Hierarchical Mixture-of-Experts for Multi-Task Sensor Analytics with Automatic Task Routing

Abstract

The proliferation of heterogeneous sensor networks in industrial and consumer environments has necessitated the development of advanced machine learning architectures capable of processing multi-modal data streams simultaneously. While Multi-Task Learning (MTL) offers a promising paradigm for leveraging shared representations across distinct sensing objectives, traditional hard-parameter sharing often suffers from negative transfer and optimization conflicts. Furthermore, the deployment of dense deep learning models on resource-constrained edge devices presents significant latency and energy challenges. This paper introduces a novel framework: Hierarchical Mixture-of-Experts for Multi-Task Sensor Analytics with Automatic Task Routing (HMoE-ATR). By structuring experts in a hierarchical manner—differentiating between low-level modality processing and high-level semantic reasoning—and integrating a sparse, learnable routing mechanism, our approach dynamically allocates computational resources based on input complexity and task relevance. We demonstrate that HMoE-ATR achieves superior predictive performance compared to monolithic baselines while reducing computational overhead by approximately 40%. The proposed automatic routing gate effectively mitigates interference between competing tasks, such as anomaly detection and activity recognition, by learning orthogonal pathways through the expert network.