Implementation and Analysis of Pulse Propagating Networks

最近的生理学和数学研究表明，大脑，特别是大脑皮层的复杂信息处理，是由神经元对非同步传入的兴奋性动作电位的精细时间结构做出敏感反应而实现的。神经元在时空输入脉冲序列中起到重合检测器的作用。在具有重合检测神经元的时空网络中，连续变量和时间是重要的，因为它们可以进行实数处理。实数运算能力很强，但一般的数字计算机无法实现，本研究采用模拟集成电路技术实现时空处理网络，该网络具有连续的时间和电压、电流等变量。首先，提出了异步脉冲传播网络的核心单元--异步脉冲神经元模型。研究了单个神经元和由多个神经元组成的神经网络的响应特性。结果表明，网络中存在单个神经元的混沌响应，并且网络中存在动态组装。动态拼接是时空处理中信息编码的一种候选方法。其次，提出了模型神经元、突触和轴突的模拟电路。轴突电路中传播脉冲的延迟时间可以连续控制。此外，突触回路的重量可以通过数字方式改变。最后，采用1.2微米的半导体工艺制作了所提出的电路。测量了电路的特性。因此，芯片确认了混沌响应。此外，还观察到轴突电路中的脉冲延迟。通过集成电路的实现，说明了利用异步脉冲传播网络进行实数处理的可能性

It is suggested, from the recent physiological and mathematical studies, that the sophisticated information processing in the brain, in particular in the cortex, is achieved by the neuron that sensitively responses to the fine time structure of asynchronously incoming excitatoly action potentials. The neuron functions as a coincident detector among spatio-temporal input pule trains. Continuous variables and time are important in the spatio-temporal network with coincident detector neurons, because real-number processing is possible with them. The real-number processing is very powerful but cannot be realized by the ordinary digital computer.In this research, the spatio-temporal processing network is implemented using analog integrated circuit technology where continuous time and variables such as voltage and current are available. First of all, an asynchronous pulse neuron model that is the core element of the asynchronous pulse propagating network is proposed. The response characteristics of the single neuron and neural network composed of numbers of neurons are investigated. As the results, chaotic responses are observed from single neuron, furthermore, dynamical assembly is organized in the network. The dynamical assembly is one candidate of the information coding in the spatio-temporal processing. Secondly, analog circuits for the model neuron, the synapse and the axon are proposed. The delay time of the propagating pulses in the axon circuits can be controlled continuously. Moreover, weight of the synaptic circuit can be altered digitally. Finally, the proposed circuits are fabricated using 1.2 mum CMOS semiconductor technology. Characteristics of the circuits are measured. As a consequence, chaotic responses are confirmed from the chip. Furthermore, pulse delay in the axon circuit is also observed. The possibility of the real-number processing using asynchronous pulse propagating network is shown through the IC implementation