Automatic Optimization of Deep Learning Training through Feature-Aware-Based Dataset Splitting

Abstract

The proliferation of classification-capable artificial intelligence (AI) across a wide range of domains (e.g., agriculture, construction, etc.) has been allowed to optimize and complement several tasks, typically operationalized by humans. The computational training that allows providing such support is frequently hindered by various challenges related to datasets, including the scarcity of examples and imbalanced class distributions, which have detrimental effects on the production of accurate models. For a proper approach to these challenges, strategies smarter than the traditional brute force-based K-fold cross-validation or the naivety of hold-out are required, with the following main goals in mind: (1) carrying out one-shot, close-to-optimal data arrangements, accelerating conventional training optimization; and (2) aiming at maximizing the capacity of inference models to its fullest extent while relieving computational burden. To that end, in this paper, two image-based feature-aware dataset splitting approaches are proposed, hypothesizing a contribution towards attaining classification models that are closer to their full inference potential. Both rely on strategic image harvesting: while one of them hinges on weighted random selection out of a feature-based clusters set, the other involves a balanced picking process from a sorted list that stores data features’ distances to the centroid of a whole feature space. Comparative tests on datasets related to grapevine leaves phenotyping and bridge defects showcase promising results, highlighting a viable alternative to K-fold cross-validation and hold-out methods.