Origin of trivial datasets in QML benchmarking

Determine whether the widespread use of trivial datasets in quantum machine learning benchmarking is primarily a consequence of positive selection bias in task choice or reflects an underlying phenomenon of physical problems that makes such datasets prevalent.

Background

The paper argues that Quantum Convolutional Neural Networks (QCNNs) often succeed because they are benchmarked on locally-easy datasets, for which their action can be simulated classically using low-bodyness observables and classical shadows. The authors emphasize the need for non-trivial datasets to assess genuine quantum advantages and caution against relying on trivial tasks in evaluations of QML models.

In their concluding remarks, the authors pose a broader meta-level question about the field: whether the prevalence of trivial datasets used in QML benchmarking arises from a positive bias in dataset selection or from deeper physical reasons, thereby directing attention to understanding the nature of practical QML tasks and their difficulty.