Paper detail

A Unifying Framework for the Identification of Motor Primitives

A long-standing hypothesis in neuroscience is that the central nervous system accomplishes complex motor behaviors through the combination of a small number of motor primitives. Many studies in the last couples of decades have identified motor primitives at the kinematic, kinetic, and electromyographic level, thus supporting modularity at different levels of organization in the motor system. However, these studies relied on heterogeneous definitions of motor primitives and on different algorithms for their identification. Standard unsupervised learning algorithms such as principal component analysis, independent component analysis, and non-negative matrix factorization, or more advanced techniques involving the estimation of temporal delays of the relevant mixture components have been applied. This plurality of algorithms has made difficult to compare and interpret results obtained across different studies. Moreover, how the different definitions of motor primitives relate to each other has never been examined systematically. Here we propose a comprehensive framework for the definition of different types of motor primitives and a single algorithm for their identification. By embedding smoothness priors and specific constraints in the underlying generative model, the algorithm can identify many different types of motor primitives. We assessed the identification performance of the algorithm both on simulated data sets, for which the properties of the primitives and of the corresponding combination parameters were known, and on experimental electromyographic and kinematic data sets, collected from human subjects accomplishing goal-oriented and rhythmic motor tasks. The identification accuracy of the new algorithm was typically equal or better than the accuracy of other unsupervised learning algorithms used previously for the identification of the same types of primitives.