When labels are expensive, as with a wet-lab hemolysis assay for a peptide or a toxicity test for a molecule, active learning lets the model choose which examples to label. One of its oldest and most effective forms is co-testing: train two classifiers that look at the data in genuinely different ways, and ask the oracle about the points where they disagree. At such a point at least one of them must be wrong, so the label is informative. Co-testing's weak spot is that it needs two such "views" of the data, and in most domains nobody knows how to build them. Quantum feature maps are circuits that embed a classical input into the state of a few qubits, and similarity between two inputs is measured as the overlap of their quantum states. One of their best-documented properties is that the similarity structure they produce differs from what classical kernels produce. This proposal uses that difference as the second view. A classical kernel classifier and a quantum kernel classifier are trained side by side on the same small labeled set, and only points where they disagree are sent to the oracle, provided the disagreement is larger than the quantum model's own measurement noise.
The point is not a quantum speedup. All experiments run at qubit counts that can be simulated classically. The claim is narrower and checkable. A quantum kernel is useful in an active learning loop to the extent that its geometry differs from the classical kernel, and we can measure that difference before spending any labels. Kernel-based co-testing also changes the cost structure. Each newly labeled point requires one row of quantum overlaps against the pool, not the full pairwise matrix. The quantum measurement bill therefore scales with the label budget rather than with the square of the pool size.
Core Claim
Calibrated quantum–classical co-testing should reach a given accuracy with fewer oracle labels than uncertainty sampling with either kernel alone. The size of that saving should be predicted by a label-free measure of how far apart the two kernels' geometries are. The method has to beat a strong control that is entirely classical: co-testing between two different classical kernels. If two classical kernels that disagree as much as the quantum and classical pair give the same savings, then the quantum map contributes nothing special, and we will report that.
Grounding in Prior Work
Prior work plays three roles here.
The first is motivation. Several quantum-kernel papers find that quantum maps separate classes that classical kernels do not, all in the passive setting where every label is already available. Quantum Feature Maps for Graph Machine Learning on a Neutral Atom Quantum Processor builds a graph kernel from the evolution of neutral atoms placed on the graph's layout. It reports an F1 of 60.4% on PTC-FM, above graphlet and random-walk kernels, and calls the resulting similarity structure geometrically original. Non-hemolytic peptide classification using a quantum support vector machine encodes 40 peptide descriptors into 6 qubits through Hamiltonian evolution. It reports 74.26% accuracy on HemoPI-2 against 72.28% for an RBF kernel, and 76.00% against 75.70% on HemoPI-3. Neither paper selects which examples to label, although hemolysis assays are exactly the kind of expensive oracle active learning exists for. These small accuracy margins also constrain what we can claim. A quantum kernel that is barely more accurate can still be valuable as a second view if it is wrong on different points, and that difference is the quantity we measure.
The second role is supplying the method. Selective Sampling with Redundant Views introduced co-testing and showed that querying points where per-view classifiers disagree beats query-by-bagging and query-by-boosting. Active Learning with Strong and Weak Views extended it to imperfect views and reduced labels by 45% to 81% on wrapper induction. The same paper is also a warning: when one view is much worse than the other, disagreement queries become skewed. Both papers split classical attributes to get their views. We replace attribute splitting with a change of embedding. That substitution is untested, and it is exactly where the "view detection" problem of co-testing is open.
The third role is constraint. Towards Efficient Quantum Anomaly Detection: One-Class SVMs using Variable Subsampling and Randomized Measurements shows that full quantum Gram matrices scale quadratically, projecting roughly 255 years on hardware for about 284,000 transactions. It works around this with uniform random subsampling, and it observes that randomized-measurement kernels have higher variance than RBF. That variance is the central danger for us. If quantum–classical disagreement comes from shot noise, co-testing will spend labels on noise. Evaluating Parameter-Based Training Performance of Neural Networks and Variational Quantum Circuits reports that trainable quantum circuits take 30 minutes on hardware for Iris against 1.8 seconds for a neural network. Repeatedly retraining such circuits inside an active learning loop is therefore impractical. We use fixed quantum kernels with classical SVM training, so each round only adds kernel rows and never retrains a circuit.
