標題: | 光學認知學習網路之研究 A Study of Optical PerceptronLearning Networks |
作者: | 鄭超仁 Chau-Jern Cheng 許根玉 Ken Yuh Hsu 光電工程學系 |
關鍵字: | 認知學習;光折變全像術;光學神經網路;perceptron learning;algorithm;photorefractive holography;optical neural network |
公開日期: | 1993 |
摘要: | 本論文主要在研究光學認知學習網路的原理、架構、實現方式以及其學習 特性。我們以光折變全像技術來演示認知學習模型並建造一種光學認知學 習網路以進行光學圖形識識別與分類。網路的神經連線係以全像光柵的形 式記錄、儲存在一個光折變晶體中,而晶體的動態特性則用用來模擬網路 的學習行為。此外,我們也結合一對光偵測器與電子相減器的干涉技術來 來建構一種雙極光學輸出神經元,以實現雙極光學認知學習網路。我們並 以數學證明方式來求解光折變全像式光學認知網路的收斂特性,而得到一 個曝光時間的參考指引。另一方 面,我們結合了電腦學習、電腦全像片 以及光學處理方式來 建構一種電腦與光學混合式認知學習圖形分類系統 ,而其中的電腦全像相關濾波片則是利用認知學習法則來設計。同時,我 們亦以實驗演示了一種即時電腦全像技術,而使得這個可學習的圖形分類 系統能進行線上處理。在認知網路的模型理論上,我們首先探討一種廣義 認知學習網路的收斂特性,這種網路涉及了連線權值更新時增益因素的影 響。由於增益 (或者衰減)因子可能影響認知學習過程之收斂,我們推導 出廣義學習法則的收斂條件。而光折變認知網路則是其中的一個重要實例 。我們也提出並分析一種具有值補償方式的修 正認知學習網路,以達到 接近於理想的認知法則。最後,我們探討一種嶄新的光學方法,它能使得 在光折變介質內衰減(或微弱)的全像光柵經由二波混合的能量耦合作用而 得到復原與增強。藉由結合一系列的讀寫時程與一個相位共軛器,可得到 一種穩態光折變全像光柵,這種穩態光折變全像能夠一直持續地維持其振 幅強度而不致於衰減。理論分析的結果顯示這種光學復原方法能夠使得復 原的光折變全像的重建影像對比度保持不變。 The thesis investigates the principles, architectures, implementations and learning characteristics of optical perceptron netetworks. The perceptron model is realized by using the technique of photorefractctive dynamic hologram and optical perceptron is used for optical pattern recognition and classification. The interconnection of the network is construructed by the amplitude of photorefractive hologram. The dynamic characteristics of the photorefractive crystal is used for simulating the learning behaviors of the optical perceptron. In addition, we used an interferometric technique which combines a pair photodetectors with an electronic subtractor for implementing the bipolar optical perceptron. The convergence properties of the optical perceptron are derived and a guideline for the exposure schedule of the photorefractive perceptron learning is presented. A novel technique for trainable pattern classification using a computer generated hologram is proposed and demonstrated. The perceptron learning algorithm is used for training, and the trained interconnection weight is transferred into a computer-generated hologram. The hologram is used as a correlation filter in the optical system for pattern classification. Furthermore, the computer generated holographic technique is used in the optical recognition system for real-time image processing. On the other hand,we consider the properties of a generalized perceptron learning network taking into account the decay or gain of the weight during the training stages. A mathematical proof is given which shows the conditional convergence of the learning algorithm. We also described a modified learning algorithm which provides a solution to the problem of weight decay in optical perceptron due to hologram erasure. |
URI: | http://140.113.39.130/cdrfb3/record/nctu/#NT820123002 http://hdl.handle.net/11536/57631 |
顯示於類別: | 畢業論文 |