AutoML and Neural Architecture Search for Gaze Estimation

Abstract

This paper focuses on employing the AutoKeras neural architecture search tool and Automatic Machine Learning (AutoML) methods to find an optimal gaze estimation system. The Network Architecture Search (NAS) means automatically tuning already existing deep neural network configurations using a dataset of interest. The algorithm will search in the architectural space for a better neural model along with its optimized parameters. In the paper context, an AutoML solution will perform similarly, but using just ‘pure’ ML models. Considering “Appearance-based Gaze Estimation in the Wild” (MPIIGaze) and “Columbia Gaze Data Set” (CAVE) datasets, the experiments showed results comparable to those of with manually designed models.