Out-of-distribution Detection in Classifiers via Generation

TLDR

Discriminatively trained neural classifiers reliably predict only in‑distribution samples, yet real‑world deployments require reliable out‑of‑distribution detection, a capability that typical classifiers lack because they do not encode the boundary between in‑ and out‑of‑distribution data. The study investigates whether it is feasible to generate “effective” out‑of‑distribution samples that cover the in‑distribution boundary and proposes a novel algorithm that uses a manifold‑learning network, such as a variational autoencoder, to produce such samples and trains an n+1 classifier where the extra class represents out‑of‑distribution data. The method employs a manifold‑learning network to generate boundary‑adjacent out‑of‑distribution samples, trains an n+1 classifier on these samples, and evaluates the resulting detector against several recent classifier‑based OOD detectors on MNIST and Fashion‑MNIST. Experimental results show that the proposed approach consistently outperforms the compared methods.

Abstract

By design, discriminatively trained neural network classifiers produce reliable predictions only for in-distribution samples. For their real-world deployments, detecting out-of-distribution (OOD) samples is essential. Assuming OOD to be outside the closed boundary of in-distribution, typical neural classifiers do not contain the knowledge of this boundary for OOD detection during inference. There have been recent approaches to instill this knowledge in classifiers by explicitly training the classifier with OOD samples close to the in-distribution boundary. However, these generated samples fail to cover the entire in-distribution boundary effectively, thereby resulting in a sub-optimal OOD detector. In this paper, we analyze the feasibility of such approaches by investigating the complexity of producing such "effective" OOD samples. We also propose a novel algorithm to generate such samples using a manifold learning network (e.g., variational autoencoder) and then train an n+1 classifier for OOD detection, where the $n+1^{th}$ class represents the OOD samples. We compare our approach against several recent classifier-based OOD detectors on MNIST and Fashion-MNIST datasets. Overall the proposed approach consistently performs better than the others.