Bubbles_glow

≡ Literary Systems ≡

 
Corner_fold
  • Citation
  • Extension
  • Response
  • Comment
Hopo_catalog
Literary Systems
 
Flowchart_grey_24

Whenever formal rules are defined...

Whenever a scholar employs a computer to reason with you about the structure of language or the structure of literature; whenever you have an encounter with rules that seem to work for the general case, but not for the specific one; whenever a linguist tells you that his model will help a computer understand the meaning of language, remember that there can be more than one species of computation.

All of the computers that we are familiar with today are based on the Von-Neumann architecture, an architecture that by its very success has stifled the development of alternative models of computation. It became influential after World War II, and has only grown more ubiquitous over the years since. The alternatives remain relatively unexplored: using hardware based on Neural Networks, Genetic Algorithms, Quantum Computing, and massively parallel systems would all provide machines that behave in a markedly different manner than the computers of today.

Theoretical models that try to make human culture intelligible for machines to understand, Computational Linguistics and the like, are almost always designing for the Von-Neumann machines. Don't trust a theory that accepts the narrow, rational Von-Neumann architecture as the foundation for all formal systems.


Photo_68
 
Photo_68
Neural Networks, although alternately championed and vilified, seem to be a promising avenue for continued research. The basic premise is that instead of writing complex rules to describe complex behavior, we should instead be simulating thousands of tiny, simple components that are designed to function as neurons.

Then the methodology takes a turn for the worse: You then train the neurons to produce a desired output, given a particular input. After a period of training, the neural nets can effectively recognize patterns that would cause errors in sophisticated algorithms. The interesting thing about the functioning of these networks, is that even with extremely low numbers of neurons (say around 20), the methods that they use to produce the results are more complicated than human minds can understand. The networks quickly develop emergent behavior that cannot be modeled by conventional means.

But training a network of neurons to work as an input-output filter completely misses the point of neurons in the first place. They are an always-on, always-shifting network of  . . . (finish the thought)

The first neural networks were built out of components called perceptrons, and had less of a direct link to neuroscience. They were simple models which produced an output as a function of a number of inputs of varying weights. A somewhat-hostile book, published by the A.I. luminaries Marvin Minsky and Seymour Papert, gave the impression to many researchers that perceptrons are incapable of being computationally interesting. Because of the respect that Minsky and Papert carried in the A.I. community, research on neural networks was effectively abandoned for 15 years, and is only now emerging with strength once again.