Artificial Intelligence: DNA sequencers to dancing robots
Published: 28 Mar 2006 11:40 BST
During the late 1980s some researchers had started to look again at a concept that had first been proposed back in 1943 by Warren McCullough and Walter Pitts of the University of Illinois. The concept involved building electronic analogues of brain cells and had been refined in the early 1960s by Marvin Minsky and Seymour Papert, who had connected them into simple networks known as perceptrons.
These simple networks could be trained to recognise patterns of input data and had found uses in image recognition but had not been further developed because of difficulties in expanding network size. The researchers who were re-examining this technology discovered that the solution to this problem was to construct a multi-layer perceptron, which we now know as a neural network. Neural networks have been one of the great success stories of AI research and in software or hardware form are now found in a wide variety of applications where the system's ability to learn is important.
Dancing Honda
The remarkable ability of neural networks to learn complex tasks is best demonstrated by the ability of Honda's Asimo
humanoid robot to not just walk but dance, and even ride a bicycle.
Actions that are only possible because the neural networks that are
connected to the robot's motion and positional sensors and control its
'muscle' actuators are capable of being 'taught' to do a particular
activity.
The significance of this sort of robot motion control is the virtual impossibility of a programmer being able to actually create a set of detailed instructions for walking or riding a bicycle, instructions which could then be built into a control program. The learning ability of the neural network overcomes the need to precisely define these instructions — instead the robot is 'taught' to perform in a particular way and creates its own instructions within its neural net. This makes neural nets particularly well suited for systems which have to adapt to changing environments.
However, despite the impressive performance of the neural networks controlling Asimo's movement the most significant applications for neural networks are currently to be found in everyday objects, such as a new fire detector that has just been launched by Siemens. This uses a number of different sensors and a neural network to determine whether the combination of sensor readings are from a fire or just part of the normal room environment. Over fifty percent of fire call-outs are bogus, and of these well over half are due to fire detectors being triggered by everyday activities as opposed to actual fires.
Neural networks
The neural network in the Siemens detector
allows the fire detector to be trained to recognise the normal pattern
of temperature, smoke colour and particle size found in the room in
which it is located, as opposed to having these parameters pre-set in
the factory. This means that a detector placed in a workshop would
ignore dust and machinery exhaust, but if it was placed in a bathroom
it would ignore steam, but in both locations it would react to an above
normal rise in temperature coupled with dark smoke.
This is just one of a vast range of current applications for a neural-network based systems. If you use a digital camera then the odds are that the auto-focussing system is based upon a neural network. In the military world neural nets are an essential component of virtually every smart-weapons system and every countermeasures system. They are used in DNA sequencing equipment, in voice recognition and response equipment. And in these days of heightened security neural nets a key part of virtually every biometric analysis system. Indeed neural networks have become a key component in the design kit of every systems engineer.
It is all a question of probability
One thing that became
clear to researchers during the 1980s was that knowledge systems using
conventional logic would never produce the kind of intelligent response
that was being sought. Real world decisions are based not upon
certainty but upon probability.
Dealing with uncertainty involved the use of statistical reasoning techniques where probabilities were assigned to every option. The most popular of these mathematical techniques is known as Bayesian statistics, and when combined with graph theory it provides a powerful means of modelling probabilities based upon continuously updated information.
Bayesian networks can dynamically learn by constantly modifying modelled probabilities using a fixed set of rules, and it is a technique that has proved very popular in a wide range of applications including diagnostic and decision-making systems, data mining, computer vision, bioinformatics and of course robotics. Bayesian networks are...
For more, click here...
Previous
Did you find this article useful?
339 out of 625 people found this useful
- Share this article:
