Zero-shot learning asks a model to recognize classes it never saw during training. It does this by predicting a vector of human-annotated attributes ("has stripes", "lives in water") from an image and matching that vector against the known attribute descriptions of new classes. The map from images to attributes is fit only on seen classes, so it drifts when applied to unseen ones. This failure is called projection domain shift. The strongest fixes are transductive: they also look at the unlabeled images of unseen classes and pull the model's predictions on those images into agreement with the unseen class descriptions. They do this by matching distributions, which quietly assumes a guess about how often each unseen class occurs. The best such method reports that it degrades significantly when unseen classes are unbalanced and the true class frequencies are unknown (Bi-Directional Distribution Alignment for Transductive Zero-Shot Learning).
We propose aligning only what does not depend on class frequencies. Team semantics is a branch of logic that evaluates formulas on a whole table of rows (a "team") instead of on a single row. It sharply separates two kinds of table properties. Some depend on how often each row occurs: the probabilistic atoms, such as "these two columns have the same distribution." Others depend only on which rows occur at all: the qualitative atoms, such as inclusion, "every value of x also appears as a value of y." The unseen class-attribute table is given in zero-shot learning. We therefore constrain the team of predicted attribute vectors on unlabeled unseen images to have the same support, meaning the same set of rows that occur, as that table. Every prediction should land on some unseen class description, and every unseen class description should be claimed by some image. These two inclusion constraints are, by construction, unaffected by any positive class prior. That makes them a principled substitute for prior estimation exactly in the imbalanced regime where current methods break.
Core Claim
In transductive zero-shot learning, alignment objectives that match marginal distributions depend on the unknown unseen class prior. A two-sided inclusion constraint does not. Its forward direction says predicted attributes lie within the unseen class rows (containment). Its reverse direction says each unseen class row is reached by some prediction (coverage). It is the team-semantic statement that predictions and class descriptions have equal support. Replacing the marginal-matching and cluster-prior components of a strong transductive model with this constraint should leave balanced-benchmark accuracy roughly unchanged. It should also substantially reduce the accuracy drop as unseen classes become long-tailed.
The new element is not a new loss family. It is the decision about which team properties to align, grounded in the qualitative-versus-probabilistic distinction, together with a test of whether that choice matters under imbalance. A secondary, more speculative claim is that model-checking the seen and unseen class tables for dependence atoms can flag attributes at risk of shortcut learning. Such an atom might hold on seen classes but fail on unseen ones, as in the co-occurrence trap.
Where This Sits
Probabilistic team semantics (Probabilistic Team Semantics) introduced the marginal identity atom $x \approx y$ and conditional independence over weighted teams. Facets of Distribution Identities in Probabilistic Team Semantics showed that marginal identity is strictly weaker than its multiset variant and conditional independence. Inclusion atoms have been studied in the qualitative setting (Complexity of Propositional Independence and Inclusion Logic). These papers supply the vocabulary and the key distinction: qualitative atoms are insensitive to row multiplicities, probabilistic atoms are not. None of them touches learning.
They also constrain the design. Model checking with these atoms is NP-complete in general. We therefore use only quantifier-free atoms evaluated on explicit finite tables of tens to hundreds of class rows, which is polynomial. The hardness in the logic comes from splitting teams under disjunction, which we never do.
On the learning side, Bi-VAEGAN (Bi-Directional Distribution Alignment for Transductive Zero-Shot Learning) trains a generator from attributes to image features and a regressor from features back to attributes. It aligns them on unlabeled unseen data with adversarial critics and a K-means-based cluster prior estimate. Its own admitted weakness under unbalanced priors motivates this proposal, and we reuse its architecture as a host. Visual-structure and visual-center methods (Transductive Zero-Shot Learning with Visual Structure Constraint; Transductive Zero-Shot Learning via Visual Center Adaptation) align projected class centers with unlabeled clusters. Clustering implicitly favors comparably sized clusters, so these methods share the prior sensitivity we target.
Two methods address entanglement between attributes, which our secondary diagnostic concerns. Semantics Disentangling for Generalized Zero-Shot Learning separates semantic from non-semantic features with a total-correlation penalty. DUET: Cross-Modal Semantic Grounding for Contrastive Zero-Shot Learning names the attribute co-occurrence trap and breaks it with attribute-level contrastive sampling. Both are inductive and orthogonal to prior sensitivity.
LGML: Logic Guided Machine Learning shows that logical invariants can guide learning, but only invariants checked on single instances. Our constraints are properties of whole collections, which single-instance checking cannot express.
