Abstract
The problem of detecting a novel class at run time is known as Open Set Detection & is important for various real-world applications like medical application, autonomous driving, etc. Open Set Detection within context of deep learning involves solving two problems: (i) Must map the input images into a latent representation that contains enough information to detect the outliers, and (ii) Must learn an anomaly scoring function that can extract this information from the latent representation to identify the anomalies. Research in deep anomaly detection methods has progressed slowly. One reason may be that most papers simultaneously introduce new representation learning techniques and new anomaly scoring approaches. The goal of this work is to improve this methodology by providing ways of separately measuring the effectiveness of the representation learning and anomaly scoring. This work makes two methodological contributions. The first is to introduce the notion of Oracle anomaly detection for quantifying the information available in a learned latent representation. The second is to introduce Oracle representation learning, which produces a representation that is guaranteed to be sufficient for accurate anomaly detection. These two techniques help researchers to separate the quality of the learned representation from the performance of the anomaly scoring mechanism so that they can debug and improve their systems. The methods also provide an upper limit on how much open category detection can be improved through better anomaly scoring mechanisms. The combination of the two oracles gives an upper limit on the performance that any open category detection method could achieve. This work introduces these two oracle techniques and demonstrates their utility by applying them to several leading open category detection methods.
Introduction: The Open Set Challenge
Most deep learning image classifiers operate under a "closed-world assumption," assuming every image encountered at runtime belongs to a class seen during training. However, real-world applications—from autonomous driving to medical diagnostics—frequently encounter novel or "unknown" categories. A robust system should flag these as anomalies rather than misclassifying them. This task is known as Open Set Detection (OSD).
OSD research has historically been bottlenecked by the simultaneous introduction of new representation learning techniques and new anomaly scoring functions. When a system fails, it is difficult to determine if the internal data representation is insufficient or if the scoring mechanism is simply failing to extract the available information. This paper proposes a diagnostic framework using "Oracles" to decouple these two components.
Methodology: Decoupling Representation and Scoring
To isolate the variables of OSD, the authors introduce two "Oracle" techniques that establish performance upper bounds:
- Oracle Anomaly Detection (Oracle AD): This method tests whether a learned representation contains enough information to detect anomalies. A supervised classifier (e.g., a Random Forest) is trained specifically to distinguish between known and unknown classes using only the latent features of a pre-trained model. Because this Oracle has access to labels unavailable at runtime, it represents the "perfect" scoring function for a given representation.
- Oracle Representation Learning (Oracle Rep): This method establishes the absolute upper limit of the network architecture. A model is trained on all classes (including the "unknown" ones) as if they were known. This reveals how well the architecture could potentially separate categories if it had seen them during training.
Experimental Results and Insights
The analysis revealed several critical findings regarding the current state of OSD across benchmarks like MNIST, CIFAR-10, and CIFAR-100:
- The "Headroom" Gap: There is a massive performance gap between standard anomaly measures (like entropy) and the Oracle AD. On CIFAR-100, standard measures achieved an AUROC of ~0.72, while the Oracle AD reached ~0.82 on the same features. This indicates that current OSD methods are limited more by suboptimal scoring functions than by the representations themselves.
- Hybrid Representations: Models combining classification loss (Softmax) with reconstruction loss (Autoencoder) consistently provided the most informative features for detecting novel classes.
- Discriminative vs. Generative: Purely generative models (standard Autoencoders) struggled on complex datasets like CIFAR. While they achieved near-perfect OSD on MNIST (~0.98 AUROC), they dropped to ~0.61 on CIFAR-10, suggesting reconstruction error is a poor novelty signal for complex images.
- Best Practical Signal: Among non-oracle methods, the Maximum Class Logit (unnormalized network output) was the most effective signal, outperforming Softmax confidence or Entropy.
Conclusion
The "Oracle" framework demonstrates that while better representations help, the most immediate gains in Open Set Detection lie in improving anomaly scoring mechanisms. On CIFAR-100, the Hybrid model's performance jumped from 0.724 (standard scoring) to 0.828 (Oracle AD), and finally to 0.857 (Oracle Representation). Future work should focus on developing deployable scoring functions that approach Oracle performance without requiring "unknown" labels during training.
Summary updated