Bayesian convolutional front-end based uncertainty-aware hybrid quantum-classical image classification.

Area of Science:

• Quantum Machine Learning
• Artificial Intelligence
• Quantum Computing

Background:

• Noisy Intermediate-Scale Quantum (NISQ) devices face challenges like noise sensitivity and overfitting.
• Existing quantum neural networks (QNNs) struggle with miscalibrated predictive confidence.

Purpose of the Study:

• To propose an uncertainty-aware hybrid Bayesian quantum neural network (BQNN) to address limitations in NISQ devices.
• To evaluate the performance and robustness of BQNN variants against QCNN baselines.

Main Methods:

• Developed a three-stage pipeline: Bayesian feature extraction, quantum state evolution via a parameterized quantum circuit (PQC), and classical decision making.
• Employed a Bayesian convolutional front-end for feature extraction from 28x28 grayscale images.
• Utilized two PQC designs for quantum classification on 4 qubits, encoding data using rotation angles.

Main Results:

• BQNN variants consistently outperformed QCNN baselines on MNIST and Fashion-MNIST datasets.
• Achieved high test accuracies: 95.46% on MNIST and 87.59% on Fashion-MNIST.
• Demonstrated improved calibration (lower ECE) and faster convergence than QCNNs, with enhanced robustness under simulated depolarizing noise.

Conclusions:

• The proposed BQNN offers a promising approach for more accurate and reliable quantum machine learning on NISQ hardware.
• BQNN exhibits superior performance, faster convergence, and better noise resilience compared to traditional QCNNs.
• The uncertainty-aware nature of BQNN leads to reduced overconfident predictions and improved model calibration.