Acceptance and Rejection Zones for a Classifier’s Predictions in Deep Learning
Adama Samake, Lahsen Boulmane · 2021
A big challenge in deep learning is to estimate the reliability of a classifier’s predictions. In this paper, we propose to approach this problem through the concept of acceptance, rejection and abstain zones. Given a new example, a reliability score is defined from the cosine similarities between the activation values of the model output layer and those of its correct and incorrect predictions on the test set. The corresponding zone to a prediction is determined by its reliability score. Then the prediction is accepted, for the acceptance zone. Otherwise it is rejected for the reject zone and not considered for the abstain zone. In order to empirically examine our proposition, we carried out a case study in text classification. The results obtained show the relevance of the proposed approach.