ADEMT-NAS: Multi-Task Neural Architecture Search Based on Adaptive Differential Evolution

Neural Architecture Search, as a significant technology for automatically designing high-performance neural networks, is rapidly becoming an active research topic in both academia and industry. However, in multi-task scenarios, the exponential expansion of the search space and increased task heterogeneity pose challenges for traditional evolutionary algorithms in effectively leveraging inter-task shared information and achieving collaborative optimization. To effectively utilize the implicit experience gained by the current population to guide evolutionary search, this paper introduces the Differential Evolution algorithm into the multi-task NAS domain for the first time. Specifically, we propose an Adaptive Differential Evolution algorithm guided by cell-based architecture similarity for MultiTask NAS (ADEMT-NAS), where the pairwise similarity between architectures within the population is harnessed as empirical knowledge to steer the evolutionary search. Concretely, ADEMTNAS employs a block-based selection strategy coupled with bitlevel mutation to achieve fine-grained mutation at the cell level. Furthermore, we develop a novel architecture similarity calculation method that integrates cell-level encoding with the Manhattan distance metric to quantify inter-task similarity, adaptively guiding the crossover phase of differential evolution and effectively mitigating negative transfer. To rigorously validate the performance of ADEMT-NAS, extensive empirical experiments are conducted on classification tasks from diverse datasets. Experimental results demonstrate that the proposed method outperforms existing state-of-the-art NAS algorithms on both CIFAR10 and CIFAR100 tasks.