Overview
Curated: · Written: · Reviewed:
A useful quantum system is hybrid by construction. CPUs and GPUs ingest and clean data, formulate problems, optimize and compile circuits, schedule QPU work, manage retries, and post-process or verify results. A QPU executes a bounded quantum kernel. Tight control feedback, near-time mitigation, and scale-out preprocessing have different latency and placement needs, so architecture must expose data movement, queueing, synchronization, calibration windows, and failure recovery rather than draw one opaque quantum box.
Quantum machine learning includes quantum data processed by quantum algorithms, classical data encoded into quantum states, quantum-generated features or kernels consumed by classical models, and variational quantum models trained in classical loops. Encoding is part of the algorithm: angle, basis, amplitude, and repeated feature maps have different qubit, depth, normalization, and data-loading costs. An exponential Hilbert space is not proof of useful expressive power or classical intractability.
Quantum kernels require many pairwise circuit estimates and can produce a noisy, asymmetric, or non-positive-semidefinite empirical kernel. QNNs and variational classifiers inherit optimization, barren-plateau, finite-shot, noise, and gradient costs. Classical preprocessing can create most of the observed gain; small datasets and weak baselines invite overfitting. Dequantization asks whether a comparable classical method can reproduce the claimed benefit under similar data-access assumptions.
Production evaluation must use leakage-safe train/validation/test splits, nested model selection where appropriate, multiple seeds, uncertainty and calibration analysis, and strong classical baselines matched on preprocessing, parameter tuning, compute, latency, and total cost. The production invariant is hybrid contribution attribution: each quantum component earns inclusion only through a preregistered ablation and held-out end-to-end result that isolates its incremental benefit after encoding, orchestration, sampling, post-processing, and classical alternatives are fully accounted.
A quantum kernel method prices itself out on sample count long before it reaches an interesting dataset, and the arithmetic is worth doing before the experiment rather than after. Training a kernel model on 1,000 samples requires the Gram matrix, which is 499,500 distinct pairs; at 4,096 shots per entry that is about 2.05 billion circuit executions, and at a sustained ten thousand shots per second — optimistic once queueing and job overhead are included — the data-preparation step alone runs for roughly 57 hours of device time. Scaling is quadratic in samples, so 10,000 samples is a hundred times that. Worse, the estimated matrix is not the ideal kernel: finite shots make it noisy and asymmetric, and hardware noise can push its smallest eigenvalues negative, so the empirical Gram matrix is not positive semidefinite and the convex solver that a support vector machine relies on either fails or silently returns a solution to a different problem. The repairs — symmetrizing, clipping negative eigenvalues, or adding a ridge to the diagonal — are legitimate but they are modelling choices that change the estimator, so they belong in the reported method, and the regularization strength has to be selected inside the cross-validation loop rather than tuned against the test set.
The question a QML result has to answer is not whether it worked but which component made it work, and the only instrument for that is a preregistered ablation. A pipeline that encodes classical features through a parameterized feature map, runs a QPU, and feeds the output to a classical classifier has at least four candidate sources of accuracy: the preprocessing and dimensionality reduction applied before encoding, the feature map itself, the quantum-estimated kernel or expectation values, and the classical model on top. Replacing the quantum stage with a random feature map of the same dimension, or with a classical radial-basis kernel on the same preprocessed inputs, is the comparison that isolates the quantum contribution, and it frequently accounts for most or all of the reported gain. The methodological failures around this are ordinary machine-learning ones rather than quantum ones, which is why they are easy to miss in a quantum paper: a dataset of a few hundred samples with a model selected on the test split, a single seed reported out of several runs, a baseline that received no hyperparameter search while the quantum model received hundreds of QPU hours, and error bars taken over shots rather than over data splits. Fix the protocol first — leakage-safe nested splits, several seeds, uncertainty over splits, a baseline tuned with comparable effort and comparable wall-clock budget — and then report the incremental held-out benefit of the quantum stage with its cost beside it. A hybrid architecture also has to state where each stage runs and what happens when one fails, since a QPU task that returns after the classical optimizer has timed out is a correctness problem rather than a latency problem.
Data loading is the constraint that decides whether a quantum machine learning proposal is about quantum data or about classical data, and the two have very different prospects. Amplitude encoding packs 2^n features into n qubits, which is the source of most exponential-speedup intuitions, but preparing an arbitrary such state takes a circuit whose gate count is exponential in n unless the data has structure or a quantum random access memory exists — and no scalable QRAM has been built, so an algorithm that assumes constant-time state preparation has assumed away its dominant cost. Angle encoding avoids that by spending one qubit per feature, which makes preparation cheap and the qubit requirement linear in dimensionality, so a 500-feature dataset is out of reach on present hardware for a different reason. This is the honest framing for a design review: state the encoding, its qubit count, its preparation depth, and whether the speedup claim survives with preparation included. Quantum data — states produced by a physical experiment or by another quantum process — sidesteps the loading problem entirely, and is where the least contested near-term arguments for quantum learning sit, because there is no classical description to load in the first place. Dequantization results sharpen the same point from the other side: several proposed exponential advantages for classical data have been matched by classical algorithms given comparable sampling access to the input, which does not close the field but does mean that a claimed speedup has to name the classical access model it is beating.