An Experimental Protocol for Neural Architecture Search in Super-Resolution

Neural architecture search has seen continual progress due to the interest in automating architecture design in deep learning following the promise of finding the best possible neural network architecture tailored for a particular task. Recently, many works focused on tackling tasks like image classification and language modeling, allowing significant developments in computer vision and NLP. As research in such directions has established standard criteria and benchmarking tasks for algorithmic performance comparison, the same cannot be said of other applications and tasks. Our work presents an experimental comparison protocol that narrows down the process of evaluating super-resolution image restoration architectures in neural architecture search approaches. Such protocol consists of two datasets for training and validation during and after the architecture search, and the application of a Bayesian statistical test for studying the observable results.