Study of 1-bit data stream digital circuit of neural network for higher order chemical sensing

各种类型的用于神经网络的大规模集成电路目前已被许多机构研究。提出了一种基于1位数据流信号处理电路的数字神经网络电路结构。在此基础上，实现了大规模集成电路，并将其应用于化学传感器的信息处理。利用石英谐振腔气体传感器的输出模式，实现了气味的学习和识别电路。采用24000个晶体管和测量电路实现了LVQ的神经网络。我们通过离线测量获得的这些模式数据证实了电路的正确行为。设计了包含自检电路、计算机接口在内的116000个晶体管的电路。由于存储器电路的完全定制设计是执行的，因此布局方法是构建块方法。该芯片将在不久的将来进行评估。另一方面，利用FPGA实现的电路，对便携式气味传感系统的简易测量系统进行了碳数相同但官能团不同的气味识别实验。结果发现，即使在进行在线识别时，气味也能被正确识别。采用1位数据流信号处理电路，利用传感器阵列信号进行气流方向估计。当离线数据输入到FPGA实现的电路时，我们证实了该电路解决了基于光流约束的线性方程。

Various types of LSI for neural network have been so far studied by many institutions. We proposed a new architecture of digital neural network circuit based upon 1-bit data stream signal processing circuit. Furthermore, we aimed to realize LSI using that method and to apply it to information processing of chemical sensors. The circuit for learning and recognizing an odor using the output pattern of quartz-resonator gas sensors was implemented into LSI fabricated utilizing the chip fabrication service. The neural network of LVQ was realized using 24000 transistors including measurement circuits. We confirmed that the circuit correctly behaved using the data of those patterns obtained by the off-line measurements. Furthermore, the circuit with 116,000 transistors including self-test circuit, computer interface was designed. Since the full custom design of the memory circuit was performed, the layout method was building-block one. The chip will be evaluated in the near future. On the other hand, the circuit implemented into FPGA was used to do the experiment on the identification of the odors with the same number of carbons but with the different functional groups using the simple measurement system for a portable odor sensing system. It was found that the odors were correctly identified even when the on-line identification was performed. Furthermore, 1-bit data stream signal processing circuit was applied to the direction estimation of the gas flow using the sensor-array signals. We confirmed that the circuit solved the linear equations based upon the optical-flow constraints when the off-line data was input to the circuit implemented into FPGA.