Abstract Neural architecture search (NAS) has emerged as a powerful solution to address the labor-intensive design of deep neural networks, particularly in scenarios with complex constraints and conflicting objectives like accuracy vs. computational efficiency. This paper presents MGCANet, an end-to-end NAS framework that integrates multi-scale channel attention and multi-objective optimization to enhance feature representation and segmentation performance for complex structural patterns in images. The proposed meibomian gland channel attention (MGCA) module–adapted for general complex structures–employs orthogonal learning and fuzzy logic-inspired refinement to suppress background noise and capture multi-scale features, while a NAS-driven multi-objective search balances segmentation accuracy and computational overhead. MGCANet further incorporates weight inheritance mechanisms and cross-domain robustness design to enable efficient knowledge transfer across diverse imaging conditions with low memory overhead. Experimental results on standard segmentation tasks show that MGCANet achieves competitive performance in metrics such as Dice coefficient and Mean Intersection over Union (IoU), demonstrating its effectiveness in complex structure segmentation. This work highlights the effectiveness of integrating NAS-based optimization with attention mechanisms for automated feature learning in complex structural analysis, offering a scalable paradigm for real-world applications in image segmentation.

Ad hoc teamwork is a crucial challenge that aims to design an agent capable of effective collaboration with teammates employing diverse strategies without prior coordination. However, current Population-Based Training (PBT) approaches train the ad hoc agent through interaction with diverse teammates from scratch, which suffer from low efficiency. We introduce Multi-Expert Distillation (MED), a novel approach that directly distills diverse strategies through modeling across-episodic sequences. Experiments show that our algorithm achieves more efficient and stable training and has the ability to improve its behavior using historical contexts. Our code is available at https://github.com/LAMDA-RL/MED.