Inherent interpretability provides inherent value in quantum machine learning
Abstract
The field of quantum machine learning (QML) evolved to value models believed to most directly rival those providing utility in classical ML, namely large-scale neural networks. This pursuit of quantum neural-network analogues has produced important theoretical insights for expressivity and trainability at scale; however, due to our inability to conduct large-scale empirical studies, this direction has contributed to skepticism that QML will provide a compelling application for quantum computers. In the meantime, classical ML has been learning a hard lesson with respect to deploying un-interpretable neural networks in the wild: model interpretability matters for human adoption. We adopt this broader ML perspective to encourage a research direction in QML that prioritizes discovering and understanding tools from quantum information theory that applied ML researchers can use to construct inherently interpretable models: models with intentional mathematical structure that guarantees domain-aligned behavior. We first exemplify this research direction through a thread of work in a neighboring area---Bayesian ML. Specifically, we discuss an efficient mathematical tool that emerged from a connection between Gaussian processes and Bayesian linear models, which has since been widely studied and used by applied researchers for problem-specific model design. We then discuss how prior QML research can be reframed as progress towards this direction, despite having been pursued with different goals in mind. Finally, we propose an initial sketch of future work in this direction, with the broader aim of better supporting the needs of applied ML researchers.
Related
- Inherent Turbulence Immunity of Vector Vortex Beams in Free Space Quantum Key Distribution
- Time-reversal symmetry breaking and inherent Josephson diodes in a multi-component order parameter superconductor
- Observation of the Inherent Chiral Smith-Purcell Effect via Symmetry Breaking
- Inherent altermagnetism in minimal tight-binding models of regular hyperbolic lattices
- Inherent Failure Modes of Physics-Informed Neural Networks for Buoyancy-Driven Flow: Role of Non-Dimensionalization in Network Training
BibTeX
@misc{gili2026inherent,
title = {{Inherent interpretability provides inherent value in quantum machine learning}},
author = {Kaitlin Gili and Zachary P. Bradshaw},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2607.13827},
eprint = {2607.13827},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2607.13827},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Inherent interpretability provides inherent value in quantum machine learning
AU - Kaitlin Gili
AU - Zachary P. Bradshaw
PY - 2026
DA - 2026/10/24
PB - arXiv
DO - 10.48550/arxiv.2607.13827
UR - https://arxiv.org/abs/2607.13827
AB - The field of quantum machine learning (QML) evolved to value models believed to most directly rival those providing utility in classical ML, namely large-scale neural networks. This pursuit of quantum neural-network analogues has produced important theoretical insights for expressivity and trainability at scale; however, due to our inability to conduct large-scale empirical studies, this direction has contributed to skepticism that QML will provide a compelling application for quantum computers. In the meantime, classical ML has been learning a hard lesson with respect to deploying un-interpretable neural networks in the wild: model interpretability matters for human adoption. We adopt this broader ML perspective to encourage a research direction in QML that prioritizes discovering and understanding tools from quantum information theory that applied ML researchers can use to construct inherently interpretable models: models with intentional mathematical structure that guarantees domain-aligned behavior. We first exemplify this research direction through a thread of work in a neighboring area---Bayesian ML. Specifically, we discuss an efficient mathematical tool that emerged from a connection between Gaussian processes and Bayesian linear models, which has since been widely studied and used by applied researchers for problem-specific model design. We then discuss how prior QML research can be reframed as progress towards this direction, despite having been pursued with different goals in mind. Finally, we propose an initial sketch of future work in this direction, with the broader aim of better supporting the needs of applied ML researchers.
ER -