Quantum Advantage of Neural Networks

2019-11-13

Recently, a team of physicists led by Renato Renner from the Swiss Federal Institute of Technology (ETH) in Zurich developed a new type of neural network. They conducted an experiment using an algorithm to detect the motion of the Sun and Mars relative to Earth, ultimately resulting in the heliocentric theory. In other words, the machine was able to learn the basic physics knowledge and related orbital information from the time of Copernicus and independently arrived at the conclusion of the heliocentric theory. It even derived the Mars trajectory formula proposed by Copernicus, making it a "Copernicus" in the field of artificial intelligence. History of Neural Networks Neural networks, also known as connection models, are algorithmic mathematical models that mimic the behavior of animal neural networks for distributed parallel information processing. By adjusting the relationships between a large number of interconnected nodes within the system, neural networks achieve the goal of processing information. In 1943, psychologist W. McCulloch and mathematician W. Pitts first proposed a mathematical model of neurons based on the analysis and summary of the basic characteristics of neurons. This model is still in use today and directly influences the progress of research in this field. In 1945, a design team led by John von Neumann successfully produced a stored-program electronic computer, marking the beginning of the era of electronic computers. In his research, von Neumann compared the fundamental differences between the structure of the human brain and stored-program computers, and proposed a network structure composed of simple neurons. In the late 1950s, the "perceptron" was invented, giving rise to multi-layer neural networks. In the early 1960s, non-linear multi-layer adaptive networks began to develop. In the early 1980s, the application of digital computers in several fields encountered difficulties, indicating that the time was ripe to seek solutions in artificial neural networks. As a result, a large number of scholars and researchers devoted themselves to the study of neural networks, which continues to this day. The researchers of the "Copernicus" team used a "slow" neural network to solve the problem of self-training in traditional neural networks. Their next step is to use the general rules and patterns derived from big data for the discovery of new physical laws and the redefinition of quantum mechanics. Advantages of Neural Networks in Quantum Science Redefining quantum mechanics may seem far-fetched, but it is not entirely impossible for the realization of quantum system simulation. As early as July of this year, researchers successfully simulated the "steady state" of a quantum system using a neural network-based algorithm. Using neural networks to estimate and simulate wave functions and density matrices can greatly reduce computational complexity and computational power requirements, laying the foundation for solving prominent problems in quantum science and information. Even in everyday life, nature is governed by the laws of quantum physics. These laws explain common phenomena in life, such as light, sound, heat, and even the motion trajectory of balls on a billiard table. These phenomena are familiar to most people and have become commonplace. However, when it comes to phenomena involving a large number of interacting particles, most of the phenomena explained by quantum physics contradict our intuition. To study quantum systems composed of a large number of particles, physicists must first be able to simulate such systems. Equations that explain the internal workings of quantum systems can be solved by supercomputers, but although Moore's Law predicts that computer processing power doubles every two years, it still falls far short of the computational power required to solve the challenges faced by quantum physics. Simulating the properties of open quantum systems is a prerequisite for solving prominent problems in quantum science and information. The difficulty lies in the exponential growth of the density matrix of the system as the system size increases. In open quantum systems, researchers aim to find the "steady state," which refers to a quantum state that does not change over time. When the system contains multiple quantum particles, difficulties may arise in computation. To describe the entire spin system, it is necessary to determine the 2^N possible states. Storing information about just 20 rotations requires approximately 8 gigabytes of RAM, and the computational power required doubles with each additional rotation. Dealing with the same number of rotations in an open system is even more challenging because rotations must be described using a "density matrix" ρ. This matrix is extremely large, with 2^N×2^N elements, requiring astonishing computational power to support it. The advantage of neural networks is that they can approximate wave functions or density matrices with very little information. It acts like a mathematical "box," taking a string of numbers (vectors or tensors) as input and outputting another string. For a specific task of simulating N quantum systems, the neural network function can serve as a "guess" for the wave function, taking the states of N objects as input. Then, researchers allow the network to "learn" from actual or simulated data or minimize the physical quantities determined by the wave function to optimize the function parameters. Once the correct guess is obtained, it can be used to calculate other physical properties, with far fewer parameters than 2^N.

RECENT AI NEWS

RECENT AI TOOLS