Quantum Machine Learning: VQC, Quantum Kernels, and Where the Hype Ends

In 2018, QML papers appeared with headlines claiming '100x quantum speedup for machine learning'. By 2024, many of the same groups published corrections. This is how science works - the important skill is reading both sides.

• Google Quantum AI uses VQC for quantum chemistry - calculating molecular energies matters for drug discovery and catalyst design, where quantum advantage is well-motivated
• BMW and Volkswagen research quantum optimization for logistics planning - practical applicability at scale remains an open question
• PennyLane (Xanadu) is the dominant open-source QML framework with 8k+ GitHub stars, widely used in quantum kernel and VQC research

Variational Quantum Circuits: Neural Networks on Quantum Hardware

2017. Researchers observe that NISQ machines are too noisy for Shor and Grover - but what if they were used as parametric circuits, analogous to neural networks? This insight spawned the VQC (Variational Quantum Circuit) research direction.

A VQC has three components: classical encoding of input data into qubits, a parametrized quantum ansatz (the trainable circuit), and a classical optimizer updating the parameters via gradient descent. Measuring the qubits at the end produces a prediction.

VQC is a hybrid classical-quantum model: the quantum circuit does the forward pass, a classical computer updates parameters. This is a necessity - quantum computers cannot yet run backpropagation efficiently - but it creates a fundamental question: if a classical computer handles optimization, where exactly is the quantum advantage?

Barren Plateaus: Why VQC Training Is Hard

The barren plateau problem is one of the central open questions in QML. As the number of qubits or circuit depth grows, gradients of the loss function decay exponentially. For a random ansatz on n qubits, the gradient variance scales as exp(-n). Training a 50-qubit VQC is practically impossible - all gradients are numerically zero.

Root cause: a sufficiently expressive random ansatz forms a unitary 2-design - the circuit explores the full unitary space nearly uniformly. Any single parameter's gradient averages to zero. Mitigation strategies: shallow circuits (1-3 layers), locally structured ansatz, problem-specific designs. But any ansatz expressive enough to potentially show quantum advantage tends to hit the plateau.

Barren plateau is not just a technical obstacle - it is a fundamental mathematical limitation. Some researchers argue it is unresolvable for general VQC. Only task-specific ansatz with strong structural prior constraints reliably avoid it.

Quantum Kernels: A Different QML Paradigm

Quantum kernel methods offer an alternative to VQC. Classical SVMs operate via a kernel k(x, x') = inner product of phi(x) and phi(x') in feature space. A quantum kernel: the feature map phi is a quantum circuit, and the inner product computation is a quantum measurement.

Quantum kernels avoid the barren plateau (no quantum parameter optimization) and have rigorous theoretical guarantees via PAC learning theory. The catch: computing K_{ij} requires a separate quantum circuit execution for every data pair - that is N squared quantum circuit runs for N training points, making large-scale training expensive.

Honest Assessment: Where QML Helps and Where It Is Hype

2023-2024: a wave of papers cools QML enthusiasm. Huang et al. (Nature, 2021): classical ML algorithms can simulate most quantum kernel methods in polynomial time. Abbas et al. (2021): VQC expressibility does not exceed classical neural networks for most tasks. This is good science - the field is self-correcting.

An honest forecast: QML has genuine prospects for quantum advantage in specific domains - quantum chemistry (drug discovery, materials science), quantum system control, combinatorial optimization problems with quantum structure. For classical ML tasks (NLP, computer vision) - classical methods will likely remain superior. This is not the end of QML - it is its realistic scope.

QML in the Quantum Technology Landscape

QML is an active research area transitioning from broad claims toward targeted applications with defensible quantum advantage.

• Computational Chemistry — Related topic
• Quantum Optimization — Related topic
• Classical ML Theory — Related topic

Итоги

• VQC is a hybrid parametric model: quantum forward pass + classical optimizer; barren plateau makes training deep VQC exponentially hard
• Quantum kernels avoid barren plateau (no quantum parameter optimization) but require N squared circuit runs for N training points
• Honest assessment: quantum advantage is real for quantum system simulation; for classical ML tasks (NLP, CV) - classical methods will likely dominate
• QML is not a GPU replacement - it targets problems where quantum mechanics is intrinsic to the task, not a tool applied from outside

Вопросы для размышления

• If forced to pick one applied task where QML will most likely deliver practical quantum advantage by 2030 - what would it be and what is the reasoning?

Связанные уроки

• qc-19-error-correction — QML on NISQ devices operates without error correction - understanding NISQ limitations is essential for honest QML assessment
• qc-17 — Quantum algorithms establish where quantum speedup is theoretically possible
• ml-15-naive-bayes — Classical kernel methods (SVM) are the foundation for quantum kernels - which are kernel SVM with a quantum feature map
• aie-09-embeddings — Quantum feature maps play the same role as embeddings: mapping data to a space where the task becomes tractable