kortge@elaine2.Stanford.EDU (Chris Kortge) (04/14/91)
I've been working on a neural net learning algorithm which extracts principal components from a sequence of input patterns (as backprop and other NN algorithms can do). It requires, say, a third as many pattern presentations as optimized versions of other NN algorithms, but this is at a cost of requiring an order of magnitude more computation per pattern. (It learns faster by reducing interference with old knowledge when a new pattern is learned; I can say more if anyone's interested.) I've been working under the assumption that there will be situations where it's good to make this tradeoff of within-pattern processing time for number of patterns required; one might guess, for example, that a creature running around in the world can do a lot of processing before the current "pattern" changes to a significant degree. But I don't know of any *existing applications* that require this sort of algorithm. Does anyone else know of any? It basically just needs to be a situation where there is some significant limitation on the rate at which inputs can be observed (e.g. the inputs aren't all stored in RAM ahead of time). A possibility might be devices which must adapt to each new user, but I don't know of any specific instances of this that would fit the bill. It would really help in writing this up if I had something in the real world to relate it to. Thanks for the help, Chris Kortge kortge@psych.stanford.edu