Finally, active learning on quantum experiments is not what we do. Active learning for quantum mechanical measurements, Active learning machine learns to create new quantum experiments, and Bayesian Phase Estimation via Active Learning apply active learning to quantum hardware. None use a quantum embedding as the representation that decides which classical data to label.
Research Program
The loop maintains two SVMs on a shared labeled set $L$. The first uses a classical kernel: RBF on descriptors for peptides, and a Weisfeiler–Lehman or graphlet kernel for molecules. The second uses a quantum kernel: the Pauli-string Hamiltonian map of the peptide paper, or a simulated quantum evolution kernel for graphs. We estimate quantum kernel entries from a finite number of shots $S$, matching how they would be measured on hardware.
We resample those shots to get a bootstrap standard deviation $\sigma_Q(x)$ of the quantum SVM's decision value $f_Q(x)$ at each pool point. A pool point $x$ counts as a calibrated contention point when
$$ \mathrm{sign}\, f_C(x) \neq \mathrm{sign}\, f_Q(x) \quad \text{and} \quad |f_Q(x)| > z\,\sigma_Q(x), $$
where $f_C$ is the classical SVM's decision value and $z$ is a confidence multiplier fixed in advance (for example 2). In words, the quantum model must disagree with the classical one and be confident beyond its own measurement noise.
Each round selects a batch of contention points. Within the batch, points are ranked by the smaller of the two normalized margins, which is meant to avoid the strong-view/weak-view skew Muslea and colleagues observed. When there are no contention points, the loop falls back to uncertainty sampling on whichever view currently has the better cross-validated accuracy. After labeling, the loop computes the new points' overlaps with the pool, adding $b \times M$ entries per round for batch size $b$ and pool size $M$, and retrains both SVMs classically.
Before any labels are spent, we compute a label-free view-difference score on the unlabeled pool. We use one minus the centered alignment between the two Gram matrices, computed on a random subsample. This alignment is the normalized inner product of the two centered matrices, so the score is zero when the kernels agree up to scale. We compute the same score for pairs of classical kernels. The program's predictive claim is that label savings over single-view uncertainty sampling increase with this score, whichever kind of kernel supplies the second view.
Baselines
The naive reference is random sampling with the quantum SVM, with the classical SVM run under the same random draws. This is what the quantum-kernel papers implicitly do, and it is the passive analogue of variable subsampling in the anomaly-detection paper. A method that does not clearly beat random sampling at small budgets has failed.
The strongest single-view competitor is margin-based uncertainty sampling with the classical kernel SVM, followed by the same strategy with the quantum kernel SVM. The classical version is what a practitioner would deploy. It is deterministic, cheap, and starts from a nearly equal passive accuracy (72.28% versus 74.26% on HemoPI-2, per the peptide paper). We also run classical Kernel Query By Committee from Query by Committee Made Real. That paper samples a committee of consistent hypotheses inside the span of the labeled points and the candidate, and reports about 0.7% error against about 6% for a passive SVM at 120 labels on 15-dimensional synthetic data, and a 3.6% average gain at 140 labels on AR Face. We found no reported result on Lacuna for any active learning method on PTC-FM or HemoPI, so all active baselines will be run by us.
The control that matters most is classical–classical co-testing. It uses the identical calibrated-contention rule, but both views are classical kernels chosen to be as different as possible: RBF on descriptors against a sequence-composition kernel for peptides, and Weisfeiler–Lehman against random-walk kernels for graphs.
Two ablations isolate the proposal's ingredients. The first is quantum–classical co-testing with the noise calibration removed (plain sign disagreement). The second is the full method at several shot budgets $S$, which shows how much of any gain survives realistic measurement noise.
Fairness is enforced in four ways. All methods share the same seed sets of 10–20 labeled points, the same batch sizes, and the same label budgets. Hyperparameters, meaning SVM regularization, kernel bandwidths, and quantum evolution times, are tuned by the same small cross-validation on the seed set only. Twenty random seeds per configuration are run. Quantum methods are also charged for their kernel evaluations in a reported second axis.
