Quantum Datasets and Benchmarks: Foundations for Evaluating Quantum Machine Learning

Table of Contents

1. Introduction
2. Why Datasets Matter in QML
3. Classical vs Quantum Datasets
4. Synthetic Datasets for Quantum ML
5. Real-World Use Cases for Quantum Datasets
6. Benchmarking in Classical ML vs QML
7. Types of Quantum Datasets
8. Quantum-Classical Hybrid Datasets
9. Dataset Formats and Representations
10. Encoding Datasets into Quantum Circuits
11. Quantum Dataset Libraries and Platforms
12. IBM Qiskit Datasets and qiskit-machine-learning
13. PennyLane Datasets and QML Benchmarks
14. TFQ Datasets and Integration
15. Notable Quantum Benchmarks
16. Quantum Dataset Generation Techniques
17. Evaluation Metrics in QML Benchmarks
18. Challenges in Dataset Standardization
19. Open Source Quantum ML Datasets
20. Conclusion

1. Introduction

Quantum machine learning (QML) requires appropriate datasets and benchmarks to compare models, evaluate algorithms, and validate performance. As the field evolves, the creation and standardization of quantum datasets are becoming increasingly important.

2. Why Datasets Matter in QML

• Provide ground truth for training and validation
• Enable reproducibility of experiments
• Support fair comparison between quantum and classical models

3. Classical vs Quantum Datasets

4. Synthetic Datasets for Quantum ML

• Iris dataset (projected into quantum encodings)
• Parity classification
• Quantum state discrimination
• XOR problem in quantum space

5. Real-World Use Cases for Quantum Datasets

• Quantum chemistry states
• Material simulations (e.g., lattice models)
• Financial time series encoded in qubit registers

6. Benchmarking in Classical ML vs QML

• MNIST, CIFAR-10 in classical ML
• No widely accepted standard yet in QML
• Most studies use simulated or re-encoded datasets

7. Types of Quantum Datasets

• Labeled qubit states
• Quantum circuits as data points
• Quantum trajectories and time evolution data

8. Quantum-Classical Hybrid Datasets

• Classical data encoded in quantum circuits (e.g., angle encoding)
• Used for hybrid models and transfer learning

9. Dataset Formats and Representations

• NumPy arrays for parameters
• Qiskit QuantumCircuit objects
• PennyLane templates with labels

10. Encoding Datasets into Quantum Circuits

• Angle Encoding: \( x_i
  ightarrow RY(x_i) \)
• Amplitude Encoding: normalize data vector and map to amplitudes
• Basis Encoding: binary feature maps to qubit states

11. Quantum Dataset Libraries and Platforms

• Qiskit’s datasets and qiskit_machine_learning.datasets
• PennyLane’s qml.datasets module
• TFQ’s tfq.datasets

12. IBM Qiskit Datasets and qiskit-machine-learning

• Ad hoc dataset loaders (e.g., ad_hoc_data)
• Iris, breast cancer, quantum-enhanced classification tasks

13. PennyLane Datasets and QML Benchmarks

• qml.datasets.qcircuits() for circuit generation
• Integration with PyTorch and TensorFlow

14. TFQ Datasets and Integration

• TFQ provides datasets in TensorFlow tensor format
• Supports quantum-enhanced layers on top of classical embeddings

15. Notable Quantum Benchmarks

• VQE on molecule datasets (H2, LiH, BeH2)
• QAOA on graph optimization
• Quantum kernel classification (synthetic vs noisy data)

16. Quantum Dataset Generation Techniques

• Generate circuits with specific entanglement properties
• Simulate Hamiltonian dynamics
• Create oracle-based classification labels

17. Evaluation Metrics in QML Benchmarks

• Accuracy, precision, recall (classification)
• Fidelity with target quantum states
• Cost function convergence and gradient norms

18. Challenges in Dataset Standardization

• Lack of large-scale quantum-native datasets
• Hardware dependence of results
• Reproducibility due to shot noise and backend drift

19. Open Source Quantum ML Datasets

• Pennylane QHack challenges
• QML community benchmarks on GitHub
• Synthetic generators like QSet and QData

20. Conclusion

Quantum datasets and benchmarks are crucial to the development and evaluation of QML models. As quantum hardware scales and software matures, more standardized and diverse datasets will become available, enabling meaningful comparisons and progress across the field.