机器学习在量子信息与其他物理中的应用

MACHINE LEARNING IN QUANTUM INFORMATION AND OTHER PHYSIC

崔子嵬

11849207

硕士

物理学

翁文康

2020-05-27

2020-07-05

哈尔滨工业大学

深圳

机器学习是利用概率论、统计学等知识，结合最新的计算机技术，通过大量数据或其他经验自动改进计算机算法的研究，在图像分类、时序预测、自我调控等方面取得了诸多进展。量子机器学习是机器学习和量子信息领域的结合学科，包括将使用量子计算加速机器学习方法，或者使用传统方法解决量子物理领域的难题两方面。本文对后一类方法进行了探索，并探索了机器学习在其他物理问题中的应用。 快速高效的传递量子信息对于实现可扩展的量子计算至关重要，其中一维自旋链上的量子态传输是量子信息领域中的热点问题。经典方法研究中有诸多方法，但皆有些许不足。增强学习是机器学习中的一种，可以在无需人类标注数据的情况下自主学习，寻找针对特定问题的调控方案。本文通过引入增强学习和另一种经典机器学习手段，提出了无需含时调控和包含含时调控的两种量子态学习方案，并且在速度和准确率上均超过了原有结果。 量子纠缠是量子信息与量子计算中重要的资源。然而即使在小型系统，例如2-3量子比特的情况下，将纠缠态与可分态进行区分依然是非常困难、需要消耗大量资源的过程。本文通过引入人工搭建的神经网络，在两比特情况下证明了机器学习的分类能力。卷积神经网络作为神经网络的一种，可以有效的提取并组合局部内容以获得更有效的信息。在本文中，一种修改了的卷积神经网络将可以把这种提取映射为量子信息中的投影测量过程，并且可以在通过数据的训练过程中获得能够解决某些特定问题的合适测量基。通过这种方法，我们不仅可以获得分辨量子纠缠的能力，同时可以在其他多种问题中取得结果，例如计算纠缠熵，或者在引入自编码器后实现量子态层析过程等。 Rayleigh-Benard对流实验在流体力学中具有重要的意义，在进行实验时，标注当前流体状态是研究其他内容的基础，然而极其费时费力。通过引入机器学习手段，可以将这一过程缩短至数秒乃至更短，并且适用于多种其他对流实验系统。

Based on probability, statistics and newest computer science, machine learning is a new subject which could improve the computer algorithm automatically by huge data or other experiences. It has already achieved several progress in many subjects, like image classification, time series prediction or self-control. Quantum machine learning is the integrated discipline combining quantum mechanics and machine learning, including two directions: using quantum computing to accelerate machine learning algorithm, or using classical machine learning to help solve the difficult problems lie on quantum physics. In this thesis, we focus on the second direction, and make one more step to explore the application of machine learning on other physics problems. Quantum entanglement is the central resources in quantum information and quantum computing. However, even in small systems contain two or three qubits, the task of classifying entangled states from separable states is still difficult and resources-consuming. By introducing artificial neural network, we proved the classification power of machine learning in this problem. Convolution neural network can extract local information and combine them into high-order information. Here a modified convolution layer is proposed to map the convolution progress into the projection measurement in quantum field, which could find the suitable measurement basis for certain problems. By this method, we can not only classify the quantum state, but also apply into other problems, like prediction entanglement entropy, or realize quantum state tomography by auto encoder structure. To transfer quantum information quickly and efficiently is a key to achieve scalable quantum computing, and quantum state transfer on one-dimension spin chain is one of the hot problems in quantum information fields. There are several traditional results, but each of them has their disadvantages. Reinforcement learning, one of the machine learning method, can automatically learning control protocols on certain problem even without human labelled data. This thesis introduced the reinforcement learning method and another classical machine learning method to propose two quantum transfer protocols, time-independent version and time-dependent version, and each of them has improved the result of original protocol. Rayleigh-Benard convection experiment plays an important part in fluid mechanics. The convection state label is the basis of researching other problems, which cost a lot of time and attention. The machine learning method can reduce the time into only a few seconds or even less, and has the ability of generation into other systems.

机器学习 神经网络 量子信息 量子纠缠 流体力学

machine learning artificial neural network quantum information quantum entanglement fluid mechanics

中文

联合培养

南方科技大学

崔子嵬. 机器学习在量子信息与其他物理中的应用[D]. 深圳. 哈尔滨工业大学,2020.