- 學位論文
相干光學神經網路設計與演算法之研究
Studies on Optical Correlator-based Neural Networks with Different Training Algorithms
摘要
人工神經網路的架構與集體式計算之概念，與現今電腦的資訊儲存與計算之架構完全不同，以電腦硬體與程式語言來運行仿生的計算，往往是耗時且耗能的工作。因此，找尋更佳的硬體製作技術，乃是人工神經網路進一步發展之關鍵研究主題。 針對目前深度機器學習以層狀交替來模擬神經網路之運作的架構，完全符合光學系統輸入與輸出平面映射的想法，再加上微電子與光電科技之硬體元件的發展，使得運用全光學物理機制來建立「光學神經網路」架構的研究變得十分熱門，也頗有進展。因為，相較於以電腦實現類神經網路之硬體架構，光學神經網路有以下的優勢：1.高速資訊處理、2.巨量的互連性、3.高密度的空間資訊、4.平行處理能力、5.低耗能。 本論文基於這樣的趨勢，運用光學信號中最常見的光學相干器，實現多類別圖形分類光學神經網路的硬體架構，並探討不同演算法對這樣的光學神經網路運作與功能的影響，期望提升光學神經網路的效能和影像辨識的能力，並希望藉由這樣概念證明之先導性研究，未來能替機器學習的應用開創更獨特的機會。
並列摘要
Neural network structure is totally different from computer architecture. The neural network computed by computer hardware is a timely and energy-consumed work. Thus, to find a better hardware technique is the crucial problem for the development of artificial neural network. Based on the similar structure and the progress of photoelectronic technique, research of optical neural network has become more and more popular nowadays. Compared to the computer neural network, optical neural network has 5 advantages: 1. High speed data processing, 2. Large number of interconnections, 3. Huge data storage, 4. Parallel computing, 5. Low energy consumption. We proposed an Optical Correlator-based Neural Network to implement hardware for image classification neural network, and discuss the capability of the optical neural networks under different algorithms. Hope this concept-proof research can provide new opportunity for machine learning in the future.