Semi-Supervised Hierarchical Open-Set Classification

TL;DR

This work tackles hierarchical open-set classification in a semi-supervised setting by enabling models to learn from unlabeled mixtures of in-distribution and out-of-distribution data. It introduces SemiHOC, a teacher–Student framework that combines ProHOC-based hierarchical predictions with subtree pseudo-labels and a novel age-gating mechanism to prevent overconfident, overly specific OOD assignments. The key contributions are the subtree pseudo-labels for robust OOD supervision and age-gating to curb late SPL overpredictions, integrated within a semi-supervised learning loop. Empirically, SemiHOC outperforms SSL baselines and matches fully supervised performance on iNaturalist19 with only 20 labeled samples per class, demonstrating strong practical impact for scale-limited open-world recognition in hierarchical taxonomies.

Abstract

Hierarchical open-set classification handles previously unseen classes by assigning them to the most appropriate high-level category in a class taxonomy. We extend this paradigm to the semi-supervised setting, enabling the use of large-scale, uncurated datasets containing a mixture of known and unknown classes to improve the hierarchical open-set performance. To this end, we propose a teacher-student framework based on pseudo-labeling. Two key components are introduced: 1) subtree pseudo-labels, which provide reliable supervision in the presence of unknown data, and 2) age-gating, a mechanism that mitigates overconfidence in pseudo-labels. Experiments show that our framework outperforms self-supervised pretraining followed by supervised adaptation, and even matches the fully supervised counterpart when using only 20 labeled samples per class on the iNaturalist19 benchmark. Our code is available at https://github.com/walline/semihoc.

Figures (13)

• Figure 1: Semi-supervised hierarchical open-set classification. The goal is to predict data from the open-world distribution (gray) to the appropriate hierarchy nodes.
• Figure 2: Average accuracy as a function of prediction confidence for samples predicted as ID and OOD. The shaded regions show the corresponding confidence histograms. The subtree confidence enables pseudo-labeling for OOD-predictions by achieving high accuracy for high-confidence predictions.
• Figure 3: Distances (the number of edges in $\mathcal{H}$) to the ground truth for data predicted as OOD (internal nodes), displayed as percentages. LCA denotes the Lowest Common Ancestor of the prediction and the ground-truth.
• Figure 4: Frequencies of subtree pseudo-label assignments across training epochs for selected hierarchy nodes, split into correct and incorrect assignments. Each node's local hierarchy structure and example images from correct and incorrect assignments are shown to the right. We see that the majority of correct SPLs are assigned early in training, while incorrect overpredictions of OOD data emerge later as distinct peaks.
• Figure 5: The flow of labeled and unlabeled data in SemiHOC. Labeled data are used to train depth-specific student models with cross-entropy losses using labels mapped to the specified depths. Unlabeled data are predicted by teacher models via ProHOC, producing hierarchical distributions that form subtree pseudo-labels. After age-gating, these subtree pseudo-labels are mapped to the appropriate depths and used to train the students with cross-entropy losses.