Electrically modifiable nonvolatile synapses for neural networks

The realization of an electronic element to simulate the synaptic interconnect in an electronic neural system is addressed. The basic interconnect or weight is an electrically reprogrammable, nonvolatile, analog conductance for the adaptive (modifiable) synapse. This weight is formed with a multidielectric, SONOS memory transistor which may be programmed at 5-V levels compatible with CMOS VLSI technology. The attractive features of this synaptic weight are its low power dissipation, small size, low voltage programmability, wide dynamic range, and ability to mimic biological synapses with respect to memory retention. Also addressed is the incorporation of the basic electronic synapse into a Widrow-Hoff delta-rule algorithm to study the electrical characteristics of this synthetic element in a single-level adaptive linear neuron. The combination of synaptic weights and a control algorithm provides a means to examine a basic functional neural building block for neural networks