Kuantum sinir ağı eğitim algoritmalarının karşılaştırılması

In recent years, the computational load of machine learning algorithms significantly increased. This fact led to the research of novel training techniques of artificial neural networks. In this study, we challenge the classical Gradient Descent algorithm and compare various classical and quantum techniques in terms of time and accuracy. Classical algorithms include classical optimization algorithms such as Gradient Descent with ADAM, Branch and Cut and Simulated Annealing. For quantum, among possible algorithms BF-DCQO, Quantum Annealing and QAOA are included. We developed a Python code that trains fully connected neural networks converting them to constrained/unconstrained binary programs. We tested Branch and Cut and BF-DCQO algorithms with small and medium sized networks utilizing the code. Eventually, we demonstrated that these model-based approaches fall short against Gradient Descent, thus iterative training methods were found to be more preferable than one-shot methods. Despite hardware limitations, quantum algorithms were found to be promising if run on sufficiently capable quantum hardware.