Visualizations for universal deep-feature representations: survey and taxonomy

Abstract In data science and content-based retrieval, we find many domain-specific techniques that employ a processing pipeline with two fundamental steps. First, entities are represented by some visualizations, while in the second step, visualizations used machine learning model to extract deep features. Deep convolutional neural networks (DCNN) became standard reliable choice. The purpose of using DCNN is either specific classification task or just feature representation visual for additional (e.g., similarity search). Whereas extraction domain-agnostic step (inference an arbitrary input), visualization design itself domain-dependent ad hoc every use case. this paper, survey analyze instances models (mostly DCNN) tasks. Based on analysis, synthesize taxonomy provides systematic overview suitable usage models. aim enable future generalization process become completely domain-agnostic, leading automation entire pipeline. As ultimate goal, such automated could lead universal representations retrieval.