THE MLP-BGD-2 METHOD Regression and classification with multilayer perceptron networks trained using batch gradient descent with adaptive, with dynamic input-dependent learning rates Radford M. Neal, 23 July 1997 This method is the same as mlp-bgd-1, except that the learning rates (stepsizes) for weights on connections out of input units are adjusted dynamically in an attempt to make weights from the more relevant inputs learn faster, thereby improving the performance of early stopping. In detail, the adjustment works as follows. The magnitude of the gradient vector restricted to the weights on connections out of each input unit is computed. The fourth power of this magnitude is then found, and these fourth powers are divided by the largest of them, producing a scaling factor between zero and one for each input. The stepsizes that would normally be used for weights out of an input are multiplied by the corresponding scaling factor. Other weights and biases have the same stepsizes as in mlp-bgd-1. All other aspects of the procedure are also the same as in mlp-bgd-1. The shell files used to implement the method work the same way as for mlp-bgd-1.