abbott@aerospace.aero.org (Russell J. Abbott) (03/17/90)
It has been pointed out to me that in my previous note on meta-nets I didn't make it clear that the intent was to train the meta-net on a number of examples of different categories. For each set of categories the meta-net would be expected to privide the correct weights for those categories. After training, if it is sucessful, the meta-net would be able to take as input a new set of category examples and provide weights for that set of categories. I realize that that sounds rather preposterous, but since weight training (excuse me) is so similar to the sort of things neural nets are used for, why can't it be done? Finding correct weights is a (highly parallel) process of finding a minimum in a space defined by the sort of continuous mappings that neural nets generally do. -- -- Russ abbott@itro3.aero.org