Semi-supervised learning trains on a few labeled examples plus many unlabeled ones, usually by having the model guess labels for the unlabeled data (pseudo-labels) and training on its own guesses. It often ends up worse than simply ignoring the unlabeled data, and this failure is sharpest when the unlabeled pool contains classes the labeled set never showed. In hierarchical label spaces such as species taxonomies, those unseen classes are rarely noise. An unknown warbler is still a bird and still a songbird; it just is not any of the labeled warbler species. A pseudo-label that commits all the way to a leaf species is wrong. A pseudo-label that stops at "songbird" is correct and useful. The question this note addresses is how a model should decide, for each unlabeled image, how deep in the tree its guess is allowed to go.
We propose to read that depth off the geometry of a hyperbolic embedding. Hyperbolic space is a curved space whose volume grows exponentially with radius, so it can hold trees naturally: general concepts sit near the center and specific ones near the edge. The radius bin an image falls into is converted into a thermometer code, a binary vector whose first $k$ bits are 1 and the rest 0. Bit $\ell$ switches on supervision at hierarchy level $\ell$ for that image. The thresholds that define the bins are not fixed in advance. Each is chosen on held-out labeled data so that unlabeled supervision at that level is only switched on where it would not do worse than a purely supervised model. In two sentences: an image's distance from the center of a hyperbolic embedding tells us how specific a label it can safely receive, and a cumulative thermometer code turns that distance into a monotone "supervise down to level $k$" mask. Thresholds are set per level so that the unlabeled data cannot push accuracy at any level below the supervised floor.
Core Claim
The claim is that radius in a taxonomy-anchored hyperbolic space is a better signal of the deepest correct ancestor than softmax confidence in a flat space. If so, gating pseudo-label depth by radius will reduce semi-supervised degradation under hierarchical class mismatch, and it will do so level by level rather than only on average.
The geometric argument for why radius should carry this signal is the following. Class prototypes are placed by a Poincaré embedding of the taxonomy, following Poincaré Embeddings for Learning Hierarchical Representations, which embeds hierarchies with low distortion in a few dimensions and puts ancestors nearer the origin. An image is trained to sit close, in hyperbolic distance, to its leaf prototype. An image that resembles several sibling leaves equally minimizes its expected distance to them by moving toward the point midway between them. In hyperbolic space that point lies inward, toward their shared parent, because shortest paths between points on separate branches bend toward the origin, as exploited in Hyperbolic Space with Hierarchical Margin Boosts Fine-Grained Learning from Coarse Labels. An unseen sibling class should therefore settle at roughly the radius of its deepest known ancestor.
We treat this as a hypothesis to test, not a fact. The research file contains no measurement showing that radius tracks correctness in image embeddings.
Relation to Prior Work
Three papers motivate the problem. Realistic Evaluation of Semi-supervised Learning Algorithms in Open Environments benchmarks 20 algorithms and finds that a mismatch between the labeled and unlabeled class sets causes the worst degradation of the shifts it studies (worst-case robustness 0.044, against 0.028 for ordinary distribution shift). It also finds FixMatch brittle under these conditions. Semi-Supervised Hierarchical Open-Set Classification (SemiHOC) is the closest prior work. It shows that confidence does not track correctness at internal tree nodes. It responds by assigning unlabeled samples to subtrees through summed softmax probabilities and by freezing assignments with "age-gating", all in a flat Euclidean feature space. We keep its target, partial-depth pseudo-labels, but change three things. The source of the depth signal becomes geometric rather than confidence-based. The depth mask becomes cumulative, so committing to a leaf requires committing to every ancestor. Thresholds are calibrated per level against a supervised floor rather than set by a heuristic.
Two papers supply methods we reuse. Don't fear the unlabelled: safe semi-supervised learning via debiasing (DeSSL) makes the unlabeled loss unbiased by subtracting the same surrogate loss computed on labeled data. We apply that correction separately at each hierarchy level. Its guarantee assumes that which examples got labels was decided at random, independently of the data. Class mismatch breaks that assumption, and depth gating is our partial answer to that weakness, not something DeSSL already covers. Thermometer Encoding: One Hot Way To Resist Adversarial Examples supplies the cumulative code. That paper applies the code to pixel intensities. We apply it to a learned scalar, hyperbolic radius, and the bits gate losses rather than feed a network. The file's catalog of thermometer work places such codes only on raw inputs or scalar attributes. We also inherit two known weaknesses of the code: rigid uniform bins and sharp behavior at bin edges. Calibrated thresholds address the first. The second is measured directly, by sensitivity to small shifts in the thresholds.
