competitive performance. These results highlight FRI as a principled mechanism for adaptive resource allocation, advancing efficient and scalable multimodal MoE/RoE systems.
Our code is open-sourced in the supplementary materials.

Primary Area: neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)

Submission Number: 9085