- 學位論文
由量子電路和混合神經元構成的深層神經網路
Deep Neural Network with Quantum Circuits and Hybrid Neurons
摘要
近年來深度學習被廣泛應用在各個領域，醫療、經濟、或氣候分析等等都有它的身影。而近年最為顯眼的發展就是深度學習模型在圍棋領域徹底擊敗人類。這事件讓人了解到深度學習具有的潛力，因此促使近年的深度學習模型的快速發展。而用量子電腦建構神經網路也是一個發展方向。但純量子的系統因為缺少 神經元不能建構深層神經網路。因此這篇論文的目的就是用Variational quantum circuit與hybrid neurons來建構深層人工神經網路。 我們引入了hybrid neurons的概念來補上量子神經網路的缺口。也讓新架構的神經網路同時具有量子和古典的優勢。量子的優勢是模型寬度隨qubits數量指數增長。古典的優勢則是模型複雜度隨網路深度指數增長。而此模型能兩者兼具。 在此文章中我們為了展示這種混和系統的威力，我們將用它來處理三類問題。第一類是回歸問題，第二類是類分類問題。最後一類則是強化學習問題。強化學習問題我們會以冰湖遊戲為例子。而從實驗的結果來看，加入hybrids neurons後量子電路能有更好的表現。這也證明hybrids neurons幫助提升模型的複雜度。在處理問題時我們大部分使用數值模擬作驗證。但是在處裡冰湖遊戲時我們有用真的量子電腦去驗證，並且達到很好的成效。這也證明此種模型在處理強化學習問題是能抵抗系統的噪音。此種特性對目前缺乏誤差修正的量子電腦而言頗具實用價值。
並列摘要
In recent years, deep learning has been widely used in various fields, such as medical, economic, climate analysis, etc. The most notable development in recent years is that deep learning models have completely defeated humans in the field of Go. This event made people understand the potential of deep learning, so it promoted the rapid development of deep learning models in recent years. The use of quantum computers to construct neural networks is also a development direction. But pure quantum systems cannot build deep neural networks because of the lack of neurons. Therefore, the purpose of this thesis is to construct deep artificial neural networks using variational quantum circuit (VQC) and hybrids neurons. We introduce the concept of hybrids neurons to fill the gaps in quantum neural networks. It also allows the neural network of this architecture to have both quantum and classical advantages. The advantage of quantum part is that the model width increases exponentially with the number of qubits. The classical advantage is that the complexity of the model increases exponentially with the depth of the network. Our hybrid can have both kinds of advantages. In order to show the power of this hybrid model, we will use it to deal with three types of problems. The first type is the regression problems, and the second type is the classification problems, the last type is the reinforcement learning problems. In the reinforcement learning problem, we will take the frozen-lake game as an example. From the experimental results, the quantum circuit can perform better after adding hybrid neurons. This also demonstrates that hybrid neurons help to increase the complexity of the model. When dealing with the problems in this thesis, we mostly use numerical simulations for verification. But in the game of frozen-lake, we also use real quantum computers to verify the model and achieve good results. This also demonstrates that this model can resist the noise of realistic quantum machines when dealing with reinforcement learning problems. This feature has practical value for quantum computers that currently lack error correction.
參考文獻
延伸閱讀
- Chiu, D. Y. (2020). 殘差網路啟發的混合量子神經網路之研究 [master's thesis, National Taiwan University]. Airiti Library. https://doi.org/10.6342/NTU202003705
- Chen, M. C. (2004). 量子類神經網路於求解滿足限制問題之研究 [master's thesis, Tatung University]. Airiti Library. https://www.airitilibrary.com/Article/Detail?DocID=U0081-0607200917233101
- 劉生武（2006）。模糊量子類神經網路於短期電力負載預測之研究〔碩士論文，中原大學〕。華藝線上圖書館。https://doi.org/10.6840/cycu200600637
- 劉彥成（2019）。應用於可調變傳輸之量子化資料壓縮類神經網路〔碩士論文，國立臺灣大學〕。華藝線上圖書館。https://doi.org/10.6342/NTU201901278
- 羅敏嘉（2013）。Development of Artificial Neuronal Network Devices〔碩士論文，國立清華大學〕。華藝線上圖書館。https://doi.org/10.6843/NTHU.2013.00292