Quantum Advantage in ML, What's Real

Quantum Advantage in ML: What's Real

The Hype vs. Reality

Quantum machine learning has generated enormous excitement and significant hype. Let's be precise about where things stand in 2025.

Where Quantum ML Has NOT Proven Advantage

Image classification: Classical CNNs (ResNet, ViT) achieve near-human accuracy on ImageNet. No quantum circuit comes close on any real dataset.

Natural language processing: Large language models process billions of tokens. Quantum circuits cannot match this scale or capability.

Recommendation systems: Classical matrix factorization and neural collaborative filtering work extremely well. The dequantization results of Tang (2019) show that the proposed quantum speedups for recommendation were illusory: classical algorithms can achieve the same complexity.

General function approximation: The universal approximation theorem applies to classical networks. Quantum circuits are also universal approximators, but classical hardware beats quantum on any practically scaled task today.

Where Quantum ML Shows Genuine Promise

Quantum chemistry simulation: Computing molecular ground state energies. VQE on near-term hardware has demonstrated small molecules (H2, LiH, BeH2), all of which are still easy for classical methods; whether VQE can ever beat the best classical chemistry techniques remains an open research question.

Quantum optimization: QAOA for combinatorial problems. No proven advantage yet, but theoretical arguments exist for specific structured problems.

Learning from quantum data: If the input is already quantum (from a quantum sensor or quantum communication channel), quantum ML naturally avoids the encoding bottleneck.

Kernel-based advantage: For specific synthetic datasets designed to match quantum feature map geometry, quantum kernels have provably outperformed classical kernels. But these datasets are constructed to be favorable, not naturally occurring.

The Dequantization Problem

The most important theoretical development for quantum ML skeptics: dequantization. Ewin Tang's 2019 result showed that many proposed quantum ML speedups could be achieved by classical algorithms using sampling methods.

The quantum recommendation system algorithm was O(polylog(N)): exponentially faster than classical. But sampling-based classical algorithms achieve similar complexity. The lesson: be very careful about comparing quantum algorithms to naive classical algorithms. Compare to the best classical algorithms.