Capacitive synapses for microelectronic neural networks

It is shown that the synaptic function of nonadaptive neural networks can be implemented with a single capacitor. The resulting synaptic matrix, being devoid of active devices, offers not only very high space efficiency, but also negligible DC power, perfectly linear dendritic functionality, negligible process noise, and simple and fast sensing. The generic capacitor matrix is analyzed on the bases of dendritic charge conservation, and the results are used to design a double-poly CMOS feedforward classifier which is capable of correcting any 3-bit error occurring in a set of 30 16-bit code-patterns. Each synapse occupies a 16.5 mu m*10 mu m field-oxide space for the very conservative 3 mu m rules used in this particular design. Electrical performance is verified through simulation.