Artificial Intelligence


Example: Natural Language Processing

Traditional neuroscientific techniques may give us some indication of the way that we understand language by mapping out the brain areas that are involved and the basic links between them, but at the moment it looks like computers offer us the best chance of gaining a deeper understanding of how the individual units of language are coded for and used to produce a coherent and semantically and syntactically correct whole.
However trying to model the neural processing of natural language is harder than it may initially appear! We run into all sorts of problems with:
Word order
For example, in the two sentences:

  1. We gave the monkeys the bananas because they were hungry
  2. We gave the monkeys the bananas because they were over-ripe,

the word “they” is in the same place in the two sentences, which are almost identical, however using our knowledge of the English language, we automatically realize that in the first sentence, the “they” refers to the monkeys, whereas in the second it refers to the bananas. This shows that word order alone is not enough to determine syntactic correctness.

Artificial Intelligence-“the science and engineering of making intelligent machines.” (John McCarthy, 1956.)

Artificial Intelligence models, on the other hand, are used to design intelligent agents, without any requirements that it follows the way in which a human brain would carry out this function.  An intelligent agent would be a system that carries actions which maximize its chances of success in its particular environment.

And the main long-term objective of this is to build robots to carry out human functions which may save on time/money/energy etc.
Although neural networks in intelligent machines can vary hugely, and there are a wide variety of artificial intelligence problems that are being solved, there are also some common factors:

  1. They generally implement a learning algorithm-one which modifies connections based on experience, so the weight between two neurons increases if they fire together, decreases every time they don’t.

Each neuron conceptually identical. This is a major advantage as it leads to massive parallelism, and can be encoded into computers easily.

Example: Pattern Recognition

Again, the designing of artificial intelligence to perform these functions is not easy. Taking the recognition of human handwriting as an example, the way individuals form letters can vary so widely that it is extremely difficult to create an algorithm for recognizing individual characters. In addition, when faced with unfamiliar writing, one often has to look at the context and surrounding letters for assistance, so this would also have to be taken into account and written into the program. However the fact that humans (and therefore biological neural networks) are able to carry out these tasks out with relative ease would suggest that the problems can be solved!

Meera Desai, September 2008