Released by
Junkai Wang
Contact email
WangTheoPhys@outlook.com
Method
Quantum Circuit Simulation and Quantum Artificial Intelligence
Challenge issue
This question is related to Issue #149 .
Across both industry and academia, everyone in quantum computing is chasing a single question: how to find the right application where quantum algorithms can demonstrate a genuine quantum advantage. For me personally, given the boom in artificial intelligence and machine learning, and the desire to understand knowledge and the vision to create silicon-based intelligence , I'm very interested in learning-related problems in this space.
In a 2025 paper from Ye et al., the authors propose Quantum Automated Learning (QAL), a quantum-native learning paradigm that departs fundamentally from the conventional variational quantum circuit framework. Instead of representing a model by a parametrized circuit (U(\boldsymbol{\theta})) and training it through repeated gradient estimation and classical parameter updates, QAL contains no variational parameters and reformulates training directly as a quantum state-preparation process. Classical data are encoded into sample-dependent unitary operations, while label-dependent perturbations iteratively drive an initially random quantum state toward a task-specific target state.
The resulting dynamics admit a transparent physical interpretation: the learning process effectively realizes an imaginary-time evolution, or equivalently a dissipative cooling process, toward the ground state of a data-encoded Hamiltonian. Under the assumptions established in the paper, this evolution converges exponentially to a state corresponding to the global minimum of the learning loss. Consequently, trainability is not inferred only from numerical optimization experiments but follows from the spectral and dynamical structure of the underlying quantum process. The method is intrinsically gradient-free and therefore avoids characteristic obstacles of variational quantum learning, including costly parameter-by-parameter gradient estimation, barren plateaus, and convergence to unfavorable local minima.
QAL also possesses provable representation and generalization properties. The authors show that the framework can represent classifications of arbitrary finite datasets with arbitrarily high accuracy, while its generalization error is bounded as $\epsilon_{\mathrm{gen}} O!\left(\sqrt{\frac{\log D}{N}}\right), $ where (D) is the Hilbert-space dimension and (N) is the number of training samples. Numerical studies on classical image datasets and quantum many-body data further support the proposed mechanism and its underlying assumptions. Conceptually, this work suggests that a genuinely quantum-native learner need not imitate a classical neural network by introducing tunable quantum gates. Instead, learning can be formulated as the autonomous physical preparation of a quantum state whose structure directly encodes the solution of the learning problem.
Another complementary landmark is the 2026 work of Zhao et al., including John Preskill, which establishes an unconditional exponential quantum space advantage for learning from massive classical data. Unlike Quantum Automated Learning, it is not parameter-free in a literal sense; rather, its key idea is to avoid explicitly materializing and storing a high-dimensional dataset, parameter vector, or matrix. Through quantum oracle sketching, streaming classical samples generate incremental coherent updates, encoding task-relevant information into an implicit and queryable quantum evolution. Learning is therefore reformulated from optimizing and storing a large classical parameterization to preparing and processing quantum states and operators. Although the proven separation concerns memory rather than generic end-to-end runtime, this work is conceptually important because it addresses the data-loading and readout bottlenecks that undermine many previous QML proposals. Together, these two examples suggest that genuinely quantum-native learning should be built around state preparation, coherent dynamics, and operator queries, rather than merely replacing a classical neural network with a parametrized quantum circuit.
Ideally, we should extend these core features to reinforcement learning, unsupervised learning scenarios, and other research areas.
Released by
Junkai Wang
Contact email
WangTheoPhys@outlook.com
Method
Quantum Circuit Simulation and Quantum Artificial Intelligence
Challenge issue
This question is related to Issue #149 .
Across both industry and academia, everyone in quantum computing is chasing a single question: how to find the right application where quantum algorithms can demonstrate a genuine quantum advantage. For me personally, given the boom in artificial intelligence and machine learning, and the desire to understand knowledge and the vision to create silicon-based intelligence , I'm very interested in learning-related problems in this space.
In a 2025 paper from Ye et al., the authors propose Quantum Automated Learning (QAL), a quantum-native learning paradigm that departs fundamentally from the conventional variational quantum circuit framework. Instead of representing a model by a parametrized circuit (U(\boldsymbol{\theta})) and training it through repeated gradient estimation and classical parameter updates, QAL contains no variational parameters and reformulates training directly as a quantum state-preparation process. Classical data are encoded into sample-dependent unitary operations, while label-dependent perturbations iteratively drive an initially random quantum state toward a task-specific target state.
The resulting dynamics admit a transparent physical interpretation: the learning process effectively realizes an imaginary-time evolution, or equivalently a dissipative cooling process, toward the ground state of a data-encoded Hamiltonian. Under the assumptions established in the paper, this evolution converges exponentially to a state corresponding to the global minimum of the learning loss. Consequently, trainability is not inferred only from numerical optimization experiments but follows from the spectral and dynamical structure of the underlying quantum process. The method is intrinsically gradient-free and therefore avoids characteristic obstacles of variational quantum learning, including costly parameter-by-parameter gradient estimation, barren plateaus, and convergence to unfavorable local minima.
QAL also possesses provable representation and generalization properties. The authors show that the framework can represent classifications of arbitrary finite datasets with arbitrarily high accuracy, while its generalization error is bounded as $\epsilon_{\mathrm{gen}} O!\left(\sqrt{\frac{\log D}{N}}\right), $ where (D) is the Hilbert-space dimension and (N) is the number of training samples. Numerical studies on classical image datasets and quantum many-body data further support the proposed mechanism and its underlying assumptions. Conceptually, this work suggests that a genuinely quantum-native learner need not imitate a classical neural network by introducing tunable quantum gates. Instead, learning can be formulated as the autonomous physical preparation of a quantum state whose structure directly encodes the solution of the learning problem.
Another complementary landmark is the 2026 work of Zhao et al., including John Preskill, which establishes an unconditional exponential quantum space advantage for learning from massive classical data. Unlike Quantum Automated Learning, it is not parameter-free in a literal sense; rather, its key idea is to avoid explicitly materializing and storing a high-dimensional dataset, parameter vector, or matrix. Through quantum oracle sketching, streaming classical samples generate incremental coherent updates, encoding task-relevant information into an implicit and queryable quantum evolution. Learning is therefore reformulated from optimizing and storing a large classical parameterization to preparing and processing quantum states and operators. Although the proven separation concerns memory rather than generic end-to-end runtime, this work is conceptually important because it addresses the data-loading and readout bottlenecks that undermine many previous QML proposals. Together, these two examples suggest that genuinely quantum-native learning should be built around state preparation, coherent dynamics, and operator queries, rather than merely replacing a classical neural network with a parametrized quantum circuit.
Ideally, we should extend these core features to reinforcement learning, unsupervised learning scenarios, and other research areas.