DDS-NAS: Dynamic data selection within neural architecture search via on-line hard example mining applied to image classification

Matt Poyser, Toby P. Breckon

Abstract: Highlights•A NAS method that uses hard example mining/curriculum learning to improve the NAS search process.•Novel consideration of image dissimilarity as an alternative metric to hardness.•Image similarity embeddings combined with a kd -tree for dissimilar image look-up.•Generation of models in a manner intrinsically robust to biased datasets.•Generation of models an order of magnitude faster than existing NAS techniques.