QuantumVE: Quantum-Transformer Advantage Boost Over Classical ML

Breaking Discovery: Vision Transformer embeddings unlock quantum machine learning advantage! First systematic proof that embedding choice determines quantum kernel success, revealing fundamental synergy between transformer attention and quantum feature spaces.

• 📂 GitHub Repository: QuantumVE
• 📄 Research Paper: Embedding-Aware Quantum-Classical SVMs for Scalable Quantum Machine Learning
• 💻 Dataset on HuggingFace: QuantumEmbeddings
• Demo: Colab

🎯 Breakthrough Results

• 8.02% accuracy improvement on Fashion-MNIST vs classical SVMs
• 4.42% boost on MNIST dataset
  
• First evidence that ViT embeddings enable quantum advantage while CNN features show degradation
• 16-qubit tensor network simulation via cuTensorNet proving scalability
• Class-balanced k-means distillation for efficient quantum processing

Project Architecture

QuantumVE/ ├── data_processing/ # Class-balanced k-means distillation procedures ├── embeddings/ # Vision Transformer & CNN embedding extraction ├── qve/ # Core quantum-classical modules and utilities └── scripts/ # Experimental pipelines with cross-validation ├── classical_baseline.py # Traditional SVM benchmarks ├── cross_validation_baseline.py # Cross-validation framework └── qsvm_cuda_embeddings.py # Our embedding-aware quantum method

🚀 Quick Start

1. Environment Setup

```bash

Create conda environment

conda create -n QuantumVE python=3.11 -y conda activate QuantumVE

Clone and install

git clone https://github.com/sebasmos/QuantumVE.git cd QuantumVE pip install -e .

For Ryzen devices - Install MPI

conda install -c conda-forge mpi4py openmpi ```

2. Download Pre-computed Embeddings

MNIST Embeddings: bash mkdir -p data && \ wget https://huggingface.co/datasets/sebasmos/QuantumEmbeddings/resolve/main/mnist_embeddings.zip && \ unzip mnist_embeddings.zip -d data && \ rm mnist_embeddings.zip

Fashion-MNIST Embeddings: bash mkdir -p data && \ wget https://huggingface.co/datasets/sebasmos/QuantumEmbeddings/resolve/main/fashionmnist_embeddings.zip && \ unzip fashionmnist_embeddings.zip -d data && \ rm fashionmnist_embeddings.zip

3. Run Experiments

Single Node: ```bash

Classical baseline with cross-validation

python scripts/classical_baseline.py

Cross-validation framework

python scripts/crossvalidationbaseline.py

Our embedding-aware quantum method

python scripts/qsvmcudaembeddings.py ```

Multi-Node with MPI: ```bash

Run with 2 processes

mpirun -np 2 python scripts/qsvmcudaembeddings.py mpirun -np 2 python scripts/crossvalidationbaseline.py ```

🔬 What Makes This Work?

Our key insight: embedding choice is critical for quantum advantage. While CNN features degrade in quantum systems, Vision Transformer embeddings create a unique synergy with quantum feature spaces, enabling measurable performance gains through:

1. Class-balanced distillation reduces quantum overhead while preserving critical patterns
2. ViT attention mechanisms align naturally with quantum superposition states
3. Tensor network simulation scales to practical problem sizes (16+ qubits)

🤝 Contributing

We welcome contributions! Help us advance quantum machine learning:

1. Fork the QuantumVE repository
2. Create your feature branch (git checkout -b feature/amazing-feature)
3. Submit a pull request with detailed description

Areas for contribution: - New embedding architectures (BERT, CLIP, etc.) - Additional quantum backends - Performance optimizations - Documentation improvements

🙏 Acknowledgements

This work was supported by the Google Cloud Research Credits program under award number GCP19980904.

📄 License

CC BY-NC-SA 4.0

📚 Citation

Paper

bibtex @article{Cajas2024_QuantumVE, title={Embedding-Aware Quantum-Classical SVMs for Scalable Quantum Machine Learning}, author={Cajas Ordóñez, Sebastián Andrés and Torres Torres, Luis and Bifulco, Mario and Duran, Carlos and Bosch, Cristian and Simón Carbajo, Ricardo}, journal={arXiv preprint arXiv:2508.00024}, year={2024}, url={https://arxiv.org/abs/2508.00024} }