Multi-Task Learning Based on the Multi-Gate Mixture-of-Experts Architecture for Automated and Explainable Edible Bird’s Nest Grading

Highlights Multi-gate mixture-of-experts for multi-attribute grading. Hierarchical multi-task learning framework. Interpretable and traceable grading decisions. Automated quality assessment for agricultural products. Abstract. Automated edible bird’s nest (EBN) grading is a critical and challenging task in the agricultural product processing industry, involving the evaluation of complex, interrelated attributes. This study proposes a hierarchical multi-task automated grading framework based on a multi-gate mixture-of-experts (MMoE) architecture. The framework solves the complex EBN grading problem by evaluating multiple, interconnected attributes through shared feature learning. Pretrained ConvNeXt model extract features from EBN images. The features are subsequently processed by specialized expert networks within the MMoE architecture to enable effective attribute-specific classification for color, shape, impurities, and cracks. MMoE gating mechanisms allow different tasks to adopt their own relevant expert combinations, and multi-task gradient signals can also guide the learning process of backbone features through dynamic weight assignment. Focal loss is used to address class imbalance and improve the model’s ability to discriminate. Final EBN grades are determined by a two-stage process that evaluates individual attributes and combines their information to generate five predefined quality grades. Experimental results show that, on a complex, multi-attribute EBN dataset, the proposed method achieves 92% grading accuracy, and all attribute classifications improve. Framework interpretability is demonstrated through performance improvements and Grad-CAM activation maps. This hierarchical, interpretable grading system follows human logic, making decisions transparent so that the reasons for each grade can be traced. The proposed approach is a practical solution to the problem of processing agricultural products.