Quantum Neural Network Research
My research focuses on understanding how Hybrid Quantum Neural Networks behave under real-world hardware imperfections, specifically different types and magnitudes of quantum noise. Using PennyLane and PyTorch, I built and trained QNN models while varying qubit counts, applying amplitude damping and depolarizing noise, and running more than 60 controlled experiments to measure changes in accuracy. I visualized these results through surface plots and trend graphs to identify patterns in how noise affects model performance and where QNNs may still outperform or fail compared to classical neural networks. This project allowed me to blend quantum theory, machine learning, and experimental design, ultimately giving me insight into the challenges of making quantum models practical and robust.
Figure: Accuracy surface plot showing how Hybrid Quantum Neural Network performance changes with increasing noise levels and different qubit counts. This visualization highlights the trade-offs between circuit size and noise resilience, revealing where QNNs maintain stability—and where accuracy begins to degrade.
Code snippet from a quick 3-epoch test run used to check that the QNN, noise channels, and training loop were working correctly before running the full 5-epoch experiments.
Figure: Parameterized quantum circuit (PQC) used in the Hybrid Quantum Neural Network model. The architecture combines learnable RY rotations with entangling CNOT layers across four qubits, enabling the circuit to encode input data and capture non-classical correlations before measurement. This structure forms the core of the QNN’s feature mapping and learning process.
Early testing output for the QNN (5 qubits, amplitude-damping noise = 0.06). This was a 3-epoch trial run used only to validate training behavior before running the full 5-epoch experiments included in the research results.