Silicon following the human brain: IBM unveils cognitive computer chips

IBM's new processing cores combine digital "neurons" and "synapses" - for a self-learning computer system that does not require prior programming

IBM unveils a new generation of experimental computer chips that mimic the perceptual abilities of the human brain, its ways of working and the processes of consciousness. IBM's new technology may make it possible to build systems that will consume orders of magnitude less electricity and space - compared to the computers we know today.

IBM's new neuro-synaptic processing chips represent a sharp departure from traditional concepts of designing and building computer systems. The new chips create an electronic equivalent to the operation of neurons and synapses in biological systems such as the brain. This comparison is possible thanks to the use of sophisticated algorithms and silicon circuits with a new design. Two new prototypes for such chips have already been produced in practice, and are now in testing processes.
Cognitive computers, which will be built around these chips, will not be programmed in the same traditional way known today in the world of information systems. Instead of such programming, cognitive computers are expected to learn from experience, identify connections and correlations, present hypotheses, remember to learn and draw lessons from the results. In fact, this process mimics the way the human brain works, and the learning processes we are familiar with.

In order to meet these tasks, IBM's researchers and developers combine principles from the world of nanotechnology, brain research and supercomputing. The project is also partly financed by a new grant from DARPA - the research body of the American defense system - for a scalable electronics system with self-adaptive neural morphology (Systems of Neuromorphic Adaptive Plastic Scalable Electronics - SyNAPSE).

The SyNAPSE project is designed to build a system capable of not only analyzing complex information received from a large number of sensor units and external sources - but also dynamically organizing its internal wiring routes, based on interaction with the environment, in processes parallel to those in which the brain manages to learn and change behavior patterns using a level low energy.
The current project goes far beyond the concept outlined by von Neumann, which has dominated the computing world for the past fifty years. Future computing applications are expected to require functional flexibility that is not offered in traditional systems familiar to us. The new chips are an important step in the development of computers from calculation systems to learning systems. They herald the beginning of a new generation of computers - and new applications of these computers in the world of business, science and government.

Although the new chips do not incorporate biological components, the prototypes of IBM's cognitive processors use silicon circuits inspired by the world of brain research. Thus, it is possible to build what IBM calls a "neuro-synaptic core". This includes an integrated memory equivalent to that of the synapses in the human brain, along with a processing capacity equivalent to that of the neurons, and communication equivalent to that of the axons in the brain.

IBM currently operates two prototypes of such processing cores. The two cores are manufactured using silicon technology on an insulator, with a separation of 0.45 microns, and each contain 256 neurons. One core offers 262,144 programmable synapses, and the other 65,536 learnable synapses. The IBM development team has already successfully demonstrated simple applications running on top of these cores, such as navigation, machine vision, uniform pattern recognition, associative memory and group sorting.

IBM's cognitive computing architecture relies on a network of cores operating on a single chip, in order to create one integrated system of hardware and software. This architecture constitutes a fundamental change from traditional computing concepts, which is expected to give birth to more efficient machines in terms of electricity consumption - and without the need for prior programming. The combination between the memory and the processor at the discrete core level actually mimics the way the brain thinks, which is guided by events and external input, and is carried out in a distributed and parallel process.

The long-term goal that IBM is striving for is to build a chip system that will include ten billion neurons and hundreds of millions of synapses - and consume only a single kilowatt of electricity, and a total volume of less than two liters.

Future chips could process data from complex environments that characterize the real world, using a wide variety of sensors - and act in different ways in order to reach a coordinated and context-dependent response pattern.

Thus, for example, cognitive systems that monitor climate data for the purpose of long-term water supply planning could analyze the data received from the field regarding temperature, barometric pressure, tidal variables - and predict the changes in precipitation and the needs of the water systems on the one hand, as well as generate tsunami warnings on the other hand. Similarly, traffic lights will be able to monitor traffic data at the intersection in order to adjust their activity cycles to actual loads, and adjust themselves repeatedly with any change in traffic.