I've been thinking about building graph-optimized hardware and how one might use analog function cells as the baseline for a function machine. What do I mean by function machine?
An intelligence system device which stores functions and their relationships between and towards other functions as well as the results of those functions applied to data sets on a periodic basis. I have a programming model for such a machine that can be modelled via 3D interactions of units that can attach to each other in simple ways, so programming functions is analogous to connecting some pipes and boxes and fittings together. So you might have a function that takes an input series and another slower changing series and outputs a differential equation. A function cell would contain both that algorithm / transformation / equation for a given function as well as add storage and caching and optimized data retrieval instructions, structures, and algorithms. Function cells are composable, i.e. they can have the connection topology equivalent to the different faces of any regular space-filling 3D tessalation, starting with the simplest and most flexible, the truncated tetrahedron, 4 - hexagon faces and 4 triangle faces. The mapping between physical object and mathematical objects and abstractions made possible by the real-world connection possibilities afforded by the physical connection restrictions creates a better interface for reliably connecting complex systems in a fast and efficient manner. Imagine pipes leading from sensors to sense maker- / detector- / observer- systems. There is enough information in the physical connection in the 3D model to automate most programming if the system uses a single common global semantic lexicon. When a connection is made the software could automatically handle the communication links required between sensors and neural-network cells implemented via the function cell topology. So programming becomes connecting these parts logically and defining their internals recursively. Then you assemble them, twiddle a few constants knobs and you can address them and query them instantly. You program instantly as state changes made to the connected cell graph or to settings for each cell flow through all cells during a single update cycle. If you look at call as a unit of composition that adds storage and hardware implementation cell rather than the code that implements a function, then some new capabilities emerge with rare benefits that are not obvious because the simplifications that arise are trans-dimensional and transitive. The equations for this simplification are equally simple to conceive because they follow Metcalf's law. This has important implications for distinction graphs and series as well as for any classification problems. But that's a better topic for the whiteboard. And I'd like to explain the implications for robotics before everyone heads off to LA. I have drawn a picture on the whiteboard in the office with a design which leverages this idea in a physical topology suitable for human robot legs. - Curtis -- You received this message because you are subscribed to the Google Groups "opencog" group. To unsubscribe from this group and stop receiving emails from it, send an email to opencog+unsubscr...@googlegroups.com. To post to this group, send email to opencog@googlegroups.com. Visit this group at https://groups.google.com/group/opencog. To view this discussion on the web visit https://groups.google.com/d/msgid/opencog/CAJzHpFrdRY1KFmBN1KenS-scwRjP_C7Wy3jdMR3pMVP9JUgXug%40mail.gmail.com. For more options, visit https://groups.google.com/d/optout.
