Hugo de Garis - Evolvable Hardware

Evolvable Hardware

From 1993 to 2000 de Garis participated in a research project at ATR's Human Information Processing Research Laboratories (ATR-HIP) which aimed to create a billion neuron artificial brain by the year 2001. The project was known as "cellular automata machine brain," or "CAM-Brain." During this 8 year span he and his fellow researchers published a series of papers in which they discussed the use of genetic algorithms to evolve neural structures inside 3D cellular automata. They argued that existing neural models had failed to produce intelligent behaviour because they were too small, and that in order to create "artificial brains" it was necessary to manually assemble tens of thousands of evolved neural modules together, with the billion neuron "CAM-Brain" requiring around 10 million modules; this idea was rejected by Igor Aleksander, who said "The point is that these puzzles are not puzzles because our neural models are not large enough."

Though it was initially envisaged that these cellular automata would run on special computers, such as MIT's "Cellular Automata Machine-8" (CAM-8), by 1996 it was realised that the model originally proposed, which required cellular automata with thousands of states, was too complex to be realised in hardware. The design was considerably simplified, and in 1997 the "collect and distribute 1 bit" ("CoDi-1Bit") model was published, and work began on a hardware implementation using Xilinx XC6264 FPGAs. This was to be known as the "CAM Brain Machine" (CBM).

The researchers evolved cellular automata for several tasks (using software simulation, not hardware):

• Reproducing the XOR function.
• Generating a bitstream that alternates between 0 and 1 three times (i.e. 000..111..000..).
• Generated a bitstream where the output alternates, but can be changed from a majority of 1s to a majority of 0s by toggling an input.
• Discriminating between two square wave inputs with a different period.
• Discriminating between horizontal lines (input on a 2D grid) and random noise.

Ultimately the project failed to produce a functional robot control system, and ATR terminated it along with the closure of ATR-HIP in February 2001.

The original aim of de Garis' work was to establish the field of "brain building" (a term of his invention) and to "create a trillion dollar industry within 20 years". Throughout the 90s his papers claimed that by 2001 the ATR "Robokoneko" (translation: kitten robot) project would develop a billion-neuron "cellular automata machine brain" (CAM-brain), with "computational power equivalent to 10,000 pentiums" that could simulate the brain of a real cat. de Garis received a US$0.4 million "fat brain building grant" to develop this. The first "CAM-brain" was delivered to ATR in 1999. After receiving a further US$1 million grant at Starlab de Garis failed to deliver a working "brain" before Starlab's bankruptcy. At USU de Garis announced he was establishing a "brain builder" group to create a second generation "CAM-brain".