Efficient Quantum Neural Networks via Knowledge Distillation

16x model compression for NISQ devices with 86.10% accuracy.

NISQ-Era Quantum Machine Learning

This project develops a teacher-student quantum learning workflow for resource-constrained quantum devices. A classical teacher transfers softened knowledge to a compact 4-qubit hybrid quantum neural network, reducing trainable parameters while preserving strong classification performance.

Knowledge distillation methodology for efficient quantum neural networks
Knowledge distillation pipeline: a classical teacher guides a compact hybrid quantum student through soft targets, hard targets, and KL divergence.
Hybrid quantum neural network student architecture
Student architecture: classical feature extraction maps into a 4-qubit parameterized quantum circuit before final classification.

Highlights

  • Achieved 16x parameter reduction with 86.10% accuracy.
  • Built with IBM Qiskit, PyTorch, parameterized quantum circuits, and custom gradient-based optimization.
  • Targets deployment scenarios where coherence time, qubit count, and circuit depth limit near-term quantum models.
  • Kept as an active project page while the manuscript is being prepared for submission.