A production-grade hybrid quantum-classical machine learning system for text classification. Combines classical NLP preprocessing with a Variational Quantum Circuit (VQC) layer using PennyLane and PyTorch — trained end-to-end via gradient descent through quantum backpropagation.
This project implements a Hybrid Quantum-Classical Neural Network for NLP — one of the most active research areas in quantum machine learning. The key idea:
- Classical computers handle text preprocessing (TF-IDF, PCA) and output classification
- Quantum circuit (VQC) performs non-linear feature transformation in high-dimensional Hilbert space — something exponentially expensive to simulate classically
- The entire pipeline is trained end-to-end using quantum backpropagation (parameter-shift rule)
Raw Text Input
│
▼
┌─────────────────────────────────────────────────────────────┐
│ CLASSICAL PRE-PROCESSING │
│ TF-IDF Vectorizer (bigrams) → PCA → MinMaxScaler [0, π] │
└────────────────────────┬────────────────────────────────────┘
│ n_qubits features
▼
┌─────────────────────────────────────────────────────────────┐
│ CLASSICAL PRE-NET │
│ Linear(n→n) → BatchNorm → Tanh │
└────────────────────────┬────────────────────────────────────┘
│ angles ∈ [0, π]
▼
┌─────────────────────────────────────────────────────────────┐
│ QUANTUM VQC LAYER ⚛️ │
│ │
│ |q0⟩ ──RY(x0)──RZ(θ)──RY(φ)──●─────────────── ⟨Z0⟩ │
│ |q1⟩ ──RY(x1)──RZ(θ)──RY(φ)──⊕──●──────────── ⟨Z1⟩ │
│ |q2⟩ ──RY(x2)──RZ(θ)──RY(φ)─────⊕──●───────── ⟨Z2⟩ │
│ |q3⟩ ──RY(x3)──RZ(θ)──RY(φ)────────⊕───────── ⟨Z3⟩ │
│ │
│ Repeated n_layers times with trainable θ, φ parameters │
└────────────────────────┬────────────────────────────────────┘
│ n_qubits expectation values
▼
┌─────────────────────────────────────────────────────────────┐
│ CLASSICAL POST-NET │
│ Linear(n→32) → BatchNorm → ReLU → Dropout │
│ Linear(32→n_classes) → Softmax │
└────────────────────────┬────────────────────────────────────┘
│
▼
Class Prediction
Each classical feature
RY(x_i)|0⟩ = cos(x_i/2)|0⟩ + sin(x_i/2)|1⟩
This maps the classical input into quantum state (Hilbert space).
Each layer applies parameterized rotations:
RZ(θ_i) RY(φ_i) on qubit i ← trainable parameters
Followed by CNOT entanglement gates creating quantum correlations between qubits.
CNOT(0→1), CNOT(1→2), CNOT(2→3), CNOT(3→0)
Creates correlations between all qubits — capturing relationships between input features.
⟨Z_i⟩ = ⟨ψ|Z_i|ψ⟩ ∈ [-1, +1]
Pauli-Z expectation value collapses quantum state back to classical values.
Parameters are optimized using quantum backpropagation via PennyLane's diff_method="backprop", computing exact gradients through the quantum circuit.
| Feature | Details |
|---|---|
| ⚛️ Quantum Layer | PennyLane VQC with angle encoding + entanglement |
| 🔀 Hybrid Design | Classical pre/post nets + quantum middle layer |
| 📝 NLP Pipeline | TF-IDF (bigrams) + PCA dimensionality reduction |
| 🎓 End-to-End Training | Quantum backpropagation via parameter-shift rule |
| 📊 Comparison | Quantum vs Classical baseline training |
| 🌐 Streamlit App | Interactive demo with circuit visualization |
| 🧪 Unit Tests | Full test suite for all modules |
| 📈 Visualizations | 6 real output plots from actual training run |
Quantum-Classical-NLP/
├── app.py # Streamlit interactive demo
├── train.py # Main training script
├── predict.py # Inference on new text
├── src/
│ ├── __init__.py
│ ├── quantum_circuit.py # VQC definition (PennyLane QNodes)
│ ├── classical_preprocessor.py # TF-IDF + PCA + dataset utilities
│ ├── hybrid_model.py # HybridQuantumClassifier + ClassicalBaseline
│ ├── trainer.py # Training loop + evaluation + early stopping
│ └── visualizer.py # All matplotlib plots (dark theme)
├── tests/
│ └── test_quantum_nlp.py # Unit tests (preprocessing + model + training)
├── outputs/
│ ├── training_history.png # Real training curves
│ ├── quantum_vs_classical.png # Model comparison plot
│ ├── confusion_matrix.png # Test set confusion matrix
│ ├── circuit_diagram.png # VQC architecture diagram
│ ├── feature_importance.png # PCA variance explained
│ └── model_architecture.png # Full hybrid architecture
├── notebooks/ # Jupyter exploration notebooks
├── data/ # Custom datasets (CSV format)
├── requirements.txt
├── .env.example
├── .gitignore
└── README.md
git clone https://github.com/NandithKumar/Quantum-Classical-NLP-Classifier.git
cd Quantum-Classical-NLP-Classifierpython -m venv venv
source venv/bin/activate # Linux/Mac
venv\Scripts\activate # Windowspip install -r requirements.txtNo API keys required — runs fully locally using quantum simulation!
python train.pystreamlit run app.pypython predict.py --text "This product is absolutely amazing!"python -m pytest tests/ -v| Aspect | Classical Model | Quantum-Classical Model |
|---|---|---|
| Feature transformation | Linear / polynomial | Hilbert space (exponential) |
| Entanglement | Not possible | CNOT gates create correlations |
| Parameter efficiency | More params needed | Fewer quantum params, same expressivity |
| Training | Standard backprop | Quantum backpropagation |
| Hardware | CPU / GPU | Quantum simulator (→ real QPU) |
| Component | Technology |
|---|---|
| Quantum Computing | PennyLane 0.36+ |
| Deep Learning | PyTorch 2.0+ |
| NLP Features | Scikit-learn (TF-IDF, PCA) |
| Visualization | Matplotlib (dark theme) |
| Interactive Demo | Streamlit |
| Language | Python 3.8+ |
- Sentiment Analysis — product reviews, social media, customer feedback
- Document Classification — spam detection, topic labeling
- Intent Detection — chatbots, virtual assistants
- Fraud Text Detection — financial NLP
Paladugu Nandith Kumar
- 🎓 B.Tech CSE (AI & ML) — RGMCET, Kadapa
- 🌐 Portfolio
- 📧 nandith1411@gmail.com
- Schuld, M., & Petruccione, F. (2021). Machine Learning with Quantum Computers. Springer.
- Bergholm, V., et al. (2022). PennyLane: Automatic differentiation of hybrid quantum-classical computations. arXiv:1811.04968
- Cerezo, M., et al. (2021). Variational quantum algorithms. Nature Reviews Physics, 3(9), 625-644.
MIT License