Quantum AI simulator using a hybrid CPU-FPGA approach.

Area of Science:

• Quantum Computing
• Machine Learning
• Computer Engineering

Background:

• Quantum kernel methods are promising for quantum machine learning but limited by current quantum hardware.
• Existing quantum computers have few qubits, restricting the number of features usable in quantum kernels.
• Efficient classical simulation of quantum kernels is crucial for practical applications.

Purpose of the Study:

• To develop an efficient, application-specific simulator for quantum kernels using classical hardware.
• To implement and evaluate a quantum kernel designed for image classification on a Field-Programmable Gate Array (FPGA).
• To demonstrate the scalability and performance of this hybrid approach.

Main Methods:

• Co-design of an application-specific quantum kernel tailored for image classification.
• Implementation of the quantum kernel estimation on a heterogeneous CPU-FPGA computing architecture.
• Numerical simulation of gate-based quantum kernels with up to 780-dimensional features.

Main Results:

• The FPGA implementation achieved a 470x speedup in quantum kernel estimation compared to CPU-only methods.
• The system successfully simulated one of the largest gate-based quantum kernels to date.
• Performance on the Fashion-MNIST dataset showed the quantum kernel is comparable to optimized Gaussian kernels.

Conclusions:

• Heterogeneous CPU-FPGA computing offers a viable and efficient solution for simulating large-scale quantum kernels.
• This approach overcomes current hardware limitations, enabling more realistic quantum machine learning research.
• The developed quantum kernel and FPGA implementation show potential for practical image classification tasks.