A new continuous-time MOS implementation of feedback neural networks

An economical, simple, and versatile MOS cell that lends itself to the MOS implementation of feedback/feedforward neural networks is introduced. The new cell achieves vector scalar product of the form Sigma T/sub ij/V/sub j/, where y=1 . . ., n, i is fixed, V/sub j/ is the output of neuron i, and T/sub ij/ is the assigned positive or negative weight that is realized through voltage levels. The new circuit comprises only one operational amplifier (op-amp) and two input MOS depletion transistors. The vector scalar product of 2 n-tuple vector inputs is achieved by using 2(n+1) MOS transistors. This offers an economical alternative for VLSI analog neural networks. The analog neural network is realized by interconnecting double inverters through the new vector scalar product circuits. Moreover, the output voltage is tunable via programmable DC control voltages. The technique is demonstrated with an example of a two neuron circuit.<>