cv-qnn-edge-ai

name

cv-qnn-edge-ai

description

Parameter-efficient Continuous-Variable photonic Quantum Neural Networks for edge AI deployment. Simplified CV-QNN architecture reduces trainable parameters by 40-45% while maintaining or exceeding classical baseline performance. Barren plateau mitigation via dimensionality reduction and encoding restriction. Use when: building quantum machine learning models for edge deployment, optimizing CV-QNN architectures, mitigating barren plateaus, parameter-efficient quantum classification.

Core Methodology

Simplified CV-QNN Architecture (Φ∘D∘U₁)

Standard CV-QNN layers (Killoran et al. 2019) use displacement (D), squeezing (S), and interferometric (U) gates. The simplified architecture Φ∘D∘U₁ removes squeezing and reduces to:

1. U₁ - Single interferometer (passive linear optics)
2. D - Displacement gates (amplitude/phase modulation)
3. Φ - Nonlinear Kerr gates (measurement)

This cuts trainable parameters by 40-45% relative to the standard layer.

Barren Plateau Mitigation Strategies

Dimensionality Reduction: Use PCA to reduce input dimensions before quantum encoding. Reducing to 16 dimensions was effective for image classification tasks.

Encoding Restriction: Restrict which qumodes receive encoded data. Don't encode into all qumodes simultaneously - selective encoding prevents gradient vanishing.

Key Result: These strategies raise loss-gradient variance by ~58 orders of magnitude, effectively eliminating barren plateaus.

Width-Dependent Performance

• 2 qumodes: Full layer has small but significant edge
• 4 qumodes: Simplified layer is significantly better with 44% fewer parameters

Parameter Efficiency Benchmarks

• 4-qumode simplified CV-QNN: only 18 trainable parameters
• Exceeds 55-parameter classical baseline with 67% fewer parameters
• Achieves 100% calibrated test accuracy across all seeds

Pipeline Architecture

Raw Input → Classical Feature Extractor (e.g., MobileNetV1) 
          → PCA Dimensionality Reduction 
          → CV-QNN Encoding (restricted) 
          → Simplified CV-QNN Layer (Φ∘D∘U₁) 
          → Measurement → Classification

Key Parameters

Parameter	Recommended Value
Input dimensions (after PCA)	16
Qumodes	4 (optimal for simplified)
Trainable parameters	~18
Encoding restriction	Partial (not all qumodes)

Activation

cv-qnn, continuous-variable quantum, photonic quantum computing, edge quantum AI, barren plateau mitigation, parameter-efficient quantum ML, quantum neural network optimization, room-temperature quantum computing