A capstone project for The Coding School — Qubit x Qubit program, continued investigating how quantum support vector machines (QSVMs) perform across three major quantum frameworks: PennyLane, Cirq, and Qiskit. This work reveals a critical insight often overlooked in quantum ML literature: framework choice significantly impacts classical baseline performance — sometimes more than the quantum model itself.
While SVM accuracy remained stable at 99%. In PennyLane and Cirq implementations, accuracy varied dramatically (45.0% → 71.5% → 88.0%) depending solely on framework integration overhead. Training time similarly spanned three orders of magnitude (3.33s → 50.27s → 1057.90s).
💡 Novelty: First empirical demonstration that framework-specific implementation details—not quantum advantage alone—drive performance variations in QSVM benchmarking. Most prior work evaluates QSVMs within single frameworks; we benchmark identical models across three.
| Framework | Model | Accuracy (%) | Time (s) | Key Insight |
|---|---|---|---|---|
| PennyLane (NumPy) | QSVM | 71.5 | 3.33 | Baseline impacted by minimal overhead |
| Classical SVM | 99.0 | 0.00 | Fastest simulation; lightweight wrapper | |
| Cirq (TFQ) | QSVM | 88.0 | 50.27 | Moderate integration overhead |
| Classical SVM | 99.0 | 0.02 | Balanced performance | |
| Qiskit (Aer) | QSVM | 45.0 | 1057.90 | ComputeUncompute bottleneck |
| Classical SVM | 99.0 | 0.00 | Full circuit simulation overhead |
-
Dataset: MNIST binary classification (digits 0 vs 1)
- 14,780 total samples → 100 training / 200 test (NISQ constraints)
- PCA dimensionality reduction to 2 features → 2-qubit amplitude encoding
- Quantum Feature Map: ZZFeatureMap (2 repetitions) with RY rotations + CZ entanglement
-
Kernel: Fidelity-based quantum kernel
$K(x_i,x_j) = |\langle 0|U^\dagger(x_i)U(x_j)|0\rangle|^2$ -
Classical Baseline: scikit-learn SVM with RBF kernel (
$C=1.0$ ) - Hardware: Google Colab runtime (identical hardware across all runs)
pip install scikit-learn numpy matplotlib pandas
pip install pennylane # For PennyLane notebook pip install cirq # For Cirq notebook pip install qiskit # For Qiskit notebook pip install qiskit-machine-learning qiskit-aer qiskit-algorithms # Qiskit ML stack