Difficulty Within Deep Learning Object-Recognition Due to Object Variance

摘要

It is one of the areas where deep learning models demonstrate possible performance bottleneck to learn objects with variations rapidly and precisely. We research on the variances of visual objects in terms of the difficulty levels of learning performed by deep learning models. Multiple dimensions and levels of variances are defined and formalized in the form of categories. We design "variance categories" and "quantitative difficulty levels" and translate variance categories into difficulty levels. We experiment how multiple learning models are affected by the categorization of variance separately and in combination (of several categories). Our experimental analysis on learning of the dataset demonstrates not only the expected way of utilization of variance of the data differs from models, the amount of learning or information gained for each data fed into models also varies significantly. Our results suggest it matters to search for a possible key explicit representation of the "invariance" part of objects (or of the respective cognitive mechanism) and for the pertinent elements and capabilities in the deep learning architectures. It can be used to make the learning models a match to humans on complex object recognition tasks.