Statistical Mechanics of Nonlinear On-line Learning for Ensemble Teachers

摘要: We analyze the generalization performance of a student in model composed nonlinear perceptrons: true teacher, ensemble teachers, and student. calculate error analytically or numerically using statistical mechanics framework on-line learning. treat two well-known learning rules: Hebbian perceptron As result, it is proven that shows qualitatively different behaviors from linear model. Moreover, clarified show each other. In learning, we can obtain solutions. this case, monotonically decreases. The steady value independent rate. larger number teachers more variety have, smaller is. have to dynamical are nonmonotonic. rate is, is; minimum