Building Generative Models: Structural Learning and Identification of the Learner
Reza Shadmehr, Sandro Mussa-Ivaldi · The MIT Press eBooks · 2012
This chapter presents a discussion on structural learning and identification of the structure of the learner. It reveals that the prior exposure to a rotation perturbation, despite being random and unlearnable, seemed to significantly enhance learning rates for a member of the same perturbation class. It shows that the problem of structural learning is that of describing a dynamical system that in principle can accurately predict the sensory consequences of motor commands, that is, learn the structure of a forward model. This chapter suggests that Expectation Maximization is an alternate approach to estimating the structure of a linear dynamical system.