Reproduce “When Quantum Meets AI: Quantum Methods for Machine Learning and Machine Learning Methods for Quantum Systems” on OpenResearch

About this paper

Tak Hur

Tak Hur’s dissertation develops methods at the intersection of quantum computing and AI, spanning quantum-enhanced representations, QNN generalization, neural surface-code decoding, and neural quantum-state optimization. Its experiments show that learned quantum embeddings can raise MNIST accuracy from 52.7% to 96.1% on IBM hardware, Mamba decoding preserves Transformer-level error rates with roughly threefold lower latency at code distance 39, and multi-shift regularization reduces validation error and update variance in neural quantum states.

Key insights

• The trace distance between class-averaged embedded quantum states imposes a lower bound on achievable binary classification risk, and quantum channels or hardware noise cannot improve this distinguishability.
• Margins were more predictive of observed generalization gaps than total or effective parameter counts in the studied QCNN phase-recognition tasks, linking representation quality to downstream generalization.
• Regularization in SR is not merely numerical stabilization: it filters poorly estimated tangent-space directions, reducing variance at the cost of shrinkage bias.

Summary by alphaXiv.