In this paper, the structure and CMOS analog circuits for implementation of multilayer perceptron with on-chip learning are presented. Two learning schemes, error back-propagation and hebbian learning, are possible. The learning rate can be controlled externally by digital signal. Also, the sigmoid and the derivative of sigmoid have adjustable slope. The number of synapse and neuron can be expanded by simple connection. Several configurations of multilayer perceptron are simulated, and these SPICE simulation results are shown.