Accelerating Evolutionary Neural Architecture Search via Multifidelity Evaluation

Evolutionary neural architecture search (ENAS) has recently received increasing attention by effectively finding high-quality architectures, which however consumes high computational cost training the encoded each individual for complete epochs in evaluation. Numerous ENAS approaches have been developed to reduce evaluation cost, but it is often difficult most of these achieve accuracy. To address this issue, article, we propose an accelerated via multifidelity termed MFENAS, where significantly reduced only a small number epochs. The balance between and accuracy well maintained suggesting evaluation, identifies potentially good individuals that cannot survive from previous generations integrating multiple evaluations under different numbers Besides, population initialization strategy devised produce diverse architectures varying ResNet-like Inception-like ones. As shown experiments, proposed MFENAS takes 0.6 GPU days find best holding 2.39% test error rate, superior state-of-the-art approaches. And transferred CIFAR-100 ImageNet also exhibit competitive performance.