qnn-survey-design-patterns - SKILL.md Agent Skill

name: qnn-survey-design-patterns description: "Quantum Neural Network (QNN) design patterns and architecture selection guide — comprehensive survey methodology for selecting, designing, and evaluating QNN architectures based on task requirements. Covers fully connected QNNs, quantum CNNs, equivariant QNNs, quantum Hopfield networks, quantum Boltzmann machines, quantum reservoir computing, and composite networks. Activation: QNN survey, quantum neural network design, QNN architecture selection, quantum machine learning survey, 量子神经网络综述"

Overview

Comprehensive QNN design patterns extracted from the survey paper "Research progress on quantum neural networks and quantum machine learning" (arXiv:2605.30724, May 2026). This skill provides a practical decision framework for selecting QNN architectures based on task requirements, hardware constraints, and performance goals.

arXiv Reference

ID: 2605.30724
Title: Research progress on quantum neural networks and quantum machine learning
Authors: Yifan Sun, Boyuan Sun, Jiameng Tian, Xiangdong Zhang
Published: May 2026
Categories: quant-ph

QNN Architecture Selection Matrix

By Task Type

Task Type	Recommended QNN	Why
Image/Pattern Recognition	Quantum CNN (QCNN)	Hierarchical feature extraction via quantum pooling
Sequence/Time-Series	Fully Connected QNN (FC-QNN)	Direct mapping of temporal features to qubits
Associative Memory	Quantum Hopfield Network	Quantum superposition enables exponential memory capacity
Optimization/Combinatorial	Quantum Boltzmann Machine	Energy-based formulation for combinatorial search
Chaotic/Dynamic Systems	Quantum Reservoir Computing	Fixed random quantum circuit, only readout trained
Symmetry-Preserving Tasks	Equivariant QNN	Built-in symmetry guarantees via group representations
Reinforcement Learning	Composite QNN-RL	QNN as policy/value function approximator
Generative Modeling	Composite QNN-GL	Quantum circuits for probability amplitude encoding
Transfer Learning	Composite QNN-TL	Pre-trained quantum feature extractors

By Resource Constraints

Constraint	Recommended Approach	Notes
Few qubits (NISQ, <20)	FC-QNN, QRC	Minimal depth circuits
Moderate qubits (20-50)	QCNN, Equivariant QNN	Structured ansatze
Many qubits (>50)	Full QBM, Deep QNN	Requires error mitigation
Limited circuit depth	QRC, FC-QNN shallow	Reservoir has fixed random circuit
Can train classically	Hybrid QNN-Classical	Classical optimizer + quantum circuit

Key Design Patterns

Pattern 1: Parameterized Quantum Circuit (PQC) Encoding

Classical Data → Encoding Circuit (Rx, Ry, Rz) → Variational Layer → Measurement → Classical Output

Encoding: Amplitude, angle, or basis encoding of classical data
Variational: Alternating layers of parameterized single-qubit gates + entangling gates
Measurement: Expectation values of observables as output

Pattern 2: Quantum Convolution + Pooling

Input State → [Quantum Convolution (entangling gates) → Quantum Pooling (measurement + post-selection)]^n → Classification

Hierarchical feature extraction analogous to classical CNN
Pooling via mid-circuit measurement and qubit reuse

Pattern 3: Quantum Reservoir Computing

Input → Fixed Random Quantum Circuit (reservoir) → Measurement → Classical Readout (trained)

Only the classical readout layer is trained
Quantum circuit is fixed and random
Exploits quantum dynamics as rich feature space

Pattern 4: Hybrid Classical-Quantum

Classical NN Feature Extractor → Quantum Circuit (nonlinear transformation) → Classical Classifier

Classical layers extract features
Quantum circuit provides quantum advantage in feature space
Classical classifier on measured outputs

Barren Plateaus Mitigation

The survey identifies barren plateaus as a critical training challenge. Key solutions:

Structured Ansatz Design: Use problem-inspired ansatze rather than hardware-efficient random circuits
Layer-by-Layer Training: Train QNN layers sequentially rather than all-at-once
Local Cost Functions: Use local observables instead of global ones
Parameter Initialization: Initialize near identity or use classical pre-training
Adaptive Learning Rates: Use quantum natural gradient or adaptive methods

Performance Comparison Summary

QNN Type	Training Speed	Expressivity	NISQ-Friendly	Scalability
FC-QNN	Fast	High	Yes	Moderate
QCNN	Moderate	High	Yes	Good
Equivariant QNN	Slow	Very High	Moderate	Good
QHN	Moderate	Very High	Limited	Poor
QBM	Slow	Very High	Limited	Poor
QRC	Very Fast	Moderate	Yes	Excellent

Practical Implementation Guidelines

Start Simple: Begin with FC-QNN or QRC for proof of concept
Match Symmetry: Use equivariant QNN when the problem has known symmetries
Consider Data Loading: The state preparation bottleneck is critical — use neural network encoding (see nn-quantum-state-encoding skill) to avoid variational optimization per data point
Hybrid Approach: For near-term, hybrid classical-quantum architectures are most practical
Monitor Expressivity: Use effective rank or entanglement entropy to measure QNN expressivity during training
Avoid Barren Plateaus: Use structured ansatze and local cost functions

Related Skills

nn-quantum-state-encoding — Neural network encoding for quantum state preparation (arXiv:2605.31006)
quantum-neural-barren-plateau — Barren plateau mitigation techniques
quantum-neural-architecture — General QNN architecture design
quantum-reservoir-computing — Quantum reservoir computing methodology
qml-framework-agnostic-design — Framework-agnostic QML design

Activation

Use this skill when:

Designing quantum neural network architectures
Comparing QNN types for a specific task
Evaluating quantum machine learning approaches
Selecting between classical and quantum neural networks
Surveying QNN literature or writing QNN-related code