Quantum Contrastive Learning

A hybrid classical-quantum framework for contrastive representation learning on MNIST, using variational quantum circuits as projection heads.

Overview

This project implements a Quantum Contrastive Learning framework inspired by SimCLR, combining:

• Classical encoders (ResNet-18, lightweight CNN) for feature extraction
• Quantum projection heads using variational circuits (Tensor Ring, MERA)
• NT-Xent contrastive loss for self-supervised representation learning
• Linear evaluation on frozen encoder features for classification

The framework follows a two-phase pipeline:

1. Contrastive pre-training — encoder + quantum head are jointly optimized on augmented image pairs
2. Linear evaluation — encoder is frozen, quantum circuit removed, linear classifier trained on frozen features

Architecture

Classical Encoders

Encoder	Output Dim	Parameters	Description
ResNet-18	512	11.4M	Standard ResNet-18 adapted for 28x28 grayscale input
CNN	2^n_qubits	~518K (4q) / ~1M (6q)	Lightweight 1D CNN with 4 conv blocks

Quantum Projection Head

Two encoding strategies map classical features to quantum states:

• Amplitude encoding — direct state preparation
• Angle encoding — re-uploading with RY gates

Two circuit ansatze are supported:

Ansatz	Qubits	Description
Tensor Ring	4 (default)	Full-entanglement ring with RY-CNOT-RY blocks, configurable depth
MERA	6 (fixed)	Multi-scale Entanglement Renormalization Ansatz with 9 unitary blocks per layer

Measurement is performed via Pauli-Z expectation values.

Classical Features -> Quantum Encoding -> [Variational Layers] x D -> Pauli-Z Measurement -> Quantum Features

Installation

This project uses UV for Python package management and requires Python 3.12.

git clone https://github.com/mattiacurri/QuantumComputing-Project.git
cd QuantumComputing-Project
uv sync

Quick Start

# Train with ResNet-18 encoder + amplitude encoding + Tensor Ring circuit
uv run src/quantum/contrastive_supervised.py \
    --encoder simclr \
    --quantum-encoding amplitude \
    --epochs 200 \
    --linear-epochs 200 \
    --batch-size 64 \
    --n-samples-per-class 500

# Train with CNN encoder + angle encoding + MERA circuit (6 qubits)
uv run src/quantum/contrastive_supervised.py \
    --encoder cnn \
    --quantum-encoding angle \
    --circuit-type mera \
    --num-qubits 6 \
    --epochs 200 \
    --linear-epochs 200 \
    --batch-size 64 \
    --n-samples-per-class 500

# Run all experiments from the report
bash run_experiments.sh          # Linux/macOS
.\run_experiments.bat            # Windows PowerShell

Running Experiments

Quick Start: Smoke Tests

Linux/macOS:

bash smoke_test_experiments.sh

Windows (PowerShell):

.\smoke_test_experiments.bat

✓ Smoke tests run all 20 configurations with minimal parameters:

• Epochs: 2 (vs 200 full)
• Samples/class: 50 (vs 500 full)
• Batch size: 32 (vs 64 full)
• No WandB logging (local output only)

Full Experiment Suite

Linux/macOS:

bash run_experiments.sh

Windows (PowerShell):

.\run_experiments.bat

Single Experiment

For standalone runs, use the training script directly:

uv run src/quantum/contrastive_supervised.py --encoder simclr --quantum-encoding amplitude --epochs 200 --linear-epochs 200 --batch-size 64 --n-samples-per-class 500

Project Structure

src/
  classical/          Classical encoder architectures (ResNet-18, CNN)
  quantum/            Quantum SimCLR model and training script
  data/               MNIST data loading and augmentation transforms
  expressivity/       Circuit expressivity analysis (Haar-random sampling)
  utils/              Plotting, WandB logging, configuration

reports/              LaTeX report and circuit figures

run_experiments.bat   Reproduces all 20 experiments from the report
run_experiments.sh    Reproduces all 20 experiments from the report (Linux/macOS)
smoke_test_experiments.bat  Runs quick smoke tests for all configurations (Windows)
smoke_test_experiments.sh    Runs quick smoke tests for all configurations (Linux/macOS)
compute_circuit_expressivity.py  Script for computing circuit expressivity

Results

See the report for detailed results and analysis.

---

🚀