Beating the baselines means the full method's area under the learning curve exceeds that of classical uncertainty sampling, KQBC, and classical–classical co-testing, with seed-level confidence intervals that do not overlap. Concretely, the method should reach the passive full-data QSVM accuracy (74.26% on HemoPI-2) or the QEK F1 (60.4% on PTC-FM) with materially fewer labels than any baseline.
Why not just…?
Why not just run uncertainty sampling with a classical RBF SVM? It may well be enough, given that the reported passive gaps are 0.30 to 1.98 accuracy points. Our argument is not that the quantum model is more accurate. It is that a second model with different geometry finds the points where the first is confidently wrong, which margin sampling cannot do by construction. The head-to-head learning curves settle this. If the classical SVM matches us, the quantum view adds nothing in this regime.
Why not just co-test two classical kernels? This is the real threat, and the proposal is built around it rather than around beating random sampling. The "geometric originality" reported for the neutral-atom kernel was measured as passive accuracy, not as error decorrelation with classical kernels. A well-chosen pair of classical kernels might disagree just as usefully. The view-difference score lets us compare the pairs on equal footing. If savings track that score regardless of whether a quantum kernel is involved, we will report that the quantum map is simply one more source of diverse views. The quantum-specific claim survives only if the quantum view supplies differences that no classical pair we try reaches.
Why not just subsample passively, as the anomaly-detection paper does? Variable subsampling cut training time by 95% on synthetic data and addresses compute cost. It does not address the number of oracle labels, which is the binding cost when each label is an assay. Random subsampling is our naive baseline, so this question is answered directly.
Why not just use foundation-model priors, as PriorAL does? Active Learning with Foundation Model Priors halves annotation on CIFAR-100 by combining CLIP-like models with a small task model. For 286 molecules or a few thousand peptides described by physicochemical descriptors, no comparable pretrained prior exists in the file. We therefore treat PriorAL as out of scope rather than as a defeated competitor.
Proposed Experiments
The first experiment uses the synthetic random-walk graph dataset from the neutral-atom paper, where the reported quantum lead over classical kernels varies from 7% to 21% with the noise level. This controlled sweep lets us test the predictive claim directly. Label savings from quantum–classical co-testing should rise and fall with both the view-difference score and that known passive gap, while classical–classical co-testing should not show the same pattern.
The second experiment uses PTC-FM with a simulated quantum evolution kernel, together with HemoPI-2 and HemoPI-3 with the 6-qubit Hamiltonian encoding. These are realistic expensive-oracle settings. We report learning curves, area under the curve, and labels needed to reach the passive full-data quantum score.
The third experiment varies the shot budget $S$ from a few hundred to the noiseless limit. It shows whether the noise calibration preserves the gains or whether the uncalibrated rule degrades into querying shot noise, the failure mode the anomaly-detection paper's variance finding predicts. For every configuration we report quantum kernel evaluations used. This shows that the cost scales with labels times pool size rather than with pool size squared.
The decisive check has two outcomes. The proposal stands if quantum–classical calibrated co-testing beats classical–classical co-testing at matched view-difference scores on at least the HemoPI and PTC-FM tasks, and if removing calibration hurts at realistic shot counts. If classical pairs match it, the result is still useful but different: quantum kernels are a convenient generator of co-testing views, with no special status.
References
- Quantum Feature Maps for Graph Machine Learning on a Neutral Atom Quantum Processor
- Non-hemolytic peptide classification using a quantum support vector machine
- Selective Sampling with Redundant Views
- Active Learning with Strong and Weak Views: A Case Study on Wrapper Induction
- Towards Efficient Quantum Anomaly Detection: One-Class SVMs Using Variable Subsampling and Randomized Measurements
- Evaluating Parameter-Based Training Performance of Neural Networks and Variational Quantum Circuits
- Active learning for quantum mechanical measurements
- Active learning machine learns to create new quantum experiments
- Bayesian Phase Estimation via Active Learning
- Query by Committee Made Real
- Active Learning with Foundation Model Priors: Efficient Learning under Class Imbalance