Memristor emulates neural learning

PORTLAND, Ore. — Synapses are the bit-cells of the brain, and they behave more like memristors than any other electronic circuit element, according to the University of Michigan researchers who recently demonstrated that a single memristor can learn using the same technique as the human brain.

Researchers demonstrate memristors emulating the learning function of a neural network by changing the strength of its synaptic connections in response to synchronized voltage spikes.

Their findings will be published in Nano Letters.

Neural networks can learn patterns that are too difficult for engineers to craft as specific algorithms, but they depend on an analog memory element called a synapse, which today is simulated on supercomputers as a numerical value. Learning occurs when simultaneous voltage spikes are generated from feature detectors in the senses, like edge detectors in the eye. When the simultaneous spikes come in, say from the edge detectors in both eyes, the receiving synapse in the brain responds by increasing its value--a digit used for supercomputer simulations.

Instead, memristors change its resistance value.

According to the University of Michigan researchers led by Professor Wei Lu, memristors respond to these simultaneous voltage pulses--called spike timing dependent plasticity--in a manner nearly identical to that of brain synapses, making them a viable alternative to supercomputer simulations. Massive crossbar networks of memristors, proposed by HP Labs researchers could create a more accurate and much faster executing emulation of brain functions than supercomputer simulations.

Last year, the Defense Advanced Research Project Agency (Darpa) signed up three teams led by HP, IBM and HRL Labs to determine the best way to develop the brain's learning element in its SyNAPSE (Systems of Neuromorphic Adaptive Plastic Scalable Electronics) program. A prototype is due by next year.

Hewlett Packard has been studying the use of memristors as synapses for the Darpa program, and will be describing its efforts later this year.

Last year IBM announced it has achieved an accurate supercomputer simulation of a cat brain, for which it received the Association for Computing Machinery Gordon Bell Prize at Supercomputer 2009. Called Blue Matter, the simulation could eventually be transferred to hardware using electronic synapses like those being developed at University of Michigan.

"The cat brain sets a realistic goal because it is much simpler than a human brain, but still extremely difficult to replicate in complexity and efficiency," said Lu. The goal would be to create memristive devices that someday achieve the performance of a supercomputer in a machine the size of a 2-liter bottle.

The University of Michigan research was funded by both Darpa and the National Science Foundation.