Two further papers constrain the design. Not All Parameters Should Be Treated Equally (SPL) argues that filtering unlabeled data at the sample level leaks hard unseen-class examples, and it protects network weights with a mask learned by two nested optimizations instead. Our method is still sample-level filtering, softened from "keep or drop" to "how deep". SPL's critique therefore predicts a specific failure: unseen samples that land at large radius. We measure that leakage directly. PE-HCM uses hyperbolic margins with cluster-based pseudo-classes but has no fallback when the clusters are wrong. We differ by making the fallback, stopping at a shallower level, the central mechanism.
Research Program
Model. The model is a standard image backbone whose output is mapped into a Poincaré ball with the exponential map, the standard projection from flat coordinates onto the curved space. The norm is clipped below 1 to avoid the numerical blow-up near the boundary that the file documents. Prototypes for every taxonomy node come from a frozen Poincaré embedding of the label tree. Labeled images receive a hyperbolic-distance loss to their leaf prototype and a cross-entropy loss at every level $\ell = 1,\dots,D$.
Pseudo-label construction. Each unlabeled image $u$ with embedding $z_u$ gets a radius $r_u = d_{\mathbb{B}}(0, z_u)$, its hyperbolic distance from the origin. Its depth code is
$$b_\ell(u) = \mathbb{I}(r_u \ge \tau_\ell), \quad \ell = 1,\dots,D,$$
where $\tau_1 < \dots < \tau_D$ are per-level radius thresholds. The pseudo-label at level $\ell$ is the ancestor at that level of the nearest leaf prototype.
Per-level debiased objective. The unlabeled term at level $\ell$ follows DeSSL's form:
$$\mathcal{L}_\ell = \frac{1}{|U|}\sum_{u \in U} b_\ell(u)\, H_\ell(u) \;-\; \frac{1}{|L|}\sum_{j \in L} b_\ell(j)\, H_\ell(j),$$
where $H_\ell$ is the pseudo-label cross-entropy at level $\ell$. The second sum applies the same gate and the same surrogate loss to labeled images. Each $\mathcal{L}_\ell$ is added to the supervised loss with a weight $\lambda$.
Threshold calibration. Each $\tau_\ell$ is set on a held-out labeled split. It is the smallest radius such that level-$\ell$ pseudo-label accuracy among held-out points above it exceeds the supervised model's level-$\ell$ accuracy by a margin. If no such radius exists, level $\ell$ is switched off entirely. This per-level floor is the safety argument. It is empirical, not a proof, because the held-out split contains no unseen classes. The decisive checks below test whether it holds anyway.
Baselines
Supervised-only reference. The do-nothing reference is the same backbone and hyperbolic head trained only on labeled data. This defines the floor at every level. Any setting where a semi-supervised method falls below it, at any level, counts as degradation.
SemiHOC. SemiHOC is the strongest competitor on exactly this question: hierarchical pseudo-labels under open-set unlabeled data. We rerun it on its own benchmarks (iNaturalist19, iNaturalist21-Aves, SimpleHierImageNet) with the same backbone, the same labeled splits (including 20 labels per class), the same number of epochs, and the same tuning budget as our method. The file reports only that SemiHOC is "comparable to a fully supervised model trained with all ID labels" on iNaturalist19. No reported result found on Lacuna for SemiHOC's numeric accuracy on iNaturalist19, iNaturalist21-Aves or SimpleHierImageNet, so our reproduction serves as the reference number. Beating SemiHOC means two things together. We need higher mean accuracy at the deepest level. We also need strictly fewer (level, seed, mismatch ratio) cells below the supervised floor, with no loss of coarse-level accuracy.
DeSSL with FixMatch. DeSSL applied to FixMatch with flat leaf labels is the strongest safety-oriented competitor. The file reports it improving worst-class CIFAR-10 accuracy by nearly 5% over FixMatch.
