Learning with incomplete information in the Committee Machine
Urs M Bergmann, Reimer Kuehn and Ion O Stamatescu
We study the problem of learning with incomplete information in a student-teacher setup for the committee machine. The learning algorithm combines unsupervised Hebbian learning of a series of associations with a delayed reinforcement step, in which the set of previously learnt associations is partly and indiscriminately unlearnt, to an extent that depends on the success rate of the student on these previously learnt associations. The relevant learning parameter $\lambda$ represents the strength of Hebbian learning. A coarse-grained analysis of the system yields a set of differential equations for overlaps of student and teacher weight vectors, whose solutions provide a complete description of the learning behavior. It reveals complicated dynamics showing that perfect generalization can be obtained if the learning parameter exceeds a threshold $\lambda_c$, and if the initial value of the overlap between student and teacher weights is non-zero. In case of convergence, the generalization error exhibits a power law decay as a function of the number of examples used in training, with an exponent that depends on the parameter $\lambda$. An investigation of the system flow in a subspace with broken permutation symmetry between hidden units reveals a bifurcation point $\lambda^*$ above which perfect generalization does not depend on initial conditions. Finally, we demonstrate that cases of a complexity mismatch between student and teacher are optimally resolved in the sense that an over-complex student can emulate a less complex teacher rule, while an under-complex student reaches a state which realizes the minimal generalization error compatible with the complexity mismatch.