Multilayer in-place learning networks: Multitask invariance and adaptive lateral connections

In the fields of neuroscience, psychology, computer science, and developmental robotics, currently there is a lack of biologically plausible general-purpose in-place learning models that incrementally learn multiple sensorimotor tasks, to develop "soft" multi-task-shared invariances in the internal representations while the human or robot interacts with its environment. The multilayer in-place learning network (MILN) (Weng and Luciw, 2006; Weng et al., 2007) is a developmental network aiming at this ambitious goal. This biologically inspired developmental model for sensorimotor pathways provides an unusually efficient learning algorithm whose simplicity, low computational complexity, and generality are set apart from typical conventional learning algorithms. It explains how a biological cortical layer uses three types of adaptive connections, bottom-up, lateral, and top-down to accomplish this very challenging goal through the miraculous developmental experience. The work presented here concentrates on multitask invariance and recent work about the adaptive lateral connections of the network.