From OpenCog
Jump to: navigation, search

OpenCog Hyperon

OpenCog Hyperon is a substantially revised, novel version of OpenCog -- which is currently (Sep 2020) in an early stage of designing and prototyping.

General motivation and a high-level overview for Hyperon are given in File:OpenCog Hyperon.pdf, and with more color and context in the talks presented during the second day of OpenCogCon (video starts with the talk Tentative Sketch of a Next-Generation OpenCog Architecture followed by several talks on more specific questions about Hyperon design, in particular, Thoughts on Atomese 2 and Toward a Distributed Atomspace).

Among other theoretical foundations, the Hyperon design tries to take General Theory of General Intelligence into account at least to some extent. Although the connection with the possible implementation may turn out to be loose, this theory (which is in a moderately advanced but certainly not completed stage) provides some background ideas, principles, and problems constituting the grand picture behind the Hyperon design.

The motivation for creating a radically new OpenCog version comes from practical experience in using OpenCog both in AGI research and in development of more applied narrow-AGI-ish systems. Thus, relevant use-cases provide very important motivation for the OpenCog redesign; in this vein see Hyperon:Motivating_Use_Cases.

Although design choices regarding different components are interconnected, the spectrum of problems and technical details considered for particular components may be quite distinct. Currently, two main layers of Hyperon are being analyzed in more detail:

The higher-level cognitive layer is of course the main purpose of the design, and has been taken into account in various design discussions so far, but has not been the core focus of the Hyperon-oriented technical documents produced so far. In that sense we are proceeding from the "bottom up", starting from discussions of knowledge repositories and underlying languages -- but conceptually our work on these lower-level aspects is driven by our experience with probabilistic reasoning, evolutionary learning, probabilistic programming, neural-symbolic interfacing, pattern mining, robot control, language processing, motivation and other cognitive tasks and processes using the current OpenCog system. Future Hyperon design documents will more directly address cognitive, learning and reasoning aspects.