Abstract: The hidden unit performance related spaces, i e representation space and error space, target space and expended space, are analyzed according to the model of least squares approximation feedforward neural networks, and further more, the criteria and the evaluation methods for the performance of hidden units are proposed It is revealed that the efficient component of output vector of a hidden unit should be lying in error space, avoiding expended space and closing the direction of maximum energy, which are independent of the nonlinear functions used by the units, and it is permitted that different units have different activation functions The quality factor of the hidden layer, efficient coefficient of the hidden layer, the redundancy of hidden units and the evaluation factor of the hidden layer for total evaluation, are proposed for the performance evaluation of the hidden layer