Blissful Brain

The Blissful Brain is published by Gaia Thinking. For more information on how to order your copy, please click here.

Guardian G2: Mind over matter by Andy Darling

"Neuroscientist Shanida Nataraja has proven meditation does more than clear your head, it can put both halves of your brain to work, improving your concentration, memory, and decision-making...". To read more, please click here.

Upcoming talk: Yoga Ananda, Reigate, Surrey on Friday the 4th of June

Shanida Nataraja will be speaking at a seminar on The Blissful Brain on Friday, 04th June 2010 at 19:30 at Yoga Ananda Ltd. 46 Albert Road North, Reigate, Surrey, RH2 9EL. For more information, please click here.

Artificial Intelligence

The human brain has been systematically dissected into ever smaller pieces until the basic building block was identified; the neurone. Each neurone was seen to function according to specific, pre-defined rules, and this led researchers to believe that some of the complex cognitive functions of the human brain could be successfully reproduced in artificial networks of computer components, if these components were programmed with the same, pre-defined rules. In the mid-20th Century, therefore, pioneers in the field of artificial intelligence began to assemble the first artificial networks out of computer components. Like the human brain, these artificial neural networks are composed of relatively simple processing elements, referred to as nodes or units, and the global behaviour of the network is determined both by the individual properties of these processing elements and by the connections between them. The simplest type of node is referred to as a perceptron. First described by Warren McCulloch and Walter Pitts in the 1940s, a perceptron is a single processing unit that has been programmed to receive inputs, integrate these inputs, and then produce a single output that reflects the data contained in all of the inputs.

These perceptrons can be built into complex networks and programmed with increasingly complex functions; however, this mechanical approach to building intelligence has still failed to produce neural networks on the same scale, or capable of the same scope of function, as the human brain. In the late 1980s, for example, an artificial speech synthesizer, NetTalk, was created, capable of converting English text into speech. This network was trained to pronounce English words through a process of trial and error: the network was not instructed how to pronounce the words correctly, but was programmed with information about how badly the words had been pronounced (i.e. the degree of error). The way in which NetTalk learned to pronounce text was likened at the time to the way in which children learn to talk: NetTalk was first seen to randomly babble; then it became aware of the difference between vowels and consonants; and then it learned of the existence of boundaries between words. A total of 203 individual perceptrons, and 4000 connections between these components, were needed to create a neural network capable of such a basic human skill, and it took 50,000 presentations of 1024 words to train the network to pronounce text correctly 95% of the time.

Although the value of this achievement should not be downplayed, the ability to converted written text into spoken language is only one skill of many mediated, often simultaneously, by the human brain. We are therefore far from modelling the full scope of human intelligence; we can only model specific, isolated skills. Furthermore, there is a fundamental difference between artificial neural networks and their biological counterparts, namely that artificial neural networks require external instruction. Not only are artificial neural networks pre-programmed with specific rules that determine the flow of information through the network, they also receive continuous input from a computer programmer on how well the network is performing, and this information subtly fine-tunes the network’s inbuilt set of rules. Much of the learning that takes place within biological neural networks occurs without supervision: the network spontaneously fine-tunes itself through a process of trial and error.

The difference between artificial and biological neural network is best illustrated by the following example. Human and artificial intelligence was famously pitted against each other in 1997 when IBM’s supercomputer, Deep Blue, played chess, and defeated, the world chess champion, Gary Kasparov. Although it became clear that it was possible to programme a computer with the analytical and strategic skills necessary to excel in the game of chess, this event also acted to highlight the still enormous gap between human and artificial intelligence. At the time, Deep Blue was capable of considering more than two million moves in a single second. It is perhaps not surprising, therefore, that Kasparov, who is reported to make approximately two moves a second, was unable to beat his 1.4 ton opponent. Deep Blue cannot, however, make a move that it has not been pre-programmed to make: it cannot improvise; it cannot make an educated guess; it cannot make inferences (i.e. read between the lines). Furthermore, its circuitry has been specifically designed and programmed with the skills necessary to play chess; however, it is not capable of producing the wealth of behaviour exhibited by its human opponent: it cannot listen to and critique a piece of music; it cannot drive a car; it cannot play squash. The human brain, on the other hand, doesn’t execute prewritten programmes; it learns from experience and encodes this learning by reconfiguring its own hardware.

Back to index