<p>A new form of computer memory might help machines match the capabilities of the human brain when it comes to tasks such as inter-preting images or video footage.<br /><br /></p>.<p>Researchers at IBM used what’s known as phase-change memory to build a device that processes data in a way inspired by the workings of a biological brain. Using a prototype phase-change memory chip, the researchers configured the system to act like a network of 913 neurons with 165,000 connections, or synapses, between them. The strength of those connections change as the chip processes incoming data, altering how the virtual neurons influence one another. By exploiting that property, the researchers got the system to learn to recognize handwritten numbers.<br /><br />Phase-change memory is expected to hit the market in the next few years. It can write information more quickly, and pack it more densely, than the memory used in computers today. A phase-change memory chip consists of a grid of “cells” that can each switch between two states to represent a digital bit of information—a 1 or a 0. In IBM’s experimental system, each “synapse” is represented by a pair of memory cells working together.<br /><br />Computer scientists have been working for some time on chips that crudely mimic neurons and synapses. Such “neuro-morphic” designs are radically different from the chips we use today. But they promise to make computers that are efficient at tasks computers normally find challenging, such as learning from experience or understanding video.<br /><br />The experimental system announced by IBM researchers, last year, is much less powerful than that chip. But the fact the new system’s 165,000 synapses are made using phase-change memory is significant.<br /><br />Phase-change memory is thought to be particularly well suited to neuromorphic computer systems because it stores data so densely, making it possible to create brain-inspired systems with many more synapses, says Burr. Phase-changememory is also simpler to reprogram. That makes it practical for building a neuromorphic system that is able to “learn” by adjusting its behaviour as it is fed new data.<br /><br />The new system, built with colleagues at IBM and Pohang University of Science and Technology, in Korea, is more than 1,000 times that size. A paper on their results was presented at the International Electron Devices Meeting in San Francisco earlier this month.<br /><br />The team was able to make a larger system because it developed techniques to measure and compensate for the natural variability in the performance of each unit of phase-change memory. <br /><br />Similar variability affects the conventional memory chips in our phones and computers today, but error-checking methods are more advanced for those devices.After being shown 5,000 labelled images of handwritten digits from a standardised data set, the researchers’ chip could recognise handwritten digits it had never seen before with an accuracy of 82 per cent. Burr says that a recent tweak to his team’s error compensation methods should allow accuracy to climb to close to 99 percent.<br /><br /></p>
<p>A new form of computer memory might help machines match the capabilities of the human brain when it comes to tasks such as inter-preting images or video footage.<br /><br /></p>.<p>Researchers at IBM used what’s known as phase-change memory to build a device that processes data in a way inspired by the workings of a biological brain. Using a prototype phase-change memory chip, the researchers configured the system to act like a network of 913 neurons with 165,000 connections, or synapses, between them. The strength of those connections change as the chip processes incoming data, altering how the virtual neurons influence one another. By exploiting that property, the researchers got the system to learn to recognize handwritten numbers.<br /><br />Phase-change memory is expected to hit the market in the next few years. It can write information more quickly, and pack it more densely, than the memory used in computers today. A phase-change memory chip consists of a grid of “cells” that can each switch between two states to represent a digital bit of information—a 1 or a 0. In IBM’s experimental system, each “synapse” is represented by a pair of memory cells working together.<br /><br />Computer scientists have been working for some time on chips that crudely mimic neurons and synapses. Such “neuro-morphic” designs are radically different from the chips we use today. But they promise to make computers that are efficient at tasks computers normally find challenging, such as learning from experience or understanding video.<br /><br />The experimental system announced by IBM researchers, last year, is much less powerful than that chip. But the fact the new system’s 165,000 synapses are made using phase-change memory is significant.<br /><br />Phase-change memory is thought to be particularly well suited to neuromorphic computer systems because it stores data so densely, making it possible to create brain-inspired systems with many more synapses, says Burr. Phase-changememory is also simpler to reprogram. That makes it practical for building a neuromorphic system that is able to “learn” by adjusting its behaviour as it is fed new data.<br /><br />The new system, built with colleagues at IBM and Pohang University of Science and Technology, in Korea, is more than 1,000 times that size. A paper on their results was presented at the International Electron Devices Meeting in San Francisco earlier this month.<br /><br />The team was able to make a larger system because it developed techniques to measure and compensate for the natural variability in the performance of each unit of phase-change memory. <br /><br />Similar variability affects the conventional memory chips in our phones and computers today, but error-checking methods are more advanced for those devices.After being shown 5,000 labelled images of handwritten digits from a standardised data set, the researchers’ chip could recognise handwritten digits it had never seen before with an accuracy of 82 per cent. Burr says that a recent tweak to his team’s error compensation methods should allow accuracy to climb to close to 99 percent.<br /><br /></p>