Capstone B: Quantum Classifier
The convergence project: quantum and hybrid classifiers on real radio signal data, benchmarked honestly against your Capstone A model. This is research-grade work - the same experiment shape as current "quantum image classification" and quantum pattern recognition papers, pointed at your domain.
The brief
Current devices and simulators cannot eat 128 raw I/Q samples. The experiment is therefore: compress classically, classify quantumly, compare honestly. Restrict RML2016.10a to 4 well-separated classes at high SNR, reduce each frame to 8-16 features, and put quantum models against matched classical baselines.
The compression pipeline
Tasks
- Build two feature reducers from your Capstone A data: (a) PCA to 8 and 16 dimensions; (b) a small autoencoder bottleneck of the same sizes. Normalize features to the encoding-friendly range.
- Sanity baseline: logistic regression and a small MLP on the compressed features. If the MLP gets 95 percent, your task is too easy - tighten it (more classes or lower SNR) until classical-on-compressed has real headroom.
- Freeze the datasets to disk. Every model from now on sees identical splits.
Definition of done
A frozen, documented 4-class dataset in 8 and 16 dimensions, with classical baseline numbers recorded.
Quantum models
Tasks
- Model 1 - variational classifier: angle encoding on 8 qubits, strongly entangling layers (your Week 22 build, scaled up).
- Model 2 - QCNN: amplitude encoding of 16 features into 4 qubits (or angle encoding on 8), then your Week 25 Cong-style QCNN.
- Model 3 - hybrid: your Capstone A CNN's early layers as a frozen feature extractor, quantum circuit as the classification head, trained end to end through PyTorch + PennyLane.
- Fairness rule: for every quantum model, train a classical model of approximately equal trainable parameter count on the same features. This comparison is the intellectual core of the capstone.
Definition of done
A results table: model, parameter count, test accuracy, wall-clock training time - quantum and classical side by side.
Noise, hardware, and the write-up
Tasks
- Rerun your best quantum model under a realistic noise model (PennyLane or Qiskit Aer noise simulation); quantify the accuracy drop.
- Stretch: run inference for a handful of test samples on a real IBM device with your smallest model. Even 4 qubits and 50 test points is a genuine hardware result.
- Write the final report in your repo: setup, results table, noise study, and a straight answer to "did quantum help, and under what conditions might it?" Negative or neutral results written up well are the norm in this field - that is what current papers look like.
Definition of done
A report a QML researcher would respect: matched baselines, frozen splits, noise numbers, no overclaiming.
You have trained CNNs on the field's standard radio benchmark, built and trained QCNNs, run on real quantum hardware, and produced an honest quantum-vs-classical study on your own domain. You can now read any paper on the QML frontier - including everything left on your original list - as a peer rather than a student. Next, Phase 7 pushes your classical skills to the current research frontier, and Phase 8 lands on the quantum technology that actually touches radio today.