Seokwon Yoo, Jeongho Bang, Changhyoup Lee, Jinhyoung Lee
Quantum-mechanical approach yields much better performance compared to the classical approach in a wide range of applications. This qualitative improvement is enabled by appropriate use of novel quantum effects. We investigate whether this paradigm is also valid for machine learning. In particular, the question we address here is: Can the machine learning be improved by using favorable quantum effects? To answer this question, we compare two controllable circuits, designed to learn a task of N-bit Boolean function: One is classical and the other is quantum. In the comparison, we highlight the local solution-region of the whole search space, so-called acceptable region. It is shown that the acceptable region of the quantum circuit can always be wider than that of the classical one, and consequently, it leads to quantum- speedup. We clarify that the underlying physics is quantum superposition principle in our analysis. To support and to corroborate the analysis, the numerical simulations are performed by considering two learning models. First, we consider a primitive model, random search, to give a clear understanding of the quantum-speedup enabled by extending the acceptable region. Then, we apply more practical learning model, called differential evolution, which is known as one of the most efficient learning methods. In this model, the quantum circuit exhibits faster learning speed and better convergence than the classical circuit.
View original:
http://arxiv.org/abs/1303.6055
No comments:
Post a Comment