50. Quantum Machine Learning
50.1What Quantum Machine Learning Actually Means
Four distinct research programs share the brand. One: QML on classical data — quantum circuits as models for ordinary datasets (this chapter's main subject; most hyped, least evidenced). Two: quantum data — learning models of states produced by quantum systems (not just defensible but arguably necessary: the data is natively quantum; honest current home: physics experiments, not products). Three: quantum speedups for ML subroutines — HHL-flavored linear algebra, gradient descent acceleration (largely dequantized or caveated to death — Ch. 49's HHL lessons). Four: ML for quantum — classical ML applied to quantum problems (decoding, calibration, pulse design — quietly the most productive branch, and very much a software engineer's game). Confusing these four is the root of most QML arguments; label your claims.
50.2Quantum Feature Maps
The load-bearing idea: a circuit Φ(x) maps classical input x into a quantum state |Φ(x)⟩, and the quantum feature space — the Hilbert space of these states — may contain structure (kernel values ⟨Φ(x)|Φ(y)⟩) hard to compute classically. Design knobs: encoding style (angle encoding: xᵢ as rotation angles — 1 qubit per feature; amplitude encoding: n features in log n qubits — deep state-prep circuits, the caveat that kills many claims), repetition (re-uploading layers), entangling structure. The provable side: Havlíček et al. (2019) constructed feature maps whose kernels are classically hard under assumptions; the cautionary side: those constructions are deliberately worst-case, and whether natural datasets have quantum-useful structure is an empirical question with, so far, mostly negative or inconclusive answers.
50.3Variational Classifiers
A variational circuit (49.2) as a classifier: encode input → parameterized layers V(θ) → measure → interpret bits as logits/class. Train with the same hybrid loop as VQE, cross-entropy loss, parameter-shift gradients — and inherit every pathology: barren plateaus (now under a loss, with data-dependence making theory harder), noise bias, measurement overhead per gradient. The comparison that must precede any claim: classical baselines of honest size (a tuned logistic regression, a small MLP, a kernel method on the same features) — because QML models of 2020–2024 vintage were routinely beaten by linear models on the datasets they were demonstrated on. Run that comparison yourself once (Ch. 50.8's protocol); it calibrates your skepticism permanently.
50.4Quantum Kernels
The more mathematical branch: skip the variational training, compute the kernel matrix K_ij = |⟨Φ(xᵢ)|Φ(xⱼ)⟩|² via circuit swaps, and hand it to a classical SVM. Elegant properties: training is convex (no plateaus — the non-convexity moved into kernel design), generalization bounds exist, and the whole thing connects to classical kernel theory where "quantum advantage = a kernel hard to estimate classically but useful on real data" is a crisp research question. The overhead reality: kernel estimation costs O(m²) circuit evaluations for m samples — m=1000 is 10⁶ evaluations; every realistic dataset size crushes the approach on hardware economics alone. Kernel methods are the intellectually cleanest corner of QML and the economically furthest from application.
50.5QNN Architectures
"Quantum neural networks" — the name is marketing; the objects are parameterized circuits. Architecture questions with no classical analogy: how to interleave data encoding and trainable layers (re-uploading beats encode-once empirically), what "convolution" or "pooling" means when you can't copy (qPCNs, tree tensor networks), how to read out (expectation values, not spikes), and how to regularize (symmetry constraints, again). Theory tells us QNNs are not more expressive than classical networks on classical data in any exploitation-ready way (any circuit's function can be simulated at exponential cost — and "efficiently representable but hard to train" claims remain unproven for natural data). Architectural research is legitimate and interesting; architectural branding is where your skepticism should live.
50.6Data Loading Bottlenecks
The structural problem under all of QML-on-classical-data: getting x into |Φ(x)⟩. Amplitude encoding's O(n) gates for n features negate every algorithmic speedup on any dataset that fits classical memory; angle encoding costs O(n) qubits; QRAM — the hypothetical device that loads data in O(log n) — does not exist and its no-go discussions are serious (Ch. 46's analog). So every "exponential speedup" QML result carries a footnote that, instantiated, costs more than the classical algorithm it beats. This is not a detail; it is the central obstacle, and any QML paper, pitch, or roadmap that doesn't address data loading explicitly has not engaged with its own feasibility. Memorize the argument; you will use it in meetings.
50.7Barren Plateaus
Ch. 49.6's result, sharpened for QML: data-encoding gates interact with the trainability question (data re-uploading helps; encoding-rich ansätze hurt), and the 2023–24 noise results land with double force since QML targets noisy hardware by definition. QML-specific mitigations under study: problem-informed encodings, ansätze drawn from the data's symmetry class, and trainable encodings (meta-learning what to load). The meta-lesson for your career: QML is where the field's mathematical maturity gets tested in public — trainability theory moved from folklore to theorems in five years, and the theorems were mostly negative. Reading that literature closely is the best possible training for evaluating any quantum pitch you'll ever see.
50.8Classical Baselines
The protocol, because you will need it: (1) fix dataset and split first; (2) establish the classical bar — tuned logistic regression, gradient-boosted trees, and one small neural net, all with honest hyperparameter search; (3) run the quantum model with equal compute budget (not equal epochs — equal wall-clock and equal tuning effort); (4) report learning curves, not final accuracies; (5) report shot counts and hardware assumptions in the table; (6) pre-register the metric. This is boring, and it is the difference between a result and a demo. The uncomfortable empirical record of 2020–2025: on standard tabular and image datasets, quantum models have not beaten well-tuned classical baselines, and the papers that said so clearly are the field's most valuable reading.
50.9Avoiding Quantum-AI Hype
The synthesis discipline. Claims to refuse: "quantum computers will revolutionize AI" (no mechanism known), "exponential speedups for machine learning" (dequantized or data-loading-caveated), "QNNs are like brains" (no). Claims to take seriously: quantum methods for quantum data (state learning, verification — live research), ML for quantum engineering (decoders! Ch. 37's neural decoders are real and shipping), and QML as a theory program clarifying what learning means without cloning or efficient classical simulation. Career guidance embedded: the engineers most valuable to QML teams are those who can build the classical infrastructure and the honest baselines — a job description that matches exactly the skills this book has been building.