Research Fabrication of Integrated Neural Computing Circuits

作为集成神经网络研究的第一步，制作了一个AD转换器。通过构造集成神经网络的AD转换器，除了突触的可塑性外，还可以研究构成一般集成神经网络的各种基本问题。设计的电路是一个由4个神经元和12个突触组成的4位系统。一个神经元包括2个巴黎的cmos晶体管，一个突触包括一对cmos和一个nMOS。突触权重的值由门的几何形状决定，L/W。电路有外部控制线来修改特性，以防过程产生一些意外的偏差。考虑到工艺中可能存在的各种困难，采用4微米规则设计了图形尺寸。所构建的神经元表现出与预期一致的S形曲线的阈值特性。突触的输出与输入成正比，但偏移量很小。可以通过改变沟道掺杂的条件来消除偏置。Hopfield提出的具有对称突触的AD转换器被发现由于解陷入局部极小而导致错误的结果。一种带有去对称化突触的新AD转换器被发现总是正确运行的。一种新的AD转换器的制造正在进行中。

An AD-converter was fabricated as the first step of research for integrated neuron networks. By constructing an AD-converter of integrated neural network one can study various fundamental problems underlying for the fabrication of general integrated neural networks except for the plasticity of the synapses. The designed circuit was a 4-bit system composed of 4 neurons and 12 synapses. A neuron includes 2 paris of CMOS transistors, and a synapse includes a pair of CMOS and a nMOS. The value of a synapse weight is determined by the geometry of the gate, L/W. The circuit has outside control lines to modify the characteristics in case the processes create some unexpected deviation. The pattern size was designed by 4 micron rule considering the various possible difficulties in the processes. The fabricated neuron showed a threshold characteristics with a sigmoidal curve as expected. The output of the synapse was proportional to the input with a small off-set. The off-set would be removed by changing the condition of the channel doping. The AD-converter with symmetric synapses proposed by Hopfield was found to operate with wrong results as the solution was trapped by a local minimum. A new AD-converter with desymmetrized synapses was found to operate always correctly. Fabrication of a new AD-converter is under way.