Research map / Probabilistic and scientific modeling / Molecules and proteins
Neural Auto-designer for Enhanced Quantum Kernels
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes QuKerNet, a neural-predictor-driven method that jointly searches quantum feature-map gate layouts (using mRMR feature selection and KTA surrogate labels) and then fine-tunes variational parameters to design problem-specific quantum kernels.
Problem
Designing effective quantum feature maps for real-world high-dimensional datasets on NISQ hardware with limited qubits, without prior task knowledge, while avoiding vanishing similarity (kernel concentration).
Approach
Frames quantum kernel design as discrete-continuous joint optimization: an MLP neural predictor trained on cheap kernel-target alignment labels ranks circuit layouts, the top-k layouts are fine-tuned on variational parameters, and mRMR feature selection keeps circuits small (NL ~ O(p)).
Key results
- On the synthetic dataset QuKerNet-2 reaches 91.33% test accuracy versus 71.67% for RBFK and 73% for TEK.
- On tailored MNIST QuKerNet-2 reaches 86.8% test accuracy versus 83.6% for RBFK; on tailored CC it reaches 94.4% versus 92.8% for RBFK.
- KTA correlates strongly with training accuracy of quantum kernels (PCC 0.9881 in QuKerNet-1 and 0.9821 in QuKerNet-2).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add feature selection (mRMR) instead of random feature selection (ablation) | QuKerNet · Tailored MNIST | test accuracy | Helped |
| fine-tune variational parameters after circuit-layout search (ablation) | QuKerNet-1 · Tailored MNIST | test accuracy: 85.87 $\pm$ 1.43 -> 86.80 $\pm$ 1.36 | Helped |
| fine-tune variational parameters after circuit-layout search (ablation) | QuKerNet-1 · Tailored CC | test accuracy: 92.20 $\pm$ 1.17 -> 94.40 $\pm$ 0.80 | Helped |
| fine-tune variational parameters after circuit-layout search (ablation) | QuKerNet-1 · Synthetic dataset | test accuracy: 84.00 $\pm$ 3.89 -> 91.33 $\pm$ 1.63 | Helped |
| search circuit layouts with QuKerNet rather than Bayesian optimization | quantum kernel design · Tailored MNIST | test accuracy: 85.60 $\pm$ 1.44 vs 86.80 $\pm$ 1.36 | Helped |
| search circuit layouts with QuKerNet rather than Bayesian optimization | quantum kernel design · Tailored CC | test accuracy: 93.00 $\pm$ 1.41 vs 94.40 $\pm$ 0.80 | Helped |
| search circuit layouts with QuKerNet rather than Bayesian optimization | quantum kernel design · Synthetic dataset | test accuracy: 89.34 $\pm$ 2.26 vs 91.33 $\pm$ 1.63 | Helped |
| use QuKerNet instead of RBFK | quantum kernel classification · Synthetic dataset | test accuracy: 71.67 $\pm$ 2.79 vs 91.33 $\pm$ 1.63 | Helped |
| use QuKerNet instead of NKN | kernel classification · Tailored CC | test accuracy: 47.67% vs 94.4% | Helped |
Surprising findings
- The neural-based Neural Kernel Network baseline with default parameters performed poorly (e.g., 47.67% on tailored CC) and overfit, with near 100% training accuracy.
- Bayesian optimization search did not surpass random search within the search space on tailored MNIST (85.6% vs 86.8%).
Limitations
- Experiments use small datasets (e.g., 300 MNIST images, 200 CC samples) and few qubits, with noisy-device runs on very small settings.
- Searching uses a restricted hardware-efficient ansatz search space and the neural predictor is a simple 2-layer MLP.
- Code will be released only as a promised GitHub repository; no released link is given.
Experimental setup
Datasets: MNIST Credit Card fraud dataset synthetic dataset relabeled from tailored MNIST
Benchmarks: Tailored MNIST Tailored CC Synthetic dataset
Metrics: test accuracy kernel variance Pearson correlation coefficient
Baselines: RBFK HEAK TEK Neural Kernel Network Bayesian optimization random search
Builds on: hardware efficient ansatz kernel-target alignment mRMR neural architecture search
Closest papers by meaning
- Rethink the Role of Deep Learning towards Large-scale Quantum Systems (ICML 2025 · similarity 0.88)
- Optimal algorithmic complexity of inference in quantum kernel methods (NeurIPS 2026 · similarity 0.86)
- TensorRL-QAS: Reinforcement learning with tensor networks for improved quantum architecture search (NeurIPS 2025 · similarity 0.86)
- One More Time: Revisiting Neural Quantum States from a Reinforcement Learning Perspective (NeurIPS 2026 · similarity 0.85)
- QCircuitBench: A Large-Scale Dataset for Benchmarking Quantum Algorithm Design (NeurIPS 2025 · similarity 0.85)
- Winning Lottery Tickets in Neural Networks via a Quantum-Inspired Classical Algorithm (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.