A probabilistic scaling approach to conformal predictions in binary image classification

Deep learning solutions for image classification are more and more widespread and sophisticated today, bringing the necessity to properly address their reliability. Many approaches exist in uncertainty quantification, and, among these, conformal prediction is one of most solid and well-established frameworks. In this paper, we study another approach, defined as deep probabilistic scaling, based on the notion of scalable classifiers, combined with probabilistic scaling from order statistics. Given a pre-trained neural network for (binary) image classification and a target class on which it is desirable to control the error, this method is able to bound that error to a user-defined level (epsilon). The method individuates probabilistic safety regions of target class samples correctly predicted in high probability. We show how the proposed method links with conformal prediction, discussing analogies and differences. By considering a (binary) convolutional neural network classifier, experiments on several benchmark datasets show a good overall performance of the methodology in controlling false negatives.