A PURELY CAPACITIVE SYNAPTIC MATRIX FOR FIXED-WEIGHT NEURAL NETWORKS

It is shown that the synaptic function of fixed-weight neural networks can be implemented with one capacitor only. The resulting synaptic matrix, being devoid of active devices, offers not only very high space-power efficiency and speed but also large synapse capacity with considerable analog depth. A purely capacitive matrix is very attractive also from the fabrication point of view because it gives an opportunity to exploit thin-film implementation styles as well as fully monolithic ones. The generic capacitor matrix is analyzed on the basis of dendritic charge conservation. The results are used to determine network limitations and to design a double-poly CMOS feedforward classifier, which is capable of correcting any 3-bit error occuring in a set of thirty 16-bit code-patterns. Each synapse occupies 16.5-mu-m x 10-mu-m field-oxide space for the very conservative 3-mu-m rules employed in this particular design. Electrical performance is verified through simulation. The work also includes a comparison between the proposed network and other switched-capacitor neural network configurations.