| name | cv-qnn-spatial-classification |
| description | Continuous-variable QNN advantage for spatial classification tasks. Controlled comparison showing CV-QNN outperforms DV-QNN by 18+ percentage points on wafer-map defect classification. CV structured layer captures fine spatial distinctions that DV misses. Use when designing QNNs for image/spatial classification, semiconductor yield, or any task requiring fine spatial pattern recognition. |
| version | 1.0.0 |
| tags | ["quantum","CV-QNN","DV-QNN","spatial-classification","image","defect-detection"] |
| source | arXiv:2607.00961 |
| authors | ["Yeonhong Kim","Jonghyeok Im","Monu Nath Baitha","Kyoungsik Kim"] |
| published | 2026-07-01T00:00:00.000Z |
| trigger_words | ["CV QNN vs DV QNN","continuous variable quantum neural network","quantum spatial classification","wafer map defect","quantum image classification","QNN representational capacity"] |
CV-QNN Advantage for Spatial Classification
Core Insight
CV-QNNs consistently outperform DV-QNNs on spatial classification tasks by 18+ percentage points. The CV advantage comes from two intrinsic properties: structured neural-network-analogue layer and continuous phase-space encoding, NOT Hilbert-space dimensionality.
Key Findings
1. Performance Gap
- At 4 qumodes/qubits: CV = 79.7% vs DV = 61.6% (non-overlapping 18-point gap)
- Gap is sharpest on spatially localized classes (Edge-Loc: CV recall 0.66, DV recall < 0.05)
- DV limitation is representational-capacity ceiling, not optimization failure
2. Why CV Wins
- Structured CV layer better captures fine spatial distinctions
- Continuous phase-space encoding preserves spatial information
- CV acts as neural-network-analogue layer (unlike DV)
3. Hardware Validation
- DV accuracy holds at shallow depth on IBM hardware
- DV degrades only at deepest circuit
- CV advantage expected to grow as noise improves
Implementation Pattern
- Use shared convolutional backbone for feature extraction
- Replace classical dense head with CV-QNN head
- Scale CV head over multiple qumode counts (3, 4, 8)
- Compare against DV-QNN head with same qubit count
- Validate on hardware at shallow depth
Practical Applications
Financial Spatial Data
- Geographic risk mapping
- Regional market pattern classification
- Spatial-temporal financial data analysis
General Spatial Classification
- Image defect detection
- Medical image classification
- Any task requiring fine spatial pattern distinction
Activation
Use when:
- Designing QNNs for spatial/image classification
- Choosing between CV and DV quantum paradigms
- Needing fine spatial pattern recognition
- Building hybrid classical-quantum classifiers
