MIT: The ability for machines to reach the performance of the human brain in areas such as image or video interpretation could potentially be provided by a new type of computer memory.
As reported in an article in MIT Technology Review, IBM researchers have used so-called phase-change memory to create a device that processes data in a way that mimics the functioning of a biological brain. Using a prototype phase-change memory chip, the researchers configured the system to operate as a network of 913 neurons with 165,000 connections, or synapses, between them.
The strength of these synapses changes as the chip processes incoming data, changing the way the virtual neurons influence each other. Taking advantage of this feature, the scientists enabled the system to recognize handwritten numbers.
Phase-change memory is expected to hit the market in the coming years. It can record information at high speed and pack it into much higher density than current memory types. A chip of such memory consists of a network of “cells” that can take on 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 cells working together.
Brain-like computers have been the subject of research by computer scientists for some time. Such designs are radically different from today's chips, promising to make computers more efficient at tasks currently considered difficult for conventional systems - such as learning from experience or understanding video.
Earlier this year, IBM announced the most complex chip of its kind yet, created using techniques and components used to build smartphone processors. The new system isn’t as powerful, but as Jeff Burr, an IBM researcher, pointed out, it’s important to note that phase-change memory is used to create the 165,000 synapses. According to him, this type of memory is considered suitable for “brain” (neuromorphic) systems because it stores data at a very high density – and it’s also easier to reprogram. In practice, this makes it easier to build a system that will be able to “learn,” adjusting its behavior appropriately as it receives new data.
Previous efforts in this field have been limited in scale, with 100 synapses or fewer. The new system, built in collaboration with researchers at Pohang University of Science and Technology in Korea, is 1,000 times larger.
The relevant paper was presented in December at the International Electron Devices Meeting in San Francisco.
Source: naftemporiki.gr
