ATR ' s CAM - Brain Project : The Evolution of Large - Scale Recurrent

This paper describes ongoing ATR's CAM-Brain Project, which is an attempt to build large-scale neural networks ('artiicial brains') in a special hardware called "CAM-Brain Machine" (CBM). At the time of writing (March 1998), the project is making eeorts on two fronts-the construction of the CBM, that is scheduled to be operational in the summer of 1998, and attempting to nd an eecient and eeective representation for the binary signaling of ATR's CAM-Brain Machine (CBM), using the so-called "CoDi-1Bit" model. The CBM is an FPGA based hardware accelerator which updates 3D cellular automata (CA) cells at the rate of 100 billion a second, allowing a complete run of a genetic algorithm with tens of thousands of CA based neural net circuit growths and hardware compiled tness evaluations. It is hoped that by using such a device, it will become practical to evolve 10,000s of neural net modules and then assemble them into humanly deened RAM based ar-tiicial brain architectures which can be run by the CBM in real time to control robots, e.g. a robot kitten. Before large numbers of modules can be assembled together, it is essential that the individual modules have a good functionality and evolvability. The "CoDi-1Bit" CA based neural net model uses 1 bit binary signaling, so a representation needs to be chosen based on this fact. This paper discusses the merits and demerits of a representation that we call "Spike Interval Information Coding" (SIIC) and presents some simulation results obtained so far concerning the evolution of simple functional modules and the performance of the SIIC representation .