Non-Linear Outlier Synthesis for Out-of-Distribution Detection

TL;DR

NCIS enhances the quality of synthetic outliers by operating directly in the diffusion's model embedding space rather than combining disjoint models as in previous work and by modeling class-conditional manifolds with a conditional volume-preserving network for more expressive characterization of the training distribution.

Abstract

The reliability of supervised classifiers is severely hampered by their limitations in dealing with unexpected inputs, leading to great interest in out-of-distribution (OOD) detection. Recently, OOD detectors trained on synthetic outliers, especially those generated by large diffusion models, have shown promising results in defining robust OOD decision boundaries. Building on this progress, we present NCIS, which enhances the quality of synthetic outliers by operating directly in the diffusion's model embedding space rather than combining disjoint models as in previous work and by modeling class-conditional manifolds with a conditional volume-preserving network for more expressive characterization of the training distribution. We demonstrate that these improvements yield new state-of-the-art OOD detection results on standard ImageNet100 and CIFAR100 benchmarks and provide insights into the importance of data pre-processing and other key design choices. We make our code available at \url{https://github.com/LarsDoorenbos/NCIS}.

Figures (7)

• Figure 1: Random outliers generated for three CIFAR100 classes by our method. Using our generated samples as auxiliary outliers during training greatly improves the OOD detection performance of modern classifiers.
• Figure 2: Comparison between (a) Dream-OOD, (b) linear invariants on our embeddings, and (c) our proposed method on a toy example. The disjoint embeddings and normalization of Dream-OOD greatly limit the flexibility of the generated outliers. Linear invariants similarly lack capacity. On the other hand, our method can generate successful outliers by modeling arbitrary distributions.
• Figure 3: The class-specific representations learned by the cVPN on toy data with three classes. Depending on the conditioning, the cVPN transforms the input data such that the current class is transformed into a representation with an invariant (the y-axis). The background color indicates the distance to the nearest training data point from the current class in the original space. Images with a white-shaded background result from rotation layers, and images with a gray background result from the conditional coupling layers.
• Figure 4: Effect of regularization on generated outliers. Outliers get progressively more OOD with stronger regularization, providing an intuitive way to control their difficulty.
• Figure 5: Nine random generated outliers by our method and Dream-OOD for CIFAR100 (top) and ImageNet-100 (bottom). Our generated outliers are closer to the intended meaning, providing better signal to learn the ID/OOD decision boundary.