Notes on Semantic Hierarchical Temporal Memory for Perceptual, Motoric and Intentional Intelligence

The Hierarchical Temporal Memory (HTM) approach to perception processing, pursued by prior authors including Jeff Hawkins and Itamar Arel, is expanded to encompass more abstract semantic forms of HTM, and to include separate, coupled hierarchies for perception, motor control and goal management. No experimental evidence or detailed calculations are presented; this is merely a conceptual paper describing some potentially promising ideas. Firstly, an extension of the HTM approach to perceptual AI is suggested, in which a traditional ”perceptual” HTM is coupled with a more abstract ”semantic-perceptual HTM”, while retaining the same essential spatiotemporal pyramidal architecture. This extension is argued to provide more flexible visual pattern recognition, as well as easier interlinkage between HTMs and semantic networks or other abstract cognition systems. Secondly, it is argued that the semantic HTM approach also lends itself naturally to the construction of a motoric hierarchy existing alongside the sensory hierarchy – the key differences being that the semantic-motoric HTM hierarchy refers to configuration space rather than visual space, and is linked to an effector hierarchy rather than a perceptual HTM. Finally, the notion of a goal HTM is presented, whose nodes contain implications of the form ”(perceptual) Context & (motoric) Procedure → Goal”, where the Goal is (spatiotemporally bounded) subgoal of an overall system goal.