A quantum classifier that ties a neural net — and why that's the honest result

11 July 2026 · quantum, qml, vqc, pennylane, honest-benchmarks, classification

A VQC is a quantum circuit with tunable knobs that learns to sort things — here, to separate two interleaving crescent-moon shapes of dots (the classic two-moons test). We train it exactly like any AI: show it examples, let it guess, measure how wrong it was, nudge the knobs, repeat. Then — and this is the whole honesty engine of the project — we race it against ordinary (classical) methods on the same split. A quantum number is never shown alone.

What we built

A configurable re-uploading VQC (angle embedding, strongly-entangling ansatz, data re-uploading on) trained with Adam on two-moons, 3-seed mean, on the CPU statevector simulator. Beside it — mandatory (D2) — logistic regression and a small MLP on the identical splits.

Results

Figures

How every QLAB post is built (on-brand, generated): an experiment folder becomes an honest, visual, reproducible post — baseline beside every quantum number, caveats auto-included.

Training curve: loss falling as the circuit’s knobs are tuned over epochs.

The VQC’s learned decision boundary carving the two crescents apart.

The boundary forming over training — the quantum circuit learning the shape.

Honest caveats

On this run the VQC (0.978) ties the best classical baseline (0.978). Two-moons is a 2-D toy: this shows the framework trains, not quantum advantage.

• Claiming quantum advantage is permanently out of scope (D1) — this is an honest-evaluation portfolio, not a quantum-wins pitch.
• The classical baseline is the credibility engine (D2): the quantum number never ships alone.
• Simulator-only, CPU. Real-hardware columns come later; the classical-simulability audit comes later still.

Reproduce it

Every number above traces to metrics.csv / seed_summary.csv (D7). Config + data + plots: experiments/20260708-105849_two-moons.