| 摘要 |
A self-organizing system providing improved performance is composed of node subcircuits in two or three dimensional arrays of nodes which behave like neurons in the brain. Improvements in the learning rules use the time-filtered output of nodes to define memory traces. Spatial summation and spatial difference functions then determine how node branches will compete to produce various memory trace topologies such as roots and junctions. Roots begin as input patterns at the lowest level of the circuit and grow towards output nodes at the highest level; roots are attracted to output nodes and to other roots as they grow. Roots connect or branch at junctions which are identified by spatial functions. By modifying node properties and branch competition of nodes at root junctions, roots interact to create Boolean logic roles. Unsupervised (classical) learning results when roots associate with each other. Supervised (operant) learning regulates root junction logic to assure that sequential or combinational system input patterns produce the proper system outputs. Punish or reward signals broadcast to all nodes are only acted on by memory trace root junction nodes. Implementation is in digital and analog circuitry as well as hardware and software embodiments.
|
