Deep Generalized Prediction Set Classifier and Its Theoretical Guarantees

Published: 20 May 2024, Last Modified: 20 May 2024Accepted by TMLREveryoneRevisionsBibTeX
Abstract: A standard classification rule returns a single-valued prediction for any observation without a confidence guarantee, which may result in severe consequences in many critical applications when the uncertainty is high. In contrast, set-valued classification is a new paradigm to handle the uncertainty in classification by reporting a set of plausible labels to observations in highly ambiguous regions. In this article, we propose the Deep Generalized Prediction Set (DeepGPS) method, a network-based set-valued classifier induced by acceptance region learning. DeepGPS is capable of identifying ambiguous observations and detecting out-of-distribution (OOD) observations. It is the first set-valued classification of this kind with a theoretical guarantee and scalable to large datasets. Our nontrivial proof shows that the risk of DeepGPS, defined as the expected size of the prediction set, attains the optimality within a neural network hypothesis class while simultaneously achieving the user-prescribed class-specific accuracy. Additionally, by using a weighted loss, DeepGPS returns tighter acceptance regions, leading to informative predictions and improved OOD detection performance. Empirically, our method outperforms the baselines on several benchmark datasets.
Certifications: Featured Certification
Submission Length: Regular submission (no more than 12 pages of main content)
Changes Since Last Submission: camera-ready version
Assigned Action Editor: ~Yiming_Ying1
Submission Number: 2143