非揮發性記憶元件應用於類神經感知機網路

Non-Volatile memory device apply in neural network

蔣岳霖(Yueh-Lin Chiang)

指導教授 ： 鄭湘原

中原大學/工學院/電子工程研究所/碩士(2014年)

https://doi.org/10.6840/cycu201400925

摘要

類神經網路(Artificial Neural Network)主要是探討如何模擬應用人類大腦運作的方式，目前已廣泛的應用音頻、圖形辨識中；隨著科技日益發達，具備可攜性的硬體已經相當普遍。本論文透過 0.25um代工製程建立4x3 之NOI 非揮發性記憶元件陣列來實現類神經積體電路，利用感知機的演算法並配合IC 測試儀來進行訓練與驗證。
在本論文中，我們使用了六種圖形透過感知機網路進行學習，藉由監督式學習與目標值比較，不斷將權重進行更新直到收斂，另外將元件與電路的測試結果建立經驗模型，並嵌入至軟體中進行模擬驗證。在模擬中針對輸入、判斷電壓以及寫入抹除的時間進行比較，找出最適合此系統的參數，並討論不能收斂的原因，最後再藉由硬體訓練驗證模擬之結果，其結果顯示訓練趨勢與硬體相仿。

關鍵字

類神經網路 ； 非揮發性記憶體 ； 感知器演算法 ； 經驗模型

並列摘要

The artificial neural network simulates the operation of the human brain. It is applied extensively in audio processing and pattern recognition. With the advancement of technology, hardware portability is becoming indispensable nowadays. This thesis uses the 0.25um CMOS foundry technology to implement the 4x3 NOI array neural network and perceptron algorithms with an IC tester to verify and train the circuit.
In this thesis, six input patterns were used for the learning algorithm in these NOI synapses. During the training process, the output signals were supervised and compared to the target by updating NOI synapse weights until the system converges. Initially, we measured the circuit and device data to establish their empirical models and embedded them in the software to simulate the neural network. In the simulation, we discussed the input, judgment and stress time, found the best parameter for the system and discussed the reasons why some results fail to converge. Finally, we verified the simulation result through hardware training, and the results show that the simulation training trend is similar to hardware training.

並列關鍵字

Perceptron ； Artificial Neural Network ； Non-Volatile Memory

參考文獻

[1-5]翁健豪,“感知器學習演算之硬體電路實現”, 碩士論文, 中原大學電子工程

[2-1]翁宇廷, “熱電洞注入對非重疊佈植記憶元件之可靠度研究”, 碩士論文, 中

[2-2]陳昭甫, “單邊非重疊離子佈植記憶元件介面特性與資料保存能力提升之研

Floating-Gate MOS Learning Array with Locally Computed Weight

for a Forward Path Neuron Circuitry of a Back Propagation Neural Network",