Quantum Convolutional Neural Networks vs Fully Classical CNNs

This repository contains a comprehensive performance comparison between Hybrid Quantum Convolutional Neural Networks (QCNN) and Fully Classical CNNs on the MNIST handwritten digit dataset.

Overview

This project implements and evaluates:

• A hybrid quantum-classical neural network architecture using PyTorch and PennyLane
• A comparable fully classical CNN architecture
• Performance metrics and visualizations for both approaches

Models

Hybrid QCNN

The hybrid model combines classical convolutional layers with a quantum circuit layer:

• Classical convolutional feature extraction
• Dimension reduction to quantum circuit input size
• Quantum circuit with parameterized rotations and entangling operations
• Classical post-processing and classification

Fully Classical CNN

For fair comparison, a classical CNN with comparable architecture was implemented:

• Same convolutional layers as the hybrid model
• Classical MLP layers in place of the quantum circuit
• Matching input/output dimensions throughout

Results

Performance Summary

• Hybrid QCNN Accuracy: 90.54%
• Fully Classical CNN Accuracy: 81.97%
• Performance Difference: 8.57% (Hybrid outperforms Classical)

Training History

Model Evaluation

Comparison

Key Findings

The repository explores:

• Training convergence of both models
• Test accuracy comparison
• Class-wise performance differences
• Resource usage considerations

Technologies Used

• PyTorch - Deep learning framework
• PennyLane - Quantum machine learning library
• TensorFlow/Keras - Data loading and preprocessing
• Matplotlib/Seaborn - Visualization
• Rich - Terminal output formatting

Getting Started

1. Clone this repository
2. Install dependencies:

   pip install torch pennylane tensorflow matplotlib seaborn rich
   

3. Run the Jupyter notebook to train and compare the models

Notebook Contents

• Data loading and preprocessing
• Hybrid QCNN implementation
• Classical CNN implementation
• Training procedures
• Evaluation metrics
• Performance visualization
• Comparative analysis

The notebook demonstrates how quantum circuit layers can be integrated into traditional neural network architectures and analyzes the performance trade-offs compared to fully classical approaches.