Speakers
Description
Understanding generalization in quantum machine learning, in particular the phenomenon of double descent, is a crucial open problem, especially as these models move into overparameterized regimes. Yet for quantum extreme learning machines (QELMs), the precise relationship between architectural choices, physical parameters, and the resulting shape of the risk curve remains poorly understood. We address this gap using random matrix theory to establish general bounds expressed through the population covariance, which we then study analytically by working in the Fourier representation of QELM feature maps. This approach reveals that stronger input encoding, additional input reuploading, and finite-shot sampling noise all contribute to improved generalization, and it lets us pinpoint the conditions under which a double descent risk curve persists.
| Topic | Hybrid quantum-classical approaches for learning and optimization |
|---|---|
| Additional Topic | Encoding and processing of data in quantum systems |