Integrated synapse devices for fluxon neural networks

提出了一种新型的神经元超导电路和两种基于SQUID的可变突触。约瑟夫森电路由于在极低功耗下的高速运行，似乎在神经网络的VLSI中具有优势。采用单结SQUID和双结SQUID相结合的方法实现了一种具有良好输入输出隔离和高阈值特性的神经元电路，单结SQUID的量子态代表神经元的状态，双结SQUID的输出电压为Sigmoid函数。其中一个可变突触电路以数字方式改变其电导值。另一种可变突触电路是可变电流源，其中的输出电流可以数字方式改变。两个突触回路都由多个分流的双结SQUID组成。除了电路特性的数值模拟外，我们还利用Nb/AIOx/Nb约瑟夫森结技术制造了超导神经芯片。成功地演示了各单元和3位A/D转换器的基本工作原理。该A/D转换器工作在模拟输入频率高达100 kHz的情况下，受我们高增益测量设备的限制。仿真结果表明，该网络对100 MHz以上的模拟输入具有良好的响应能力。在如此高的速度下测试网络是我们未来面临的挑战之一。本文还讨论了一个基于Hebb规则的可变电流源型突触学习系统。

Novel superconducting circuits for a neuron and two types of variable synapses, which are based on SQUIDs, are presented. Josephson circuits seem to be superior for VLSI of the neural networks because of the high-speed operation under very low power dissipation. A neuron circuit with good input-output isolation and steep threshold characteristics is accomplished using a combination of a single-junction SQUID coupled to a double-junction SQUID.The quantum state of the single-junction SQUID represents the neuron state, and output voltage of the double-junction SQUID,which is operated under a nonlatching mode with shunt resistors, is a sigmoid-shaped function. One of the variable synapse circuits changes its conductance value digitally. Another variable synapse circuit is a variable current source in which the output current can change digitally. Both synapse circuits consist of multiple shunted double-junction SQUIDs. Besides numerical simulations of the circuit characteristics, we have fabricated superconducting neural chips using a Nb/AIOx/Nb Josephson junction technology. The fundamental operation of each element and 3-bit A/D converter are successfully demonstrated. This A/D converter was operated with analog input frequency as high as 100kHz, limited by our high-gain measurement equipment. Simulation shows that this network responds to analog input at over 100MHz. Testing the network at such high speeds is one of our challenges in the future. A learning system based on Hebb's rule with variable-current-source type of synapse is also discussed.