以類比特徵輸入於類神經單層感知機之評估研究

Evaluation of the analogical feature input in hardware perceptrons

江昆諺(Kun-Yan Jiang)

指導教授 ： 鄭湘原

中原大學/電機資訊學院/電子工程研究所/碩士(2015年)

https://doi.org/10.6840/cycu201500838

摘要

類神經網路(Artificial Neural Network)已廣泛的應用在圖形辨識(Pattern Recognition)技術之中。其應用範圍多為資訊安全、身份認證、醫學影像處理、智慧型電子產品等方面。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比式圖形學習。 不同於先前所發表的數位式圖形特徵，本論文中的類比式圖形特徵為一多值輸入特徵形式。為了使系統更穩定，更快速的學習，本論文採用不同輸入特徵並根據Matlab軟體模擬結果分析其收斂趨勢。另外，本論文也藉由軟體模擬全晶片抹除機制(Chip Erase)與單條神經元抹除(Single Neuron Erase)機制探討不同的權重更新方法對於系統的影響。除此之外，本論文也針對判斷電壓、圖形特徵電壓範圍、圖形覆蓋率進行討論，找出最適合系統的參數。最後再藉由硬體訓練驗證模擬之結果。 由硬體訓練結果可知，使用類比圖形特徵能使系統的學習效率增加25倍以上，意味著訓練時間與系統總體耗能減少了25倍。雜訊測試方面，類比特徵所訓練的系統辨識率皆可增加約15.85%。證明使用類比特徵作為系統訓練時的輸入不但能增加系統收斂速度，也可以增加類神經系統的辨識能力。

關鍵字

類神經網路 ； 類比式圖形特徵 ； 非揮發性記憶體

並列摘要

Artificial neural network has been applied extensively in audio processing and pattern recognition, such as information security, authentication, medical image processing, intelligent electronic products, etc. We implement the 4x3 NOI neural arrays by using 0.25um CMOS process, with the perceptron algorithms for analogical pattern recognition.
Analogical features were used as multi-value inputs which were different from digitized feature inputs. In order to study the learning performance of hardware system, the MATLAB software is applied for simulating the training trend with different input pattern features. Moreover, we also simulate two kind of different weight updatings by software to evaluate system’s efficiency. These methods include whole Chip Erase and Single Neuron Erase. Besides, the optimization parameter of the system judgment, input voltage range and coverage of pattern are also discussed. Finally, the simulation results through hardware training are also verified.
From the results in our hardware training, we found that using the analogical feature can speed up the system learning, and the learning efficiency increase 25 times. Furthermore, in the noise testing, the recognition rate of system also improves 15.85%. In summary, it has been found that analogical features improve the system not only the training speed but also the recognition rate.

並列關鍵字

Artificial Neural Network ； Analogical pattern feature ； Non-Volatile Memory

參考文獻

【3】 古佳琳, “陣列式表面聲波感測器在有機氣體辨識之研究”, 碩士論文, 中原大學電子工程所, 2007.

【7】 Fopefolu Folowosele, Tara Julia Hamilton, Ralph Etienne-Cummings, “Silicon Modeling of the Mihala¸s–Niebur Neuron”, IEEE, 2011.

【11】 Chan, T. Y., K. K. Young, and Chenming Hu. "A true single-transistor oxide-nitride-oxide EEPROM device." Electron Device Letters, IEEE 8.3 (1987): 93-95.

【12】 Eitan, Boaz, et al. "Can NROM, a 2Bit, Trapping Storage NVM Cell, Give a Real Challenge to Floating Gate Cells?." Extended abstracts of the... Conference on Solid State Devices and Materials. Vol. 1999. 1999.

【15】 Eitan, Boaz. "Two bit non-volatile electrically erasable and programmable semiconductor memory cell utilizing asymmetrical charge trapping." U.S. Patent No. 6,011,725. 4 Jan. 2000.