Benchmarking Neural Networks For Quantum Computations

Nam H. Nguyen, Elizabeth C. Behrman, Mohamed A. Moustafa, James Edward SteckIEEE TNNLS 2020 · DOI

TL;DR

This paper benchmarks classical neural networks (real and complex-valued) against a quantum neural network (QNN) on both classical and quantum computing tasks. The QNN demonstrates significantly superior efficiency and performance, particularly in learning complex quantum phenomena like entanglement, suggesting a powerful new paradigm for machine learning and quantum computation.

What Problem Does This Paper Address?

The paper tackles the challenge of understanding and harnessing the potential of quantum computers. While quantum computers promise exponential speedups for certain problems, identifying "quantum advantage" and developing truly quantum algorithms remains difficult. This work explores how machine learning techniques, specifically neural networks, can be adapted for quantum computation. The authors aim to benchmark the performance of classical neural networks against a quantum neural network (QNN) to understand where and how quantum approaches offer advantages, both for classical tasks and for inherently quantum problems.

Methods

The paper compares three types of neural networks:

  • Classical Real-Valued Neural Networks (RVNN): Standard artificial neural networks using real numbers for weights and activations. These are known universal approximators for continuous functions.
  • Classical Complex-Valued Neural Networks (CVNN): An extension of RVNNs that use complex numbers. This allows for scaling and rotation of signals, potentially leading to more powerful and efficient architectures, as demonstrated by their ability to solve problems like XOR in a single layer, which RVNNs require a hidden layer for.
  • Quantum Neural Network (QNN): A model proposed by the authors that leverages quantum mechanical principles. The QNN's "computation" is driven by the time evolution of a quantum system governed by Schrödinger's equation. Parameters are adjusted using a quantum analogue of backpropagation. The key difference is that the QNN directly manipulates quantum states and their evolution, rather than simulating them.

The benchmarking involves training these networks on:

  1. Classical problems: Logic gates (AND, NAND, OR, NOR, XOR, XNOR) and Iris flower classification.
  2. Quantum problem: Learning an entanglement witness for a two-qubit system.

Datasets

  • Logic Gates: A synthetic dataset representing the four possible input pairs for each logic gate (e.g., {(00), (01), (10), (11)}).
  • Iris Flower Classification: The standard Iris dataset, containing 150 samples with four features (sepal length, sepal width, petal length, petal width) across three species.
  • Entanglement Witness: A synthetic dataset consisting of pure two-qubit states (initially 4 states, later expanded to 100) used to train the networks to identify entanglement.

Key Insights and Ideas

  • CVNNs offer a significant advantage over RVNNs for classical tasks: Complex-valued networks can solve non-linear problems like XOR with a single layer and require far fewer training epochs, demonstrating greater efficiency.
  • QNNs are exceptionally efficient for classical tasks: For logic gates, the QNN achieved <1% RMS error in a single epoch, vastly outperforming both classical networks. This suggests that even for problems traditionally solved classically, a quantum approach can be remarkably faster to train.
  • QNNs excel at inherently quantum problems: The QNN was able to learn an entanglement witness with minimal training (100 epochs for 100 states) and achieved very low testing errors. Classical networks struggled significantly to learn this, even with larger training sets and more epochs, highlighting the QNN's unique ability to leverage quantum phenomena.
  • Quantum advantage is not just about speedup, but also about how problems are solved: The QNN's ability to directly manipulate quantum states and leverage interference is key to its performance, especially for quantum tasks.

Key Results

  • Logic Gates:
    • RVNNs and CVNNs required thousands of epochs to reach 1% RMS error for linear gates, and struggled to learn non-linear gates (XOR/XNOR) in a single layer.
  • CVNNs were significantly faster than RVNNs, requiring 30-50 times fewer iterations for linearly separable gates.
  • The QNN achieved <1% RMS error in just one epoch for all logic gates, including non-linear ones.
  • Iris Classification:
    • CVNNs showed improved testing accuracy compared to RVNNs as the hidden layer size increased.
  • While RVNNs required 50,000 epochs and CVNNs 1,000 epochs, the QNN achieved comparable results with only 100 epochs.
  • Entanglement Witness:
    • Classical networks (RVNN, CVNN, NeuralWorks) struggled to generalize and had high testing errors for entanglement detection, even with large training sets.
  • The QNN achieved very low training and testing RMS errors, demonstrating its ability to learn and generalize entanglement properties efficiently.

Limitations and Future Work

  • Theoretical guarantees for QNNs: The authors acknowledge that while theorems exist for classical neural networks (universal approximation), precise theoretical guarantees for the complexity and power of their specific QNN model are still under development.
  • Comparison of QNN architectures: The paper focuses on a specific QNN model. Further work is needed to explore and benchmark other QNN architectures.
  • "Quantum-inspired" vs. fully quantum: The paper distinguishes between "quantum-inspired" (like CVNNs) and fully quantum networks. While CVNNs show improvements, the QNN demonstrates a more profound advantage, especially for quantum tasks.
  • Broader applications: The authors suggest that exploring other quantum machine learning architectures and identifying a wider range of problems that benefit from fully quantum mechanical approaches is crucial. They also note that physical implementations will introduce natural nonlinearities that can be leveraged.

Summary updated

BibTeX
@article{nguyen2020benchmarking,
  author = {Nam H. Nguyen and Elizabeth C. Behrman and Mohamed A. Moustafa and James Edward Steck},
  title = {{Benchmarking Neural Networks For Quantum Computations}},
  journal = {IEEE Trans. Neural Networks Learn. Syst.},
  year = {2020},
  doi = {10.1109/tnnls.2019.2933394},
}
Figure