Tuesday February 8, 2011
We represent the brain’s operation using concepts that are almost mathematical in their simplicity. Representations occur at three different layers, each sitting on top of and built from the one below:
1) The Semantic Network represents the brain’s “software.” It contains volumes of inactive “causes”, in an efficient offline memory, for each fact manager. The configuration settings from the reenactment simulator characterize the being’s current thought state. Each node of the semantic network is considered to be one such state. The reenactment simulator acts as a player and recorder, as described in the book Godel Echer Bach by Hoftstadter.
2) The Reenactment Simulator characterizes the current world situation and re-creates it in a simulation, built from one or more universal simulators. Each is parametrically controlled by configuration settings, and together they can reenact the current world’s circumstances, inside the being. Models learned from past actions are used to pick the future actions, and therefore to increase the being’s viability. Thus, the being can pick good plans and react quickly. The Reenactment Simulator is built from fact mangers.
3) Facts are defined as the way the brain holds all its world knowledge. Facts include everything from “pixel red” to “want water”. Each fact is managed by a “fact manager” (perhaps related to a cortical minicolumn), which builds a model of the world by attempting to predict the future. Facts automatically learn and discover more sophisticated causes for sensory input (Jeff Hawkins, 2005). A prediction tree driven by standard MSEE techniques is used. This construct is reminiscent of Stoller splits, but has two passes.
A Situation Tree: One of the neural like structures in the book is a tree used to record and decode the sensory experiences presented it. The tree that is growing in the video is an "Improvement Tree" (as described in Section 4.8 of the book). The video shows a tree dynamically accepting and sorting out experiences, and growing to record them. Initially there is a single root. Periodically a packet of sap (purple ball) goes up from the trunk. The purpose of the tree is to distinguish the various kinds of experiential information (IN) the packet might have. As the sap goes up the tree through a crux (green box), selected bits of IN choose the exit branch. If there is confusion, the packet is transformed into a new branch and crux that get added to the tree to record this experience.