1. Paul Werbos had talked about error back propagation in his doctoral thesis "Beyond regression: new tools for prediction and analysis in behavioral sciences" (1974) before Rumelhart et al (1986).
2.
A basic problem of statistical pattern recognition, viz., the dimensionality -sample size problem also arises in artificial neural systems. In the design of multilayer feedforward networks one question is how many hidden units to use.A few techniques are reviewed - but they appear an over-kill (in the best case) and clearly inapplicable (in the worst case) because in my experiments the size is not crucial - the order of free parameters in the network remains constant. there is only one hidden layer - and no. of hidden units are of order of 1 to 10.
3.
While the generation of artificial neural networks excite us, we should keep in mind that:4.
(1) As has been shown by [Comparing hierarchical statistical classifiers with error back propagation neural network; Kanal et al (1989)], often fairly simple statistical decision tree methods give equivalent or better results;
(2) the various neural network paradigms for pattern classification introduced in recent years have close connections with stochastic approximation, estimation and classification procedures known in statistical pattern recognition; and
(3) rather good algorithms have been developed in recent years for large combinatorial optimization problems whereas neural networks have so far only been demonstrated on much smaller problems. It remains to be shown that combinatorial optimization is a good area for artificial neural networks.
...the problem of scalability remains one of the basic concerns for employing various pattern recognition, parallel processing, and machine intelligence tools on real world problems.5. "They were AI as long as it was unclear how to make them work." After a very interesting discussion on what is AI, based on AI Magazine, Roger Shank (1991)
6. Theorem of the Ugly Duckling
(by Watanabe)
separate post.
No comments:
Post a Comment