[LR] quEEGNet: QAI for Biosignal Processing

[This review is intended solely for my personal learning]

Paper Info

arXiv: 2210.00864v1
Title: quEEGNet: Quantum AI for Biosignal Processing
Authors: Toshiaki Koike-Akino, Ye Wang

Prior Knowledge

Biosignal Processing in Human-Machine Interfaces (HMI)

Biosignals such as EEG (electroencephalogram), EMG (electromyogram), and ECoG (electrocorticogram) are fundamental to developing brain-computer interfaces (BCI) and HMIs. However, biosignals are subject to high variability and noise, requiring robust machine-learning models for feature extraction and classification.

Challenges in Biosignal Processing

• High inter-subject variability necessitates frequent recalibration.
• Computational inefficiency of deep learning models in resource-constrained environments.
• Scalability and generalization remain concerns in BCI applications.

Goal

The paper introduces a hybrid quantum-classical deep learning model called quEEGNet, which integrates a variational quantum circuit (VQC) into a deep neural network (DNN) for biosignal analysis. The key objectives are:

1. Demonstrating quantum-enhanced feature extraction for EEG, EMG, and ECoG.
2. Reducing the number of trainable parameters while maintaining high classification accuracy.
3. Providing a proof-of-concept study showcasing quantum computing’s applicability in HMI/BCI.

Method

Quantum Artificial Intelligence (QAI) Framework

1. Quantum Neural Network (QNN) Architecture:

   • The model integrates VQC-based feature extraction into a classical DNN.
   • A hybrid training strategy is used where VQC layers are optimized alongside DNN layers.

2. Variational Quantum Circuit (VQC) Design:

   • Employs Pauli-Y rotations and controlled-Z entanglement layers for quantum state evolution.
   • Uses Simplified 2-Design (S2D) ansatz to mitigate barren plateau issues.

3. Hybrid Quantum-Classical Processing:

   • EEG/EMG/ECoG signals are preprocessed and embedded into quantum states.
   • QNN performs feature transformation before passing the output to EEGNet, a well-known deep learning model for biosignal classification.

Experimental Setup

• Datasets: Multiple publicly available physiological datasets (EEG, EMG, ECoG) were used for evaluation.
• Baseline Comparison: The model is benchmarked against EEGNet, a classical convolutional neural network.
• Training Details:
  • Optimized using Adam optimizer with a learning rate of 0.1.
  • Implemented using PennyLane and PyTorch.
  • Quantum circuits simulated due to current hardware constraints.

Results

• Performance Improvement: quEEGNet outperformed EEGNet across all datasets, demonstrating the effectiveness of quantum-assisted feature extraction.
• Parameter Efficiency: The model reduces the number of trainable parameters while maintaining competitive performance.
• Computational Trade-offs: While QNN offers advantages in parameter efficiency, the practical speedup depends on the maturity of quantum hardware.

Conclusion

The study presents quEEGNet as a pioneering effort in applying quantum AI to biosignal processing. By integrating variational quantum circuits into deep learning models, the approach achieves state-of-the-art performance across multiple datasets while maintaining computational efficiency. This work establishes QML as a viable direction for future BCI and HMI applications.

Limitations

1. Quantum Hardware Constraints:
   • Experiments were conducted on simulated quantum processors, which may not fully reflect real-world hardware performance.
   • The practical deployment of quEEGNet depends on advancements in near-term quantum computing.
2. Scalability:
   • While the hybrid model reduces parameter count, the benefits of QNNs over classical DNNs are not yet definitive.
   • Further optimization, such as AutoQML, is necessary to enhance generalization.
3. Limited Variants of QNN:
   • Only one type of VQC architecture was explored.
   • Future work should investigate alternative quantum ansatz and hybrid architectures.

References

• The paper: https://arxiv.org/abs/2210.00864
• This review was written with the assistance of Generative AI and is based on the content and results presented in the original paper.