Dharmendra Modha Mimics Brain's Efficiency

IBM's Dharmendra Modha has taken another big step toward developing a computer that can match the human brain in efficiency.

IBM researcher Dharmendra Modha has received $21 million more from Defense Advanced Research Project Agency (DARPA) to develop a computing device built with chips that mimic the way a biologic brain works.

The so-called Synapse Project aims to show the way to build intelligent machines that can function with the energy efficiency of the human brain, thereby escaping the limits imposed on conventional computers by their power-consumption and resulting heat.

Modha and his team have already developed brain chips — dubbed neural cores — with electronic synapses and memory circuits that mimick two aspects of the brain: the proximity of circuitry for memory and computation and the ability to strengthen and weaken connections between these parts as new data input results in “learning”.

Compared with conventional sequentially processing von Neumann-style processors used in virtually every computer in common use, Modha’s brain chips are extremely energy-stingy, with very little of the waste heat that limits the ability to pack more computing power into conventional chips. The energy efficiency results in large part from the fact that the brain is a million to a billion times more efficient than von Neumann machines in processing complex data.

“This is a major initiative to move beyond the von Neumann paradigm that has been ruling computer architecture for more than half a century,” said a statement by Modha. “Future applications of computing will increasingly demand functionality that is not efficiently delivered by the traditional architecture. These chips are another significant step in the evolution of computers from calculators to learning systems, signaling the beginning of a new generation of computers and their applications in business, science and government.”

This effort to mimic the biologic brain must be done using conventional chipmaking materials and techniques — i.e. 45 nanometer CMOS with silicon-on-oxide doping, the same process IBM uses to etch Power7 processors. So far IBM has created two prototype neurosynaptic chips, each with 256 simulated neurons. One has 262,144 programmable synapses and the other has 65,536 learning synapses. The neurosynaptic cores replicate the function of brain synapses, neurons and axons to produce memory, computation and communication.

These chips have about the brainpower of a slug, but Modha’s team has already used them for tasks like pattern recognition, navigation, machine vision and associative memory. Over time the Synapse project aims to use neurosynaptic chips to build computers with 10 billion electronic neurons and 100 trillion synapses — about ythe complexity of a human brain — inside a two-liter package that burns one kilowatt. That isn’t quite as energy-efficient as a human brain, which uses only about 10 watts, but it’s thousands of times more efficient than current super-computers.

Modha’s research team and collaborators at various universities will present two papers at September’s Custom Integrated Circuits conference in San Jose, California to reveal the neural cores’ low power requirements and how they will work with neural-circuit-mimicking software. Experiments show a neural core learning to play Pong, navigating a car on a simple race track and recognizing images. These are impressive feats considering that not even biologists fully understand how the brain works.

Modha attacked the problem by using a supercomputer to simulate a neural network as complex as a mouse brain, then moved up progressively to a rat, a cat and a monkey. Each step required a more powerful supercomputer. Even then he was unable to run the simulations in real time due to the inherent limitations imposed on conventional computers by the separation between memory and processor.

“Our eventual goal is a human-scale cognitive-computing system,” says Modha.