- 學位論文
CMOS硬體感知機之辨識率與權重持續性的研究
Study of the recognition rate and weight retention in CMOS hardware perceptrons
摘要
類神經網路(Artificial Neural Network)主要是模擬生物大腦運作與訊號處理的方式，隨著科技與生活水平的提升，已廣泛的運用於各行各業，如氣體辨識、影像辨識與語音辨識…等。本論文透過tSMC 0.25um製程建立4*3之NOI非揮發性記憶元件陣列來實現類神經積體電路。此晶片透過測試機台來實現硬體類神經網路應用。系統根據感知機演算法的運算控制著NOI非揮發性記憶元件陣列的操作，藉由寫入與抹除機制改變權重，使系統能夠順利進行學習與辨識。 本論文提出一種創新的學習強化方式以提升現有類神經晶片的辨識率與資料保存能力。此訓練方式與傳統感知機演算法不同的地方在於轉換函數部份。學習強化方式在傳統定義的系統收斂後，讓類神經網路再次進行訓練，將判斷準位上下拉開形成一個窗口，使系統讀取時形成一塊容錯區域。透過隨機雜訊摻雜的圖形辨識，經過學習強化訓練後的樣品可大幅提升硬體類神經系統的辨識率從原本的78.81%增加至94.44%。另外，在150。C高溫烘烤下，經過學習強化訓練後的樣品其使用年限可從原本的200秒增加至1萬秒。最後，本論文將硬體的訓練數據帶入軟體模擬，比較兩種訓練方式於烘烤期間辨識率的變化。
並列摘要
The artificial neural network (ANN) simulates the operation of the human brain which dominates the signal handing. With the advancement of technology and living standard, ANN has been widely used in various industries. Hardware portability is also becoming indispensable nowadays. In this thesis we use the 4*3 NOI synapse array silicon to implement neural network circuit. And supervised learning algorithm is accomplished by tester machine for neural network application. According to the selected algorithm, the tester controls the NOI synapse array operations. Weight updating is done by programming and erasing mechanism and also determines the performances of neural network system. In order to enhancing the recognition rate and data retention of the existing neural chip, we propose a novel learning method which is called enhancement learning. Different from the traditional perceptron algorithm, after the convergence, the initial judgment level of the perceptron algorithms will be enhanced to a marginal level, which is equivalent to a tolerant region during recognition. In comparison of the conventional for perceptron learning, the enhancement learning can achieve better data retention and recognition. The pattern recognition rate is evaluated by random shading. The enhancement learning can significantly improve the hardware neural network system whose recognition rate increases from 78.81% to 94.44%. In addition, its useful life can be increased from 200 seconds to 10,000 seconds at 150 ℃ high-temperature baking. Finally, the training data of hardware experiment has also been compared with our software simulation for its recognition rate.