Each core consists of computing components analogous to their biological counterparts—core memory functions similar to the brain’s synapses, processors that provide the core’s nerve cells (or neurons), and communication capabilities handled by wiring akin to the brain’s axon nerve fibers. The IBM and Cornell University researchers heading this project published their results in the August 8 edition of Science.
Researchers from the two institutions are working together as part of the Defense Advanced Research Projects Agency’s (DARPA) Systems of Neuromorphic Adaptive Plastic Scalable Electronics (SyNAPSE) project. SyNAPSE seeks to reverse-engineer the brain’s computational abilities to make a computer that can mimic our ability to sense, perceive, act, interact and understand different stimuli. DARPA has spent about $53 million so far on this approach to SyNAPSE since 2008. (HRL Laboratories is leading DARPA’s other SyNAPSE project.)
IBM and Cornell designed their new chip to behave like an energy-efficient spiking neural network. Unlike a normal neural network that processes data at regular intervals, a spiking neural network is more efficient in that it fires only when an electrical charge reaches a specific value. The firing, in turn, influences the charge at other artificial neurons—much like what happens in the real brain, according to the Science study. Each of the chip’s neurosynaptic cores features 256 input lines that work like axons and 256 output lines that work like neurons. IBM’s new chip, meanwhile, includes 4,096 neurosynaptic cores, producing more than one million programmable spiking neurons and 256 million configurable synapses.
The ultimate goal of IBM and its SyNAPSE partners is to build shoebox-sized neurosynaptic supercomputers with 10 billion neurons and 100 trillion synapses that consume a single kilowatt of power. (The human brain has about 100 trillion synapses but uses only about 20 watts, or roughly what it takes to power an oven light.)
The researchers tested their chip by having it analyze a video clip, picking out people, bicyclists and other objects from background and then identifying each object. Compared with a simulator running on the same type of artificial neural network using a modern general-purpose microprocessor, the new chip configuration consumed 176,000 times less energy while performing 100 times faster each frame, the researchers report. The chip likewise improves upon its predecessor, consuming 100 times less power per core than the original version IBM introduced in 2011.
IBM envisions neurosynaptic chips as the building blocks for a new breed of supercomputer that mimics the human brain using microchips only a few millimeters in size and are energy efficient enough to be embedded into eyeglasses, wristwatches and other wearable accessories. These chips are also designed to be highly proficient at sorting through sensory input, making them good candidates for use in medical diagnoses For example, IBM researchers envision a digital thermometer outfitted with a cognitive sensor that could scan and sense a mix of chemical signals in a sick child’s mouth and quickly deliver a diagnosis.
More immediate efforts to deliver cognitive technology are focusing on IBM’s Watson computer. In February 2011 Watson defeated two former champions on TV’s Jeopardy! quiz show thanks to its ability to search a factual database in response to questions, determine a confidence level and, based on that level, buzz in ahead of competitors. Late last year the University of Texas M. D. Anderson Cancer Center and the Cleveland Clinic Lerner College of Medicine of Case Western Reserve University began testing smaller, more powerful version of Watson as a tool for data analysis and physician training. A January Wall Street Journal article spelled out several challenges IBM has faced as it exposes Watson to the rigors of real-world data analysis. Even as IBM works through these early kinks the company plans to make Watson available as a cloud service, potentially turning any Internet-connected device into a cognitive computer.
source- Scientific American
