Efficient Out-of-Distribution Detection via CVAE data Generation
Recently, contrastive loss with data augmentation and pseudo class creation has been shown to produce markedly better results for out-of-distribution (OOD) detection than previous methods. However, a major shortcoming of this approach is that it is extremely slow due to significant increase in the data size and the number of classes and the quadratic complexity of pairwise similarity computation. This paper proposes a novel and simple method that can build an effective data generator using Conditional Variational Auto-Encoder (CVAE) to generate pseudo OOD samples. Based on the generated pseudo OOD data, a flexible and efficient OOD detection method is proposed through fine-tuning, which achieves results comparable to the state-of-the-art OOD detection techniques, but the execution speed is at least 10 times faster. Also importantly, the proposed approach is in fact a general framework that can be applied to many existing OOD methods and improve them via the proposed fine-tuning. We have combined it with the best baseline OOD models in our experiments to produce new state-of-the-art results.
PDF Abstract
CIFAR-10
CIFAR-100
SVHN
Tiny ImageNet
LSUN
