Sacrificial Learning in Nonlinear Perceptrons

Using the cavity method we consider the learning of noisy teacher-generated examples by a nonlinear student perceptron. For insufficient examples and weak weight decay, the activation distribution of the training examples exhibits a gap for the more difficult examples. This illustrates that the outliers are sacrificed for the overall performance. Simulation shows that the picture of the smooth energy landscape cannot describe the gapped distributions well, implying that a rough energy landscape may complicate the learning process.