arXiv:2609.28229
Distilling Datasets into Shallow Circuits for Quantum Machine Learning
Guang Lin, Qibin Zhao
Abstract
In quantum machine learning, training a quantum model requires each sample to be prepared as a quantum state by a loading circuit that must be re-executed for every shot at every training step. The total burden therefore scales with both the number of samples and the cost of preparation. Existing approaches reduce the loading cost of individual inputs or distill data and compress their representations before applying a separate quantum encoding. This separation can leave resulting samples costly to prepare or cause additional loss of the information during subsequent compilation into shallow circuits. We propose quantum dataset distillation (QDD), which distills the full dataset directly into a small set of shallow circuits. Each synthetic sample is parameterized as a staircase circuit corresponding to a low-rank tensor network, making the sample and its loading circuit the same object. The circuit parameters are optimized classically with exact gradients using distribution matching under an explicit loading budget, without post-hoc state-preparation synthesis. On MNIST and Fashion-MNIST, only 10 circuits per class ($0.0017\times$ the full dataset size) achieve accuracy comparable to full-data training and outperform selection baselines under matched sample and loading budgets. In finite-shot training, QDD reaches 95\% of the full-data accuracy with more than $100\times$ fewer cumulative shots. Additionally, we validate QDD on real quantum hardware, demonstrating its practical deployment potential.
AI Summary
Quantum dataset distillation (QDD) addresses the recurring cost of loading classical training data into quantum models: both the number of samples and their preparation circuits contribute to execution costs. It replaces a dataset with a small labeled set of synthetic samples parameterized directly as shallow loading circuits, avoiding post-distillation state-preparation synthesis. Each sample uses an $L$-layer nearest-neighbor staircase circuit with $2(n-1)L$ CNOTs and MPS bond dimension at most $2^L$. Offline classical optimization uses exact simulated gradients to match class-conditional measurement distributions across randomly initialized quantum models, together with class-mean density matrices. The resulting loaders remain fixed during downstream model training. On eight-qubit, ten-class MNIST and Fashion-MNIST tasks, 100 synthetic circuits—ten per class, each with 42 CNOTs—yield test accuracies of 81.62% and 69.67%, respectively, averaged over three training seeds. Comparisons include six selection or clustering baselines and matched-objective distill-then-compress controls. Joint optimization gives its largest reported advantage over the latter at the tightest loading budget: 9.25 percentage points on MNIST with ten samples per class and 14 CNOTs. Finite-shot simulated training reaches 95% of full-data reference accuracy with 134–576-fold fewer shots than full-batch training; savings versus matched-batch-size training are 1.9–2.3-fold at the two higher shot settings. Hardware inference on 128 MNIST images across three IBM processors gives approximately 79–82% accuracy. The stated limitations are scaling classical distillation to larger quantum systems while retaining useful information in shallow circuits, and the untested extension to regression and generative modeling.
Central claim may not hold: The abstract advertises >100× fewer training shots without specifying the full-batch baseline, whereas the matched-batch full-data comparison establishes only approximately 2.3× and 1.9× savings at the two shot settings where both methods reach the target.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.