Artif Life Robotics (1998) 2:56-61 9 ISAROB 1998 ORI(JlNAI+ AR II('LI~ Hugo de Garis 9Felix Gers 9 Michael Korkin 9 Arvin Agah Norberto "Eiji" Nawa "CAM-brain" ATR's billion neuron artificial brain project: A three-year progress report Received: March 2, 1998 / Accepted: April 2, 1998 Abstract This paper reports on recent progress made in ATR's attempt to build a 10000-module evolved neural net artificial brain to control the behavior of a life-sized robot kitten. Key words Artificial brain 9 CAM-brain 9 Artificial neural networks 9 Evolvable hardware 9 CAM-brain machine (CBM) Introduction This paper reports progress in ATR's artificial brain (CAM- brain) projectY The broad aim of this project is to build/ grow/evolve an artificial brain containing a billion artificial neurons by the year 2001. The basic ideas of the CAM-brain project are as follows. Use cellular automata (CA) as the foundation upon which to grow and evolve neural network circuits with user-defined functionality. The state of each cellular automata cell can be stored in one or two bytes of RAM. Since nowadays it is possible to have a gigabyte of H. de Garis ([]) - N. Nawa Brain Builder Group, ATR, 2-2 Hikaridai, Seika-cho, Soraku-gun, Kyoto-fu, Japan Tel: +81-7749-5-1079; Fax +81-7749-5-1008 e-mail: degaris@hip.atr.co.jp F. Gers IDSIA, Lugano, Switzerland M. Korkin Genobyte Inc., Boulder, CO, USA A. Agah Department of Electrical and Computer Science, University of Kansas, USA This work was presented, in part, at the Second International Symposium on Artificial Life and Robotics, Oita, Japan, February 18- 20, 1997 RAM in one's workstation, there is a huge space in which to store the CA cell states, i.e., more than enough to contain an artificial brain. The next consideration in the CAM-brain project was how to evolve these neural networks quickly enough for "brain building" (i.e., the assemblage of 10000 and more of such evolved neural network modules into humanly defined artificial brain architectures). We chose to use evolvable hardware techniques that our team in- vented. 6'7 Special reconfigurable hardware chips (Xilinx XC6264) have been programmed to grow neural circuits literally in nanoseconds, according to a simplified neural growth and signaling model that we call "CoDi. ''8'9 This special hardware will update 100 billion CA cells per second, which is fast enough to complete a full run of a genetic algorithm (e.g., with a population of 100, for hun- dreds of generations) in about a second. A user-specified fitness definition for each module is compiled directly into programmable hardware in order to automate the grown circuit's fitness measurement at hardware speeds. This hardware is called a CAM-brain machine (CBM), I~ and will be built by the spring of 1998. Our team is currently involved in several parallel tasks. The first is the construction of the CBM. The second is the behavioral specification and design of a life-sized robot kitten which a 10000 evolved neural net module artificial brain will control. The third is the design of the brain itself. The kitten robot, called "Robokoneko" (Japanese for "ro- bot kitten") will be controlled by an on-body radio link to the artificial brain consisting of some 80 megabytes of RAM, which is updated 2500 times per second by the CBM, which is fast enough for real-time control. The year 1998 will be spent in gaining experience with the CBM and evolv- ing a large number of modules. At the same time, the kitten robot will be built and tested, and the brain architecture will be designed. In 1999, the brain and the kitten should be brought together, hopefully in time to show the world, at the turn of the millennium, the first ever functioning artifi- cial brain. The CBM's incredible speed should allow brain building to become practical. Our team hopes that the CBM will give birth to a new research field, that we call simply "brain building." If the kitten robot and its brain are suc-