NEURAL NET USING DIFFERENTIAL CAPACITIVELY COUPLED INPUT AND OUTPUT LINES

Neural nets using capacitive structures are adapted for construction in complementary metal-oxide-semiconductor integrated-circuit technology. In each neural net layer synapse input signals are applied to the inverting and non-inverting input terminals of each of a plurality of differential-input non-linear amplifiers by a respective pair of capacitors, which non-linear amplifiers generate respective axon responses. Alternatively, fully differential input amplifiers in each neural net layer apply their push-pull responses to synapse input signals via a pair of input lines. Each push-pull response is applied via a pair of capacitors of complementary capacitance values and an output line to the input of each of a plurality of non-linear output amplifiers in that neural net layer, which generate respective axon responses for that neural net layer. In certain of these neural nets, arrangements are described that make the capacitive structures bilaterally responsive, so that back-propagation calculations can be performed to alter the relative values of capacitors in each pair thereof, which is done during training of certain of the neural nets described.