Self-Supervised OOD Detection

The model is trained to classify the rotation angle (0°, 90°, 180°, 270°) applied to an input image. This forces the network to learn object-centric features like orientation and shape semantics. At inference, OOD inputs produce high rotation prediction error because the model has not learned the geometric structure of unknown classes. The rotation loss serves as a strong anomaly score, often outperforming softmax-based methods on benchmarks like CIFAR-10 vs SVHN.

Frameworks like SimCLR and MoCo pull augmented views of the same sample together in embedding space while pushing all other samples apart. This creates a highly structured feature manifold where in-distribution data clusters tightly. OOD detection is performed by measuring the cosine similarity or Euclidean distance to nearest training embeddings. Key advantages:

• No class labels required during pretraining
• Learns instance-level discrimination
• Naturally separates OOD inputs that fall outside the learned manifold
• Works with KNN-based scoring for non-parametric detection

The input image is divided into a grid of patches, shuffled, and the model must predict the correct permutation index. This pretext task teaches the network to understand spatial relationships and contextual coherence between object parts. OOD inputs disrupt these learned spatial priors, resulting in low permutation prediction accuracy. The jigsaw loss captures semantic novelty that pixel-level reconstruction often misses, making it effective against near-OOD examples.

Inspired by MAE (Masked Autoencoders), this approach randomly masks large portions of the input and trains the model to reconstruct the missing patches. The reconstruction loss on held-out validation data serves as the OOD score. In-distribution samples are reconstructed with high fidelity because the model has learned their statistical regularities. OOD inputs exhibit elevated reconstruction error due to unfamiliar patch statistics. This method scales well to high-resolution inputs and requires no negative sampling.

A neural network is trained to map all normal training samples into a minimal-volume hypersphere centered at a fixed point in feature space. The objective penalizes the distance of each sample's embedding from the center. At inference, the distance to the center becomes the anomaly score—OOD inputs map far from the learned hypersphere boundary. This method is fully unsupervised, requires no auxiliary outlier data, and provides a compact, parametric description of normality.

Combining multiple pretext tasks—such as rotation prediction, colorization, and relative patch location—into a single training objective creates a diverse feature representation that captures complementary aspects of the data distribution. Each task head provides an independent OOD signal. The aggregated loss across all tasks yields a more robust detection score than any single pretext task alone. This ensemble effect reduces false positives and improves calibration on challenging near-OOD distributions.