Complex behavior is thought to be generated by a small number of movement primitives. Multiple definitions of motor primitives have been given in the literature, each one translating into different generative models and different techniques for their identifications. While all these approaches differ from each other in the underlying generative model assumptions and specific parameter priors, for each of them the number of primitives used to approximate the original data has to be decided a priori.

Model comparsion on human gait data. Analysis of emotional gait data recorded in our lab with PCA, ICA and Anechoic demixing [Omlor2007b] for different numbers of sources. The bars represent model evidences computed with Laplace approximation, relative to the lowest observed model evidence (PCA, 1 source). The anechoic analyses were carried out either without temporal smoothing (black) or with the optimal f0=7Hz (blue) of a wave smoothing kernel (see [EndresModelSel2013]). Error bars are standard errors, computed across trials. The best model (highest evidence) is the anechoic mixture with three sources and f0=7Hz, followed by the SIM (smooth instantaneous mixture) model with f0=7Hz. PCA and ICA are significantly worse for any number of sources.