大鄰近層細胞非線性網路之分析與設計

Abstract

在這篇論文中，一個可以在超大型積體電路實現大鄰近層細胞非線性網路﹝也稱為大鄰近層細胞神經網路﹞的新架構被提出來並且分析。在這個被提出來的新架構中，在一般CMOS製程虛擬BJT被用來替代寄生雙載子接面電晶體的使用，根據ν-BJT的物理元件特性，使用虛擬BJT的新神經元架構便被提出並分析，此外，少了寄生雙載子接面電晶體的缺點，使得神經元的設計變得容易，並且在功率消耗方面也較為節省。用來連結神經元的神經腱結被稱之為”路徑”，因為使用的電源供應為1V，再加上同時使用了NPN以及PNP的虛擬BJT，可以讓”路徑”在無輸入的情況下，不會有直流的功率消耗。 基於大鄰近層細胞非線性網路的觀念理論，大於一個鄰近層的細胞非線性網路可以被實現。因為對於大距離細胞層的連結並不需要直接連結，因此可以在超大型積體電路上實現。因為連結用的”路徑”是最大的面積消耗者，在提出的新架構中，可以縮減晶片的面積，也因此提高陣列的大小。另一方面，功率的消耗卻是跟陣列的大小成正比，這也是為什麼需要低功率的神經元以及神經腱結。藉由新架構設計的電路，已經成功地在HSPICE下模擬驗證數個細胞神經網路典型的功能，如物體的擴張與侵蝕、以及「Muller-Lyer」的箭頭錯覺等。 藉著新架構的使用，可以使得神經腱結的控制能力更加簡單，並且可以實現更多的功能。因此大鄰近層細胞非線性網路在實現多樣訊號處理的通用機器上有很大的潛力。

In this thesis, a new structure for the VLSI implementation of Large-Neighborhood cellular nonlinear network (LN-CNN), also call cellular neural network, is proposed and analyzed. In the proposed LN-CNN structure, the pseudo-BJT (P-BJT) is used to replace the use of the parasitic lateral bipolar junction transistor in the CMOS process. Based on the device physic properties of the neuron-BJT (υ-BJT), a new neuron structure with P-BJT is proposed and analyzed. Besides, without the disadvantage of the parasitic lateral BJT, it makes the design of neuron easier and economizes on the power consumption. The synapses to connect the neuron are called “path”. The power supply of these paths is 1V. Before any input current flow into the path, the low power supply makes these paths off. So there is no DC standby current and no DC power consumed by the paths. Basing on the concept of LN-CNN, the neighborhood of the template can be realized to reach more than one layer. Because the direct connection between neurons is not needed for far-distanced neurons, it makes the VLSI implementation of LN-CNN easier and smaller. The paths provide most of the chip area. But because of the proposed architecture, the chip area of the paths for interconnection is reduced and the array size can then be increased. The power consumption will increase as square times as the CNN array increases. So the power problem will be very serious. This is why a low power synaptic path and neuron is needed. Using the proposed LN-CNN structure, the LN-CNN functions, such as dilation, erosion and Muller lyer, have been successfully realized and verified in HSPICE simulation. Negative and asymmetrical template all can be realized. The pseudo lambda BJT neuron is proposed to implement a low power neuron. This neuron is also able to overcome the noise when the function of current mode Schmitt trigger is used. This current mode Schmitt trigger is evolved from the output function of the neuron. The output function is a hard-limited step function but is weak to discriminate the state when the state voltage is about zero with a random noise. By the new structure of LNCNN, it makes the template generation much easy and it can almost to generate most of the templates. So the proposed LN-CNN has great potential in the implementation of the CNN universal machine for various signal-processing applications.