Core Concepts
Leveraging Hyperbolic geometry for Out-Of-Distribution (OOD) detection improves FPR95 in visual data.
Abstract
Abstract:
OOD detection is crucial for deploying deep learning models in safety-critical applications.
Conventional methods based on Euclidean geometry struggle to capture hierarchical structures in visual data.
Proposed metric framework leverages Hyperbolic geometry for improved OOD detection, showing better results than Euclidean methods.
Introduction:
Deep Learning excels within training data but faces challenges with unforeseen data.
OOD detection is vital for safe deployment in scenarios like medical imaging and autonomous vehicles.
Existing approaches suffer from overconfidence or reliance on external datasets.
Preliminaries:
Training set D = {(xi, yi)}N drawn i.i.d., Pid as the marginal distribution on X (in-distribution).
Out-Of-Distribution Detection aims to distinguish between two classes during testing.
Hyperbolic Space for Out-of-Distribution Detection:
Compact Representation of In-Distribution Embeddings using a feature encoder and projection onto the hyperboloid.
Classification of In-Distribution Embeddings involves determining distance to margin hyperplanes within the Lorentz model.
Regularization with Synthetic Hyperbolic Outliers aims to improve model boundaries between ID and OOD samples.
Experimental Results:
Setup includes CIFAR-10, CIFAR-100, and ImageNet-200 datasets with evaluations on FPR95 and AUROC metrics.
HOD outperforms Euclidean methods in OOD detection tasks, especially at lower dimensions.
Related Work:
Research on OOD detection methodologies and regularization-based approaches discussed.
Limitations:
Impact of Hyperbolic embeddings in architectures of varying depths not explored.
Effects beyond natural image datasets not investigated.
Conclusions:
HOD leverages Hyperbolic space effectively for improved OOD detection, surpassing state-of-the-art methods in FPR95 on CIFAR datasets.
Stats
Extensive experiments show that our framework improves the FPR95 for OOD detection from 22% to 15% and from 49% to 28% on CIFAR-10 and CIFAR-100 respectively compared to Euclidean methods.
The loss function Lhsup promotes low intra-class variation and high inter-class separation in the Hyperbolic space.
Quotes
"We hypothesize that the gap between Euclidean and Hyperbolic space arises from differences in the training objectives employed."
"Synthetic outliers do not bring any benefit in Hyperbolic space."