New Albany, Indiana, April 14, 2013. Peter presented his Synthetic Neuro-Anatomy chip for Artificial Intelligence applications at the Midwest Artificial Intelligence and Cognitive Sciences Conference in Louisville, Kentucky today. While other presentations focused on mathematical methods to extract patterns from complex data sets, his presentation focused on structural errors in traditional Artificial Intelligence technologies and how to overcome these problems with new technology.
The focus on intelligence has been lost in mathematical analysis, which is used to extract patterns from data sets. The meaning of the word ‘intelligence’ has been watered down to something much less than its original intent. Intelligent behavior can never be created by such analysis. What is now known as ‘Artificial Intelligence’ should be classified as ‘Computational Analysis and Control’. ‘Genetic Algorithms’ were once hailed as the way to create intelligence, but have proven to be nothing but optimizing tools that are loosely based on an evolutionary model. After 70 years of research and development, the human brain remains the only example of an intelligent system. It is time to return to the original intent of Artificial Intelligence, and reverse engineer the brain to put the quest for intelligence back in Artificial Intelligence.
Computers have a difficult time emulating the functions of the brain. A supercomputer like the IBM Blue Gene /P can emulate only a fraction of it. A simulation of 1.6 * 109 neurons and 8.87 * 1012 synapses runs 643 times slower than real time. That is 0.155% of the brain’s performance at 0.5% of its capacity. It consumes 100,000 times more power its more than 1000 times the brain’s size. Peter stated that “The greater the differences in architecture are, the greater the complexity is to emulate one computer system on another computer system”. In other words, software complexity increases linear with differences. The brain is about as different from a computer as can be. Therefore it takes an infinite amount of effort to emulate the function of the brain on a computer.
Peter started his talk by highlighting the differences between brain architecture and computer architecture. The brain does not have any processors, nor does it have an address bus. The multiple data buses of the brain are the equivalent of millions of bits wide. Every cell in the brain has synapses which are electro-chemical analog memory. Hundreds of Terabytes of storage are distributed throughout the brain. This memory is addressed directly by millions of input pulses, constituting a huge associative memory. Synapses are everywhere, so memory is everywhere. When the brain is formed, DNA builds the scaffolding for learning to take place. Nearly everything we do and say is learned. This is evident in the brains of identical twins who share the same DNA. The connectome of adult brain of such twins is different, proving that DNA is not responsible for its wiring.
Peter then proceeded to show two short video clips. C3PO (Starwars), Rosie (The Jetsons) , Andy (iRobot) and R2D2 are just fantasies, but could such robot technology become a reality? The next video clip showed a Toshiba robot playing a simple tune on a violin. The robot plays this simple tune all in first position and open string, which does not impress a violinist (Peter’s wife plays the violin and several other instruments). The robot is basically a CNC machine, programmed to make the movements that play the tune. It is not intelligent. To have sufficient processing power to create complex behavior as seen in the movies would require a computer some 4000 times faster than the fastest supercomputer.
Peter then introduced his latest invention, a circuit that perfectly mimics the actions of a brain cell. He said “The solution that we have found is not a computer in the traditional sense. Instead, it is a synthetic neuro-anatomy that is modeled on the bran, simulating the functions of brains cells and its memory method in about 8000 binary logic gates”.
The design was tested in a FPGA chip and it has proven to learn, rather than to be programmed, to perform a specific task. Putting 10,000 of these artificial neural cells on a single chip makes a compact, power efficient brain emulator component that can be used in either stand-alone mode or as a coprocessor to a microprocessor. It makes real-time brain emulation possible for the first time, even on small PCs. A fully trained device contains a ‘training model’, which can be copied to a library using the microprocessor interface. Building up such function libraries provides a growth path that will eventually lead to intelligent robots.
The component can be used in prosthesis, to vastly improve speech recognition, in medical implants, in autonomous robots, as a coprocessor in gaming PCs and in many other tools that require the instant recognition of learned patterns. It opens up a new era of cognitive computing and may well be the most significant break-through of the 21st century.
Peter has described the development, and the biological basis of his invention in his new book “Higher Intelligence” that is available as an e-book or as a paperback from Amazon.com.