Researchers in Germany have discovered what they say is a way to get computers to do more than execute all the steps of a problem-solving calculation as fast as possible – by getting them to imitate the human brain’s habit of finding shortcuts to the right answer.
Most processors are designed to break complex problems into smaller bits that can be processed simultaneously – but identically – using the parallel-processing capabilities designed into the silicon itself. Processors are able to run problems in parallel, but most programs still present problems serially, wasting much of the potential efficiency already built into existing chips.
A team of scientists from Freie Universität Berlin, the Bernstein Center Berlin, and Heidelberg University have refined the idea of parallel computing into one they describe as neuromorphic computing. In their design, a whole series of processors designed as silicon neurons rather than ordinary CPUs are linked together in a network similar to the highly interconnected mesh that links nerve cells in the human brain. Problems fed into the neuro mesh are broken up and processed in parallel, but not always using the same process. The method by which neuromorphic processors handle problems varies with the way they’re linked together, as is the case with neurons in the brain.
Once connected in an efficient way, the neuromorphic network is able to optimize its own way of approaching a problem, according to Michael Schmuker, lead author of the study. The result is a processing bed that is able, all by itself, to classify data according to different features of the data itself – a problem that is extremely difficult for ordinary processor setups and software designs, according to an announcement of the experiment’s publication in the Proceedings of the National Academy of Sciences (PNAS).
Systems able to classify data without requiring humans to do it for them first could make many processing challenges far more efficient in devices ranging from cell phones to supercomputers, according to the paper. One problem in designing such systems, however, is that no two silicon-based neurons are identical, just as no two biological neurons are the same.
The chips are designed to copy the layout and functions of brain cells, but the way they’re interconnected is based on another highly efficient biological model. “The design of the network architecture has been inspired by the odor-processing nervous system of insects,” Schmuker said. “This system is optimized by nature for a highly parallel processing of the complex chemical world.”
In tests using real-world datasets, the prototype was able to match the performance of specialized Bayeseian pattern-matching systems. Even better, the stable decisions reached by “output neuron populations” take approximately 100 milliseconds, which is the same speed required by the insect nervous systems on which the network design is based, according to the paper. With some improvement and standardization of the silicon, it should be simple to apply the same techniques to other problems and datasets as well, the researchers concluded.
Image: Bernstein Center for Computational Neuroscience Berlin