Training Objectives Compared for OOD Detection in Image Classification
A new study published on arXiv examines various training objectives for detecting out-of-distribution (OOD) samples in image classification. Researchers analyzed four different loss functions: Cross-Entropy Loss, Prototype Loss, Triplet Loss, and Average Precision Loss, focusing on their effectiveness within the framework of OpenOOD protocols. Utilizing a ResNet-18 model, the researchers found that Cross-Entropy Loss consistently yielded superior performance, particularly in distinguishing near and far OOD samples. While Prototype Loss and Average Precision Loss showed competitive results under certain conditions, the overall findings emphasize the critical role of training strategies in applications where safety is paramount.
Key facts
- Four training objectives compared: Cross-Entropy Loss, Prototype Loss, Triplet Loss, and Average Precision (AP) Loss.
- Evaluation conducted under standardized OpenOOD protocols.
- ResNet-18 architecture used in the experiments.
- Cross-Entropy Loss, Prototype Loss, and AP Loss achieve comparable in-distribution accuracy.
- Cross-Entropy Loss provides the most consistent near- and far-OOD AUROC overall.
- Other objectives can be competitive in specific settings.
- Paper available on arXiv with ID 2603.07571.
- Study addresses the influence of training objectives on OOD detection, which is critical for safety-sensitive applications.
Entities
Institutions
- arXiv