Quantum Machine Learning: A Map
Quantum machine learning studies how quantum computers might represent, process, and learn from data using superposition, interference, and entanglement.
What the field actually is
Quantum machine learning (QML) is the intersection of quantum computing and statistical learning. The central hope is that a quantum processor can prepare and manipulate states in a Hilbert space whose dimension grows exponentially with the number of qubits, and that some learning problems become easier to represent or solve in that space. The central caution is that measurement collapses those states to a handful of classical bits, so the exponential room is not freely readable.
QML divides along two axes. The first is what is quantum: the data, the model, or both. Most near-term work uses classical data with a quantum model. The second is whether the quantum device is fault-tolerant or noisy. Fault-tolerant algorithms such as HHL assume error-corrected qubits that do not yet exist at scale; near-term methods target today's noisy machines.
The two main near-term families
- Kernel methods: encode data into quantum states and use the quantum device only to estimate inner products, feeding a classical support vector machine.
- Variational methods: use a trainable parameterized circuit optimized by a classical loop, as in variational classifiers and quantum neural networks.
Where the difficulty lives
Three obstacles recur across the field. Loading classical data into a quantum state (data encoding) can cost as much as the speedup it hopes to enable. Training suffers from barren plateaus, where gradients vanish exponentially. And many proposed advantages have been matched by clever classical algorithms, a process called dequantization.
Honest framing matters. There is no general theorem that quantum models learn better than classical ones on classical data. There are constructed problems with provable separations, and there are heuristics that may help on structured or quantum-native data. Reading QML claims well means asking which regime, which hardware assumption, and which classical baseline is being compared against.
The rest of this section builds the machinery piece by piece: how data enters a circuit, how circuits are parameterized and trained, why training is hard, what generative and kernel variants look like, and how to weigh claims of quantum advantage against near-term feasibility.