Speaker
Description
Data loading is a bottleneck for near-term quantum machine learning: angle, amplitude, and basis encodings either underuse the exponential Hilbert space or need circuit depths beyond noisy intermediate-scale quantum (NISQ) coherence budgets. We introduce Shot-Based Quantum Encoding (SBQE), which embeds a datum in a data-dependent classical distribution over a fixed basis-state pool, realised by allocating shots rather than coherent gates. The distribution is learned from a few principal components through a trainable affine layer, giving a vector over the $256$ basis states of eight qubits. A fraction of the shots then starts in each basis state before the circuit runs; no data-dependent gates are needed, and coherent depth is set by the ansatz.
Because the encoded state $\rho(\mathbf{x})=\mathrm{diag}(\mathbf{p}(\mathbf{x}))$ is diagonal, measured expectation values are linear in $\mathbf{p}(\mathbf{x})$. A single SBQE layer is thus equivalent to a perceptron whose weight matrix is supplied by the circuit; stacked layers form a multilayer network, all weights inside it.
On three image benchmarks (exact statevector simulation, ten seeds), SBQE reaches $89.1\%$ (Semeion), $80.95\%$ (Fashion-MNIST), and $90.25\%$ (MNIST), exceeding amplitude encoding on all three (paired $t$-test, $p<0.01$) and matching a width-matched classical network. A rank sweep shows it reaches this accuracy with only $120$ trainable angles, matching classical maps with an order of magnitude more parameters; the separation is carried by the circuit, not the preprocessing. Accounting for finite-shot and scaling effects, shot allocation is a practical, gate-free channel for data encoding on near-term hardware.
| Topic | Quantum variational circuits and their applications |
|---|---|
| Additional Topic | Quantum algorithms for machine learning applications |