Hybrid Quantum-Classical Neural Networks Explained Interactively

The three-stage pipeline

Hover or focus any symbol to see what it does. (a) A lattice of qubits prepared in a many-body state with X and Z stabilizers. (b) A parameterized circuit of Y and CZ rotations followed by Z-basis readouts. (c) A classical dense network mapping bits to a phase label y.

Why hybrid?

Quantum feature extraction

The parameterized circuit acts directly on the many-body quantum state, no classical description needed.
Trainable Y and CZ rotation angles learn measurements that expose nonlocal order.
Can access correlations that are exponentially costly to compute classically.

Classical post-processing

A small dense network turns noisy readout bits into a robust decision y.
Cheap, stable gradient-based training on ordinary hardware.
Gradients flow through both parts, so the whole pipeline is optimized end-to-end.

Phase transition simulation

Drag the field strength h. The circuit's learned expectation value stays near +1 in the topological phase and falls after the critical point h_c = 0.5, where the topological entanglement entropy S_topo drops to zero. Data is illustrative, styled after the paper's plots.

Field strength h = 0.00

Phase: topologicalExpectation: 1.00S_topo: 0.69

Worked example: 16 Z-readout bits to a prediction

Click bits to flip measurement outcomes (mapped to +1 or -1), then run the forward pass through a tiny trained dense network to get y, the probability the state is topological.

Press "Run forward pass" to compute y.

Check your understanding

Five quick questions with instant feedback.

FAQ

What is quantum phase recognition?

Deciding which phase of matter a quantum state belongs to, for example topological versus trivial, often using order parameters that are nonlocal and hard to measure directly.

Why not a purely classical network?

The input is a quantum state, not a bitstring. A quantum circuit must first convert it into measurement statistics; the hybrid design lets that conversion itself be learned.

How is it trained?

Circuit angles are updated with parameter-shift gradient estimates while the classical weights use backpropagation, jointly minimizing one classification loss.