The CAM-Brain Machine (CBM) is a hardware implementation of a brain-inspired, recurrent, digital neural network. It is an experimental machine composed of reconfigurable (evolvable) hardware, capable of training and evaluating cellular automata based neural network modules directly in silicon.
The networks of the CBM were originally intended to be built with a genetic algorithm. However, currently the implemented genetic algorithm is not powerful enough to evolve satisfactory networks for applications of a meaningful complexity. To that end, the training technique should be considerably enhanced.
This paper addresses the problem of using the CBM more efficiently, still based on genetic evolution, but using a much more efficient gene pool. The paper focuses on the identification of frequently used primitive functions, and the hand crafting of these functions into efficient basic network patterns. Functions range from simple delay lines over logic gates, to adjustable timers and switches. Furthermore a technique is introduced to build neural structures that can generate arbitrary pulse trains.
Eventually these basic structures will be combined in a library, that will serve as a potent gene pool.