Automated deep learning architecture design using differentiable architecture search (DARTS)
Date
2019
Authors
Sharma, Kartikay, author
Anderson, Chuck, advisor
Beveridge, Ross, committee member
Kirby, Michael, committee member
Journal Title
Journal ISSN
Volume Title
Abstract
Creating neural networks by hand is a slow trial-and-error based process. Designing new architectures similar to GoogleNet or FractalNets, which use repeated tree-based structures, is highly likely to be inefficient and sub-optimal because of the large number of possibilities for composing such structures. Recently, neural architecture search algorithms have been able to automate the process of architecture design and have often attained state-of-the-art performances on CIFAR-10, ImageNet and Penn Tree Bank datasets. Even though the search time has been reduced to tens of GPU hours from tens of thousands of GPU hours, most search algorithms rely on additional controllers and hypernetworks to generate architecture encoding or predict weights for sampled architectures. These controllers and hypernetworks might require optimal structure when deployed on a new task on a new dataset. And since this is done by hand, the problem of architecture search is not really solved. Differentiable Architecture Search (DARTS) avoids this problem by using gradient descent methods. In this work, the DARTS algorithm is studied under various conditions and search hyperparameters. DARTS is applied to CIFAR-10 to check reproducibility of the original results. It is also tested in a new setting — on the CheXpert dataset — to discover new architectures and is compared to a baseline DenseNet121 model. The architectures searched using DARTS achieve better performance on the validation set than the baseline model.
Description
Rights Access
Subject
convolutional neural networks (CNNs)
differentiable architecture search
CheXpert dataset
neural architecture search
DARTS