Abstract
In this paper, we present a novel approach that combines deep metric learningand synthetic data generation using diffusion models for out-of-distribution(OOD) detection. One popular approach for OOD detection is outlier exposure,where models are trained using a mixture of in-distribution (ID) samples and``seen" OOD samples. For the OOD samples, the model is trained to minimize theKL divergence between the output probability and the uniform distribution whilecorrectly classifying the in-distribution (ID) data. In this paper, we proposea label-mixup approach to generate synthetic OOD data using Denoising DiffusionProbabilistic Models (DDPMs). Additionally, we explore recent advancements inmetric learning to train our models. In the experiments, we found that metric learning-based loss functionsperform better than the softmax. Furthermore, the baseline models (includingsoftmax, and metric learning) show a significant improvement when trained withthe generated OOD data. Our approach outperforms strong baselines inconventional OOD detection metrics.