Thomas G. Dietterich's "Statistical Tests for Comparing Supervised Classification Learning Algorithms" compares five tests of learning algorithms. He concludes that McNemar's test is best for algorithms executed only once (e.g., because of time or cost constraints); a new 5x2cv test of 5 iterations of 2-fold cross-validation is slightly more powerful for repeated trainings. . [, connectionists, 10/16/96.]

Don't forget the comp.ai.neural-nets FAQ, with seven sections covering Intro, Learning, Generalization, Books (& data), Free Software, Commercial Software, and Hardware. , or browse /pub/usenet/news.answers/ai-faq/neural-nets on rtfm.mit.edu. [Warren Sarle , comp.ai.neural-nets, 11/22/96.]

Bob Massey offers an interactive NN demo based on maximizing entropy. . [, comp.theory, 11/22/96.]

Here's a useful application: NeuroDiet 1.0 asks you to enter the foods you eat each day and any medical symptoms you experience. It trains a neural network and can then advise you on how your health depends on the foods you eat. (Are you getting headaches from eating peanuts? Red wine? Aged cheese? Something like this could also track weather factors and pollen-related allergies.) NeuroDiet is Windows 95 shareware from Stephen Wolstenholme , on or . [comp.archives.ms-windows.announce, 10/9/96.]

-- Ken