Abstract
Multi-task neural architecture search (NAS) enables transferringarchitectural knowledge among different tasks. However, ranking disorderbetween the source task and the target task degrades the architectureperformance on the downstream task. We propose KTNAS, an evolutionarycross-task NAS algorithm, to enhance transfer efficiency. Our data-agnosticmethod converts neural architectures into graphs and uses architectureembedding vectors for the subsequent architecture performance prediction. Theconcept of transfer rank, an instance-based classifier, is introduced intoKTNAS to address the performance degradation issue. We verify the searchefficiency on NASBench-201 and transferability to various vision tasks on MicroTransNAS-Bench-101. The scalability of our method is demonstrated on DARTssearch space including CIFAR-10/100, MNIST/Fashion-MNIST, MedMNIST.Experimental results show that KTNAS outperforms peer multi-task NAS algorithmsin search efficiency and downstream task performance. Ablation studiesdemonstrate the vital importance of transfer rank for transfer performance.