Abstract
Neural architecture search (NAS) has gained significant attention for automatic network design in recent years. Previous NAS methods suffer from limited search spaces, which may lead to sub-optimal results. In this paper, we propose UENAS, an evolution-based NAS framework with a broader search space that supports optimizing network architectures, pruning strategies, and hyperparameters simultaneously. To alleviate the huge search cost caused by the expanded search space, three strategies are adopted: First, an adaptive pruning strategy that iteratively trims the average model size in the population without compromising performance. Second, child networks share weights of overlapping layers with pre-trained parent networks to reduce the training epochs. Third, an online predictor scores the joint representations of architecture, pruning strategy, and hyperparameters to filter out inferior combos. By the proposed three strategies, the search efficiency is significantly improved and more well-performed compact networks with tailored hyper-parameters are derived. In experiments, UENAS achieves error rates of 2.81% on CIFAR-10, 20.24% on CIFAR-100, and 33% on Tiny-ImageNet, which shows the effectiveness of our method.
Abstract (translated)
URL
https://arxiv.org/abs/2203.04300
