Let the Void Be Void: Robust Open-Set Semi-Supervised Learning via Selective Non-Alignment

TL;DR

SkipAlign tackles open-set semi-supervised learning by introducing Selective Non-Alignment (SNA), a third contrastive operator that pulls only confident ID samples toward their prototypes while applying angular repulsion to uncertain or OOD samples, creating dense ID clusters within a diffuse OOD void. A dual-gate confidence mechanism and a projection head enable SNA to refine representations and prototypes, balancing ID discrimination and OOD detection. The method combines three losses—L_CC, L_OD, and L_SNA—with adaptive prototype refinement, achieving strong generalization to unseen OOD across CIFAR-10/100, ImageNet-30, and TinyImageNet, with notable improvements in overall OOD-AUC and competitive closed-set accuracy. Theoretical analysis links SNA-driven angular dynamics to a larger geometric margin γ_SNA, which reduces Rademacher complexity and yields robust performance for OOD detectors in OSS scenarios, while qualitative analyses illustrate the resulting ID galaxies and interstellar OOD void.

Abstract

Open-set semi-supervised learning (OSSL) leverages unlabeled data containing both in-distribution (ID) and unknown out-of-distribution (OOD) samples, aiming simultaneously to improve closed-set accuracy and detect novel OOD instances. Existing methods either discard valuable information from uncertain samples or force-align every unlabeled sample into one or a few synthetic "catch-all" representations, resulting in geometric collapse and overconfidence on only seen OODs. To address the limitations, we introduce selective non-alignment, adding a novel "skip" operator into conventional pull and push operations of contrastive learning. Our framework, SkipAlign, selectively skips alignment (pulling) for low-confidence unlabeled samples, retaining only gentle repulsion against ID prototypes. This approach transforms uncertain samples into a pure repulsion signal, resulting in tighter ID clusters and naturally dispersed OOD features. Extensive experiments demonstrate that SkipAlign significantly outperforms state-of-the-art methods in detecting unseen OOD data without sacrificing ID classification accuracy.

Figures (9)

• Figure 1: Overview of SkipAlign. Left: Illustration of the Selective Non-Alignment (SNA) concept. Confident ID samples are attracted toward their corresponding class prototypes, while uncertain or OOD samples are softly repelled and remain unaligned. Right: The overall framework of SkipAlign. A shared encoder feeds into a closed-set classifier ($CC(\cdot)$) and an OOD detector ($OD(\cdot)$), which jointly perform dual-gate ID selection. Projection embeddings are used to form class prototypes, enabling SNA to refine representation learning through prototype-based alignment and repulsion.
• Figure 2: Effect of the SNA module. (a) t-SNE visualization of projection embeddings $\mathbf{z}$. (b) Average feature norm for each class category. (c) Average cosine similarity between sample categories and ID class prototypes.
• Figure 3: Adaptive Prototype Generation. For each class, the labeled prototype $\mu_l$ is the mean embedding of labeled samples, while the unlabeled prototype $\mu_u$ is obtained from confident ID samples among the unlabeled data. The final prototype $\mu$ is computed as a weighted sum of the two.
• Figure 4: Per-dataset OOD detection AUC for various OSSL methods across different experimental configurations (CIFAR-10, CIFAR-100 with varying ID/OOD splits and 25 labels, and ImageNet30). SkipAlign consistently achieves strong and balanced detection performance on most OOD datasets.
• Figure 5: t-SNE visualization of feature embeddings from the OOD detector (CIFAR-10, 6/4-50). Red contours highlight problematic regions in prior methods: SkipAlign yields compact, class-specific ID clusters and clear ID–OOD separation, with unseen OODs naturally positioned in non-ID regions or around the OOD boundary.