Entangled Datasets for Quantum Machine Learning

Louis Schatzki, Andrew Arrasmith, Patrick J. Coles, Marco CerezoarXiv 2021 · arXiv

Introduction: The Need for Quantum-Native Benchmarks

Researchers from Los Alamos National Laboratory are proposing a fundamental shift in how we evaluate Quantum Machine Learning (QML). Currently, most QML research relies on classical datasets, such as MNIST handwritten digits, which are "embedded" into quantum states. However, this approach introduces a significant bottleneck: the choice of embedding can drastically alter the difficulty of a problem, often obscuring the actual performance of the Quantum Neural Network (QNN). Furthermore, it remains theoretically unclear if QML offers a definitive advantage for processing classical data.

To address these issues, the researchers introduced NTangled, a new class of "quantum-native" datasets. These datasets are composed of quantum states where the primary features are physical properties like entanglement, providing a more rigorous and relevant metric for measuring QML performance and potential quantum advantage.

The NTangled Framework: Generation and Classification

The core of this research is a generative-then-discriminative pipeline designed to force QML models to interact with the unique features of quantum mechanics.

Overview of the quantum dataset generation and classification pipeline, showing the generator producing entangled states and the subsequent QNN classification process.

1. Generation

The process begins with a "Generator QNN" trained to produce states with specific levels of Concentratable Entanglement (CE)—a measure of global connectivity between qubits. The researchers tested various circuit architectures, including Hardware-Efficient (HWE), Strongly Entangling (SEA), and Convolutional (CONV) ansatzes, to ensure the dataset's diversity.

2. Classification

Once generated, these states are fed into a Quantum Convolutional Neural Network (QCNN). The model's task is to classify the states based on their entanglement levels or the complexity (depth) of the circuit that created them. A key finding was the "Two-Copy Trick": providing the QCNN with two copies of a quantum state significantly boosted classification accuracy from approximately 70% to over 90%, as entanglement is a nonlinear property that is easier to extract from multiple copies.

Benchmarking for Quantum Advantage

By shifting the focus from classical bits to quantum states, the researchers aim to bypass the "embedding bottleneck" and create a clearer path toward proving quantum advantage.

A comparison table highlighting the differences between classical and quantum data benchmarking, noting the clearer potential for quantum advantage in the latter.

As shown in the comparison above, classical data requires an embedding step that complicates benchmarking. In contrast, using quantum-native data allows for a direct evaluation of the QNN's capabilities. This approach also helps identify scalability issues, such as "barren plateaus," in a context that is physically meaningful for quantum processors.

Key Results and Insights

  • High Generation Success: The Generator QNNs achieved over 95% success in producing states with target entanglement values across systems ranging from 3 to 12 qubits.
  • Robust Classification: The QCNN classified the NTangled dataset with over 90% accuracy when using two-copy inputs. It also distinguished between "shallow" and "deep" circuits with up to 96.5% accuracy in an 8-qubit system.
  • Generalization: Remarkably, a Generator QNN trained on simple computational basis states (0s and 1s) could generalize to produce correct entanglement levels for the entire, infinite space of product states.
  • Hardware Efficiency: These datasets are defined by parametrized circuits, meaning they can be stored as a list of gate angles and reconstructed on quantum hardware without the need for massive density matrices.

Conclusion and Future Directions

While the current study was conducted in noiseless simulations, it establishes a scalable framework for benchmarking QML on systems up to 12 qubits (a 4096-dimensional Hilbert space). Future work will focus on how these entanglement-based labels perform on noisy hardware and exploring more complex entanglement measures, such as "n-tangle," to differentiate between specific types of quantum states like GHZ and W-states. By providing a "MNIST for the quantum era," NTangled offers a standardized way to measure progress toward true quantum advantage.

Summary updated

BibTeX
@misc{schatzki2021entangled,
  author = {Louis Schatzki and Andrew Arrasmith and Patrick J. Coles and Marco Cerezo},
  title = {{Entangled Datasets for Quantum Machine Learning}},
  year = {2021},
  eprint = {2109.03400},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2109.03400},
}
Figure