Fig. 10
(From [51]). Growth of a GP module, based on the coincidence between an inner-flow and incoming stimuli. Alternative encodings of a given stimulus by a downstream module are shown in the two grey boxes at the bottom; the choice between two patterns of growth (dotted arrows) depends on the level of activity of coincident cells, itself controlled by the current weight w5ht of the inner-flow. Here, the filling level of a cell stands for its activity, whereas a dashed-line features its response threshold. Thanks to its tree-like structure that results from several encoding steps, a given GP module can anticipate the next stimuli to occur. If a stimulus occurs which was not expected in the ongoing context, the learning algorithm is triggered. The more boosted the inner-flow then, the more Generalization takes the lead over Differentiation