A brain that makes light of our problems

An artificial brain powered by light that has the answers to many of our most pressing problems has been designed and demonstrated by British scientists.

Nicknamed 'the brain box' it uses laser light to provide solutions to a range of vexing questions ranging from curing congestion on the Internet - problems which have led many to call it the 'World Wide Wait' - to getting trains to run on time or even screening programs for cancerous cells.

This is claimed to be a world first for the Optically Interconnected Computing (OIC) researchers at the Department of Physics at Heriot-Watt University in Edinburgh.

They have already shown a somewhat startled world a fully functional opto-electronic 'brain that works on light' that can solve any complex scheduling problem.

This innovation is the result of six years of continuous effort involving both its construction and programming.

"The underlying strength of this opto-electronic system is that it uses laser light beams to weave incredibly complex interconnection patterns and carry huge volumes of information in a way simply not feasible in conventional electronics," said Dr John Snowdon, the project team leader.

Principally, the system has been designed to overcome one of the fundamental problems in communications. Paradoxically, this is the transferring of data from one person to another.

Although connecting two people may appear a relatively simple task, in reality it is extraordinarily complex.

Not only do our systems have to handle many connections simultaneously but also this convoluted undertaking often has to be carried out under adverse conditions such as localised overloading or hardware failure.

It is these sorts of scheduling problems that lead to hold-ups in networks such as the Internet, with immense frustration for users.

Neural networks have the ability to solve such scheduling problems efficiently but limitations on the complexity and scalability of electrical interconnects on a conventional silicon chip have so far hindered the construction of any hardware.

Heriot-Watt's opto-electronic neural network has overcome these inherent difficulties by using optics to provide high-speed, high-volume interconnections, combined with off-the-shelf electronics to add information processing facilities, or 'neuron functionality'.

In the system, arrays of detectors and vertical cavity surface emitting lasers (VCSELs) act as neuron inputs and outputs with complex neural interconnection patterns woven through free space using a single diffractive optical element.

Neural summation is simply the amount of light incident on a neuron's detector. All that electronics need do is choose the neuron's next response based on input light and communicate with the outside world.

Diffractive optics are key to the power and flexibility of the system as they enable light to be manipulated into complex patterns in a controlled manner.

Conventional computers are used to perform 'reverse calculations' from the desired light energy distributions.

The results are then stored in an element of hardware that is etched into the required form by a process similar to that used in silicon chip manufacture.

Once created, the element can perform its function indefinitely and may be easily reproduced. The technology for these key elements was also developed in Heriot-Watt's OIC group by Mohammed Taghizadeh and Andrew Waddie.

This technology is advancing rapidly, both in terms of functional abilities and, perhaps more importantly, from an industry point of view, in terms of packaging.

This is graphically shown by the fact that the first-generation demonstrator produced, in collaboration with British Telecom (BT), took up almost the whole of the surface of an optical bench.

By comparison, the recently completed second-generation demonstrator, as constructed by Dr Keith Symington, could fit easily into a shoebox.

If that were not enough, the soon-to-come third-generation version will integrate many of the components used in the last system onto a single chip, and they expect to be able to fit two neural networks into a demonstrator the size of a pencil case.

At this level of integration they believe they will be getting near to the size of a system required for commercial viability.

To sum up the project so far, the adaptation and optimisation of algorithms for the specific hardware used has considerably increased system scalability and performance.

Doubling the size of the packet-switch routing problem only increases the time required to reach a solution by a few per cent. Since the range of problems that a neural network can solve is vast, minimal alteration allows adaptation to a variety of tasks.

These range from image recognition to general optimisation and task allocation problems. "We are all excited about our recent success in these opto-electronic systems," said Dr Snowdon. "In technical terms we have found that algorithmic adaptation and optimisation, specifically for the hardware used, has increased system scalability and performance. This would, therefore, allow continued expansion of networks such as the Internet or the currently topical Computational Grid systems. This result is unprecedented and could not have been imagined in a traditional, purely digital electronic system."

He believes that the range of problems that neural networks can solve is potentially as great as those only normally possible for the human brain.

"Our system can provide solutions to a variety of tasks that are relatively intractable to even the fastest digital machines. These range from image recognition, for example differentiating between healthy and cancerous cells, to general optimisation and task allocation problems like distributing jobs efficiently among multiple processors in a cluster or supercomputer system.

"Other successful applications could include heart-rate monitoring and stock-market prediction. The versatility and fault tolerance of this type of circuitry makes it a very attractive computational component of the near future, and Heriot-Watt is leading the way in this exciting new field."

Heriot-Watt's OIC Group is now starting on two new projects. One, a European Union-funded project, is to construct (with a number of collaborators) an optically interconnected processor-memory bus for a multi-processor machine.

The second is to integrate optical interconnects with reconfiguable silicon electronics in the form of field programmable gate arrays. Both of these will build on the successes of the group's previous projects.