Speaker
Description
The double descent phenomenon challenges traditional statistical learning theory by showing that increasing model size beyond the interpolation threshold need not worsen performance on unseen data and can, in some regimes, improve it. While this behavior has been widely studied in classical machine learning, its role in quantum machine learning remains less understood. In this work, we investigate double descent across two central quantum learning paradigms: quantum kernel methods and gradient-trained parameterized quantum circuits. For quantum kernel methods, we analytically show that linear regression in quantum feature spaces can exhibit double descent by drawing on tools from classical linear regression and random matrix theory. For parameterized quantum circuits, under the assumption of successful training, we provide analytical results showing that increasing the number of trainable parameters relative to the training set size can improve performance on unseen data. We support these findings with numerical experiments across multiple real-world datasets, circuit architectures, and system sizes, observing characteristic test-error peaks and subsequent performance improvements in overparameterized regimes. Together, our results provide evidence that quantum models can enter modern, highly parameterized learning regimes without necessarily suffering from overfitting. This offers a principled perspective on the scaling behavior of quantum machine learning models and suggests that increasing model capacity may be beneficial for generalization in both quantum feature-space methods and deep parameterized quantum circuits.
| Topic | Quantum variational circuits and their applications |
|---|---|
| Additional Topic | Quantum kernel methods and their applications |