| name | cv-qnn-spatial-classification |
| description | Controlled comparison methodology showing continuous-variable (CV) QNNs outperform discrete-variable (DV) QNNs on spatial pattern recognition tasks like wafer-map defect classification. |
| category | quantum-ml |
| trigger_words | ["CV versus DV","continuous variable QNN","discrete variable QNN","wafer map classification","quantum spatial encoding","CV advantage","Fock cutoff"] |
CV versus DV Quantum Neural Networks: Spatial Classification
Paper: arXiv:2607.00961v1
Authors: Yeonhong Kim, Jonghyeok Im, Monu Nath Baitha, Kyoungsik Kim
Core Insight
Under controlled conditions, CV-QNNs consistently outperform DV-QNNs on spatial pattern recognition tasks. At 4 qumodes/qubits, CV reaches 79.7% accuracy vs DV's 61.6% — an 18-point gap.
Key Results
- CV Advantage: Sharpest on spatially localized classes (Edge-Loc recall 0.66 vs DV ≤0.05)
- Representational Ceiling: DV limitation is capacity-bound, not optimization failure
- Structured Encoding: CV's continuous phase-space encoding captures fine spatial distinctions
- Hardware Validation: DV accuracy holds at shallow depth, degrades at deepest circuits
Methodology
Controlled Architecture
- Shared convolutional backbone (~4.3M params)
- Interchangeable heads: classical dense, CV-QNN, DV-QNN
- Scaled over 3 sizes (3, 4, 8 qumodes/qubits)
Encoding Strategy
- CV: structured neural-network-analogue layer + continuous phase-space
- DV: Hilbert space dimensionality with Fock cutoff d=2
Applications
- Industrial QA: Wafer-level defect screening for semiconductor yield
- Pattern Recognition: Tasks requiring fine spatial distinction
- Quantum Advantage Search: Identifying where structured quantum heads help
