The Truth is Out There: Focusing on Smaller to Guess Bigger in Image Classification

Abstract

In Artificial Intelligence (AI) in general and in Machine Learning (ML) in particular, which are important and integral components of modern Industry 4.0, we often deal with uncertainty, e.g., lack of complete information about the objects we are classifying, recognizing, diagnosing, etc. Traditionally, uncertainty is considered to be a problem especially in the responsible use of AI and ML tools in the smart manufacturing domain. However, in this study, we aim not to fight with but rather to benefit from the uncertainty to improve the classification performance in supervised ML. Our objective is a kind of uncertainty-driven technique to improve the performance of Convolutional Neural Networks (CNNs) for image classification. The intuition behind our suggested “decontextualize-and-extrapolate” approach is as follows: any image not necessarily contains all the needed information for perfect classification; any trained CNN will give for the entire image (with some uncertainty) the probability distribution among possible classes; the same CNN may also give similar probability distribution to the “part” of the image (i.e., with the higher uncertainty); one may discover the trend of the probability distribution change with the change of uncertainty value; a better (refined) probability distribution could be computed from these two distributions as the result of their extrapolation towards the less uncertainty. In this paper, we suggested several ways and corresponding analytics to discover reasonable part(s) of the images and to make the extrapolation to get better (refined) image classification results. We have considered image representation at the level of pixels as well as at the level of the discovered features. Our preliminary experiments show that the suggested refinement techniques (applied during the testing phase of the CNNs) can improve their classification performance.