Research Program
Let $r_\theta(x)$ be the attribute vector the model regresses from image features $x$. Let $\mathcal{A}_U = \{a_c\}$ be the given attribute rows of the unseen classes. Let $B$ be the pool of unlabeled unseen images, held in a memory bank so rare classes are not missed by small batches. Let $\mathrm{softmin}_\tau$ be a temperature-$\tau$ smooth minimum (negative scaled log-sum-exp). The two constraints are
$$L_{\text{contain}} = \frac{1}{|B|}\sum_{x \in B} \mathrm{softmin}_{\tau,\,c}\, \lVert r_\theta(x) - a_c \rVert^2, \qquad L_{\text{cover}} = \frac{1}{|\mathcal{A}_U|}\sum_{c} \mathrm{softmin}_{\tau,\,x \in B}\, \lVert r_\theta(x) - a_c \rVert^2.$$
As $\tau \to 0$, these measure violation of the inclusion atoms $r_\theta(B) \subseteq \mathcal{A}_U$ and $\mathcal{A}_U \subseteq r_\theta(B)$. Together they measure violation of equal support.
The invariance argument is short. Take the population versions of the two losses, where the batch average is replaced by an expectation over images. Multiplying a class's frequency by any positive factor rescales its images' terms in $L_{\text{contain}}$. It does not change which predictor drives the loss to zero, because zero requires every image to sit on some row regardless of weights. $L_{\text{cover}}$ is a minimum over images, so it changes only if a class's frequency reaches zero. A marginal-matching loss is different: its minimizer moves when the prior moves. This invariance holds at the zero-violation optimum. The finite-temperature, finite-sample objective only approximates it, and measuring how close the approximation is forms part of the study.
For generalized zero-shot learning, where seen and unseen classes compete at test time, containment is taken over seen and unseen rows together. Coverage is imposed on unseen rows. This directly counters seen-class bias: every unseen class must claim at least one image.
We build this by replacing Bi-VAEGAN's cluster prior estimation, and its unseen-side marginal critic, with $L_{\text{contain}} + \lambda L_{\text{cover}}$. Its generator, regressor, seen-class training and L2 feature normalization stay unchanged. Any difference is then attributable to the alignment term. We also test the constraints on a plain regressor with no generator, to see whether they carry the gain alone.
The secondary diagnostic model-checks dependence atoms $\mathrm{dep}(a_i, a_j)$ and pairwise co-occurrence support on the binarized seen table and the unseen table. It flags atoms that hold on seen classes and fail on unseen ones. We then ask whether flagged attributes show larger per-attribute prediction error on unseen images. CUB and SUN provide image-level attributes for this check.
Baselines
The naive reference is the same regressor and generator trained inductively on seen classes only, with no use of unlabeled unseen images. It shows how much any transductive alignment buys at each imbalance level. A second naive reference adds unseen-side alignment that assumes a uniform prior, which is what one does absent any prior estimate.
The strongest existing method is Bi-VAEGAN with cluster prior estimation. It is the most direct competitor because it targets the same transductive setting on the same benchmarks and is the method whose acknowledged failure mode we attack. On balanced proposed splits it reports transductive top-1 accuracy of 95.8% and generalized transductive accuracy of 90.4% on AWA2, 93.9% and 89.1% on AWA1, and 74.2% and 54.1% on SUN (Bi-Directional Distribution Alignment for Transductive Zero-Shot Learning). It will be reproduced in our codebase with its published settings, and we will confirm we approach these numbers before running imbalanced variants. Both it and our method get the same features, unlabeled pools, tuning budget over their respective loss weights and temperatures, and seeds. We also run it with the oracle true prior as an upper reference. We found no reported result on Lacuna for Bi-VAEGAN or any transductive zero-shot method under controlled long-tailed unseen priors, so that regime is measured fresh.
For inductive context we include SDGZSL, which reports a harmonic mean of 75.1% on CUB (Semantics Disentangling for Generalized Zero-Shot Learning), and DUET, which reports 45.8% harmonic mean on SUN (DUET: Cross-Modal Semantic Grounding for Contrastive Zero-Shot Learning), using its released code (zjukg/duet). These methods do not see unlabeled unseen data, so they place transductive gains in context rather than compete head-on.
Three ablations isolate the key ingredient. Containment only is essentially nearest-prototype self-training. Coverage only is the second ablation. The third keeps both but swaps the support constraint for a maximum mean discrepancy between predicted attributes and uniformly weighted unseen rows. That is a marginal-matching term of matched cost and form.
Beating the strongest baseline means the following. At the most severe imbalance we test, our method's unseen accuracy and harmonic mean should exceed Bi-VAEGAN with estimated prior by a margin larger than the seed-to-seed spread on at least three of the four datasets. It should also close a substantial share of the gap to the oracle-prior variant. On balanced splits it should stay within about two points of the reproduced Bi-VAEGAN.
Why not just…?
Why not just use a mutual-information or kernel penalty such as MMD or HSIC, as SDGZSL does with total correlation? These are distribution-level measures, so aligning with them inherits the prior sensitivity we are trying to remove. The matched-MMD ablation tests this directly. If MMD with uniform weights degrades as little as the support constraint under imbalance, our central claim fails and we will report that. Total correlation in Semantics Disentangling for Generalized Zero-Shot Learning targets a different problem, independence between feature parts, and is compatible with our term rather than a replacement for it.
Why not just estimate the unseen prior better and keep marginal matching? This is a serious alternative. If the prior could be estimated accurately, marginal matching would be at least as informative as support matching, since it uses more information. The oracle-prior run bounds what perfect estimation could achieve. The gap between the estimated-prior and oracle runs shows how much estimation actually costs. Bi-VAEGAN's authors report that cluster-based estimation breaks on unbalanced data. Our claim is that support alignment recovers most of the oracle's accuracy without estimating anything. If a better estimator closes the gap more cheaply, that is a competing answer and we will say so.
Why not just do pseudo-label self-training or entropy minimization on unseen images? Containment alone is close to this, and it is known to collapse onto a few dominant classes or hubs. Coverage is the piece self-training lacks. The containment-only ablation is exactly this baseline, so the experiment directly shows whether coverage adds anything. If containment alone matches the full method, the team-semantic framing reduces to self-training and the contribution shrinks accordingly.
Why not just use a linear compatibility model with L2 feature normalization? Zero-Shot Learning - A Comprehensive Evaluation of the Good, the Bad and the Ugly. shows that such models are competitive at scale but suffer from seen-class bias in the generalized setting. Bi-VAEGAN reports that L2 normalization is a large free gain, and we keep it in every run. We also attach our constraints to a plain linear regressor. This tests whether the benefit requires a generator at all, which would answer this objection directly.
Why bring logic in at all, if the loss is a pair of soft minima? The honest answer is that logic contributes the choice, not the arithmetic. The distinction between qualitative and probabilistic atoms tells us which collection-level properties survive an unknown reweighting. For the diagnostic, dependence atoms ask whether any seen row violates a determination relation, not merely whether two attributes correlate. We compare the diagnostic against the simple alternative of flagging attributes whose seen and unseen correlation differs most. If that simple flag predicts per-attribute error as well, we will present the diagnostic as a formal restatement rather than a better tool.
Proposed Experiments and Decisive Checks
We use the proposed splits of AWA1, AWA2, CUB and SUN from Zero-Shot Learning - A Comprehensive Evaluation of the Good, the Bad and the Ugly., with its standard features and attribute vectors, reporting unseen top-1 accuracy and the harmonic mean of seen and unseen accuracy. Imbalance is introduced by subsampling unseen images to Zipf-distributed class frequencies at several head-to-tail ratios, with five random class orderings per ratio. This keeps imbalance from being confounded with which classes are rare.
The decisive check is the slope of accuracy against imbalance ratio. If our method's slope is clearly flatter than Bi-VAEGAN with estimated prior and than the matched-MMD ablation, while balanced accuracy is preserved, the claim holds. If the slopes are indistinguishable, support alignment offers no advantage and the idea fails.
A second check probes the invariance argument itself. It measures how our method's accuracy changes with temperature and memory-bank size at fixed imbalance. The prediction is that lower temperature and a full-pool bank make the method more prior-invariant, and that small batches erode coverage for the rarest classes. We also report per-class accuracy for tail classes, since the advantage should concentrate there.
For the secondary diagnostic, we compute the rank correlation between atom-violation flags and per-attribute unseen error on CUB and SUN, against the correlation-difference flag. This is judged independently of the main result.
References
- Bi-Directional Distribution Alignment for Transductive Zero-Shot Learning
- Probabilistic team semantics
- Facets of Distribution Identities in Probabilistic Team Semantics
- Complexity of Propositional Independence and Inclusion Logic
- Transductive Zero-Shot Learning with Visual Structure Constraint
- Transductive Zero-Shot Learning via Visual Center Adaptation
- Semantics Disentangling for Generalized Zero-Shot Learning
- DUET: Cross-Modal Semantic Grounding for Contrastive Zero-Shot Learning
- LGML: Logic Guided Machine Learning
- Zero-Shot Learning - A Comprehensive Evaluation of the Good, the Bad and the Ugly