Quantum Machine Learning Overview

Why This Matters

Quantum machine learning promises to use quantum hardware to speed up or improve classical learning tasks. The honest picture as of 2026 is that this is still mostly an exploratory research field. There is no proven quantum advantage on classical-data ML problems: no benchmark where a quantum algorithm clearly beats the best classical method on data that comes from the classical world. Most demonstrated speedups assume idealized fault-tolerant hardware that does not yet exist, or use input models (like coherent quantum-state access) that hide the cost of loading classical data.

This page is an overview, not a treatment from a hardware expert. The goal is to give an ML audience the right mental model for what quantum ML actually is, what the main approaches are, and what is and is not currently practical. The four main approaches are quantum kernels, variational quantum circuits, amplitude-encoded data pipelines, and quantum optimization heuristics like QAOA. All four are studied actively. None has produced a clear classical-data advantage at the scale of contemporary deep learning.

Core Ideas

Quantum kernels. A quantum feature map sends a classical input $x$ to a quantum state $|\phi(x)\rangle$ via a circuit. The quantum kernel is the inner product $k(x, y) = |\langle \phi(x) | \phi(y) \rangle|^2$, estimated by running the circuit and measuring overlap. The interest comes from the Hilbert space of quantum states being exponentially large in the number of qubits, so the feature space is high dimensional almost for free. Havlicek et al. (2019) and Schuld and Killoran (2019) framed this view simultaneously. The catch: not every quantum kernel gives an advantage, and Schuld et al. (2021) showed that the expressive power depends sharply on the data-encoding choice, not just on circuit depth.

Variational quantum circuits. A parametrized circuit $U(\theta)$ produces a quantum state $|\psi(\theta)\rangle$. A loss function is defined as the expectation of some observable on this state, and the parameters $\theta$ are optimized classically using gradient estimates from the quantum device. Cerezo et al. (2021) is the standard review of variational quantum algorithms. Two persistent problems: barren plateaus, where gradients vanish exponentially in the number of qubits for random initializations, and noise on near-term devices that corrupts gradient estimates beyond what classical optimizers can handle.

Quantum data and amplitude encoding. Loading a classical vector $x \in \mathbb{R}^N$ into a quantum state requires preparing $|x\rangle = \sum_i x_i / \|x\| \, |i\rangle$, which uses $\log_2 N$ qubits but in general $\Omega(N)$ gates. This loading cost cancels most theoretical speedups for tasks defined on classical input. The few cases where quantum ML has clear theoretical wins (Harrow-Hassidim-Lloyd for sparse linear systems, certain principal component flows) all assume the data is already given as a quantum state, or arrives from a quantum sensor.

Quantum optimization (QAOA). The Quantum Approximate Optimization Algorithm of Farhi, Goldstone, and Gutmann (2014) targets combinatorial problems like Max-Cut by alternating a problem Hamiltonian and a mixing Hamiltonian for $p$ rounds. Empirically, classical heuristics like Goemans-Williamson and modern SAT solvers remain hard to beat on the problem sizes current hardware can run. Theoretical separations exist for specific structured instances; practical advantage on real workloads has not been demonstrated.

Hardware caveat: NISQ era. Preskill (2018) coined "Noisy Intermediate-Scale Quantum" to describe what is currently buildable: 50 to a few hundred qubits with gate error rates around $10^{-3}$ and no full quantum error correction. NISQ devices cannot run the textbook quantum algorithms (Shor, HHL, Grover at meaningful scale) because those require error rates orders of magnitude lower. Variational and hybrid algorithms exist precisely to do something useful within these limits, with the honest understanding that "useful" has not yet meant "beats classical at a real ML task."

Watch Out

Quantum speedup is not the same as exponential parameter count

A quantum state on $n$ qubits lives in a $2^n$-dimensional space, but you cannot read out $2^n$ amplitudes. Measurement collapses the state to one classical outcome. To estimate any property to precision $\varepsilon$ requires $O(1 / \varepsilon^2)$ repetitions. The "exponential capacity" framing is misleading: usable information per shot is at most $n$ bits.

Watch Out

Quantum kernels are not automatically classically hard

A quantum kernel only gives a real advantage if the inner products $|\langle \phi(x) | \phi(y) \rangle|^2$ are hard to compute classically. Many proposed quantum feature maps turn out to be classically simulable, and the resulting kernels have efficient classical descriptions. Whether a particular quantum kernel is actually beyond classical reach is itself a research question, not a default property.

References

Related Topics

• Kernels and RKHS
• Convex Optimization Basics
• Wave-Based Neural Networks
• Adversarial Machine Learning