The paper deals with the approximation of continuous functions by feedforward neural networks. In the first part of paper are presented some main results of Ito and Cardaliaguet-Euvrard regarding universal approximators implementable as four-layer neural networks. In the second part is presented an explicit formula similar to Cybenko expression for approximating a continuous multivariate function using sharacteristic function as a particular bell-shaped function in place of sigmoidal function. This approximation formula is implementable as three-layer feedforward neural networks that, surprisingly, have in the hidden layer the same number of neurons as to Ito and Cardaliaguet-Euvrard four-layer neural networks have in the second hidden layer.