M2M-Net: multi-objective neural architecture search using dynamic M2M population decomposition

Evolutionary multi-objective neural architecture search (MO-NAS) is an efficient solution for automating the design of deep neural network architectures, aiming to explore a diverse range of objectives. However, the sensitivity of objectives in MO-NAS varies over generations, leading to a search imbalance, and local search plays a crucial role in this combinatorial optimization problem with a discrete search space. In this paper, we propose M2M-Net, a dynamic self-adaptive (Multi-objective to Multi-objective) M2M population decomposition-based evolutionary algorithm for NAS. M2M-Net leverages dynamic self-adaptive M2M population decomposition to overcome the search imbalance in MO-NAS. The subpopulation-based search within M2M-Net facilitates local search through crossover and mutation. Additionally, M2M-Net incorporates a proxy model to reduce computational cost in architecture evaluation and utilizes the channel attention mechanism to improve the accuracy of proxy model evaluation. Experimental studies on CIFAR-10 and CIFAR-100 datasets validate the effectiveness and efficiency of M2M-Net. Comparisons and analysis demonstrate that M2M-Net achieves comparable performance to state-of-the-art NAS methods while utilizing fewer computational resources.