- 學位論文
深度神經層在手寫圖形辨識之硬體化評估
Evaluation on Deep Neural Layers in Hareware Implementation for Handwritten Pattern Recognition
摘要
本論文目的是探討手寫圖形辨識應用於非重疊離子佈植(Non-Overlapped Implantation, NOI)非揮發性記憶元件所組成的深度神經網路之可行性。圖形為32x32像素的數字0~9。首先會在Matlab建立五項純軟體的類神經架構，包含: (1)無負權重感知機(2)正負權重感知機(3)單層隱藏層倒傳遞(4)雙層隱藏層倒傳遞(5)三層隱藏層倒傳遞。接著會將NOI硬體電性模型代入上述五項模型。由於NOI突觸有權重範圍、輸入類型與無負電流的限制，因此在本論文中會藉由調整訓練過程、校正轉換函數、結合差動電路產生負權值等方式解決上述限制。 為了比較軟體與NOI類神經網路的辨識率，本論文根據MNIST圖像資料庫所提供的0~9，共10個數字對五項不同的神經網路架構進行圖像訓練，並透過1000個辨識樣本進行辨識率測試。 NOI硬體模型參數的類神經架構其辨識率在無負權重感知機、正負權證感知機、單層隱藏層、雙層隱藏層、三層隱藏層辨識率的辨識率分別為74.5%、78.2%、80.8%、80.7%、79.6%；純軟體的辨識率為67.5%、77.1%、76.1%、71.6%、65.1%。根據辨識結果得知，受限制的NOI硬體類神經網路不管在任何架構下比較，其辨識能力皆高於純軟類神經系統。
並列摘要
This paper is to investigate the feasibility of the Deep Learning Neural Network circuit using Non-overlapped implantation (NOI) for the hand-written pattern recognition. Each pattern has are 32x32 pixels with digits from 0to9. Initially, five standard neural architectures are established using Matlab that including: (1) No negative weights Perceptron (2) Standard Perceptron (3) Single hidden layer back-propagation (4) Double hidden layers back-propagation (5) Three hidden layers back-propagation. In addition, the measured NOI device data is embedded in the software to simulate these five neural networks. The NOI synaptic devices have three limits on the weight range, input type and no negative current. To overcome the limits, we try to modify the training flow, correct the transfer function, add the differential circuit to introduce negative weights. In order to make comparison between recognition rates of the software and NOI neural network, the numbers from 0 to 9 which are trained based on the MNIST image database for testing five standard neural architectures. Additional 1000 samples are therefore used to obtain the recognition rates. The recognition rates for five architecture with the NOI model are 74.5%、78.2%、80.8%、80.7%、79.6%, respectively.