SPL. SPL is the strongest method for class-mismatch safety. It reports raising Curriculum Labeling from 82.1% to 84.1% on CIFAR-10 at 60% mismatch. We run it on CIFAR-100 with superclasses, using its own protocol for mismatch ratios, so the comparison stays on its home ground.
Ablations. Four ablations isolate the key ingredients.
The first is a confidence gate. It replaces radius with the per-level summed softmax confidence, using identical calibrated thresholds, cumulative mask and debiasing. Running it in both the hyperbolic and a Euclidean head separates "geometry" from "calibration plus cumulative mask".
The second is a one-hot gate. Each image supervises a single selected level instead of all levels up to it, which tests whether the cumulative structure matters.
The third uses uniform thresholds, which tests whether calibration matters.
The fourth removes the debiasing term.
Why not just…?
Why not just use SemiHOC's softmax subtree confidence? This is the real threat. If summed softmax confidence, calibrated per level, predicts the deepest correct ancestor as well as radius does, then the hyperbolic part adds nothing. SemiHOC's own finding that confidence is unreliable at internal nodes is suggestive. However, it was measured with uncalibrated thresholds, so it does not settle the question. The confidence-gate ablation settles it. We compare how well each signal ranks unlabeled images by their true deepest-correct depth, measured as area under the curve per level, holding everything else fixed. If confidence matches radius, we drop hyperbolic geometry and report the method as calibrated cumulative depth gating in Euclidean space.
Why not just use a high-dimensional Euclidean embedding with its norm as the depth signal? Flat space has no inward-bending shortest paths, so nothing forces ambiguous images toward a parent-like position. The file's evidence that flat space handles hierarchies worse comes from graph embedding: on WordNet with 10% of edges, Hyperbolic Entailment Cones reaches 80.1 F1 in 5 dimensions against about 70 for Euclidean baselines. That evidence does not transfer automatically to image features. We therefore run the Euclidean-norm gate at 512 dimensions as a direct comparison.
Why not just apply DeSSL or SPL with flat labels? Both protect against degradation by keeping unseen-class images away from the classifier. Neither can extract the correct coarse supervision such images carry. If our method beats them only at coarse levels and ties at leaves, that is still the claimed benefit. If it loses at leaves, the depth gate is leaking, which is exactly SPL's critique of sample-level filtering.
Decisive Checks
The first check comes before any training comparison. On iNaturalist19 with held-out unseen species, we test whether hyperbolic radius ranks unlabeled images by their true deepest-correct ancestor better than calibrated confidence does. If it does not, the hyperbolic half of the proposal is abandoned as described above.
The second check measures leakage. We record the fraction of unseen-class images whose thermometer code switches on supervision below their deepest true ancestor, as a function of the mismatch ratio. The same run is repeated on CIFAR-100 with 20 superclasses at SPL's mismatch ratios and on BREEDS LIVING-17. For context on BREEDS, PE-HCM reports 53.09% all-way accuracy there, though under a coarse-to-fine protocol rather than ours.
The third check is a degradation count. We tally the cells below the supervised floor at each level, alongside the open-environment robustness measures of Realistic Evaluation of Semi-supervised Learning Algorithms in Open Environments as the mismatch ratio sweeps from 0 to 1.
The fourth check measures sensitivity to the bin edges. We shift each $\tau_\ell$ by small amounts and record how accuracy changes, to detect the sharp cliff behavior reported for thermometer codes.
References
- Poincaré Embeddings for Learning Hierarchical Representations
- Hyperbolic Space with Hierarchical Margin Boosts Fine-Grained Learning from Coarse Labels
- Realistic Evaluation of Semi-supervised Learning Algorithms in Open Environments
- Semi-Supervised Hierarchical Open-Set Classification
- Don’t fear the unlabelled: safe semi-supervised learning via debiasing
- Thermometer Encoding: One Hot Way To Resist Adversarial Examples
- Not All Parameters Should Be Treated Equally: Deep Safe Semi-supervised Learning under Class Distribution Mismatch
- Hyperbolic Entailment Cones for Learning Hierarchical Embeddings