Year: 2020
Type: dissertation
Abstract: This thesis investigates how well a neural architecture search can find competitive image classifiers on the CIFAR-10 data set with limited computational resources. Most work done on architecture search either uses vast computational resources or narrow and strongly informed space of possible solutions. The solution space proposed consists of residual convolutional networks with separable convolutions and is explored by a primitive evolutionary procedure. When regularizing the evolution it proceeds to find an architecture outperforming the ResNet-110 baseline. Further evolution is shown to find networks with significantly stronger average performance than random search. In comparison with Bayesian optimization and local search, evolution show competitive final performance. Bayesian optimization however gives unmatched sample efficiency. One pressing matter in the process of reducing computations is how extensive networks must be trained before evaluation. Small scale experiments shows a tolerable rank-correlation between early performance and final performance of residual networks but indicates a bias for shallow networks.
Author: Samuel Lundberg
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W3042520409