FlowCon: Out-of-Distribution Detection using Flow-Based Contrastive Learning

Saandeep Aathreya, Shaun Canavan
University of South Florida, USA
[ Accepted at ECCV 2024]
Paper arXiv Code

Abstract

Identifying Out-of-distribution (OOD) data is becoming increasingly critical as the real-world applications of deep learning methods expand. Post-hoc methods modify softmax scores fine-tuned on outlier data or leverage intermediate feature layers to identify distinctive patterns between In-Distribution (ID) and OOD samples. Other methods focus on employing diverse OOD samples to learn discrepancies between ID and OOD. These techniques, however, are typically dependent on the quality of the outlier samples assumed. Density-based methods explicitly model class-conditioned distributions but this requires long training time or retraining the classifier. To tackle these issues, we introduce FlowCon, a new density-based OOD detection technique. Our main innovation lies in efficiently combining the properties of normalizing flow with supervised contrastive learning, ensuring robust representation learning with tractable density estimation. Empirical evaluation shows the enhanced performance of our method across common vision datasets such as CIFAR-10 and CIFAR-100 pretrained on ResNet18 and WideResNet classifiers. We also perform quantitative analysis using likelihood plots and qualitative visualization using UMAP embeddings and demonstrate the robustness of the proposed method under various OOD contexts. Code will be open-sourced post decision.

Highlights

  1. A new density-based OOD detection technique called FlowCon is proposed. We introduce a new loss function which contrastively learns class separability in the probability distribution space. This learning occurs without any external OOD dataset and it operates on fixed classifiers.
  2. The proposed method is evaluated on various metrics - FPR95, AUROC, AUPR-Success, and AUPR-Error and compared against state of the art. We observe that FlowCon is competitive or outperforms most methods under different OOD conditions. Additionally, FlowCon is stable even for a large number of classes and shows improvement for high-dimensional features.
  3. Histogram plots are detailed along with unified manifold approximations (UMAP) embeddings of the trained FlowCon model to respectively showcase it's OOD detection and class-preserving capabilities. We also show FlowCon's discriminative capabilities.

Results




Visualizations

UMAP Embeddings