Efficient guided evolution for neural architecture search
Vasco Lopes, Miguel Santos, Bruno Degardin, Luı́s A. Alexandre
Abstract
Neural Architecture Search methods have been successfully applied to image tasks with excellent results. However, NAS methods are often complex and tend to quickly converge for local minimas. In this paper, we propose G-EA, a novel approach for guided NAS. G-EA guides the evolution by exploring the search space by generating and evaluating several architectures in each generation at initialisation stage using a zero-proxy estimator, where only the highest-scoring architecture is trained and kept for the next generation. By generating several off-springs from an existing architecture at each generation, G-EA continuously extracts knowledge about the search space without added complexity. More, G-EA forces exploitation of the most performant architectures by descendant generation while at the same time forcing exploration by parent mutation and favouring younger architectures to the detriment of older ones. Experimental results demonstrate the effectiveness of the proposed method. Results show that G-EA achieves state-of-the-art results in NAS-Bench-101 and in all NAS-Bench-201 search space data sets: CIFAR-10, CIFAR-100 and ImageNet16-120, with mean accuracies of 93.99%, 72.62% and 46.04% respectively.