An iterative method for training multilayer networks with threshold functions

摘要: Concerns the problem of finding weights for feed-forward networks in which threshold functions replace more common logistic node output function. The advantage such is that complexity hardware implementation greatly reduced. If task to be learned does not change over time, it may sufficient find correct a function network off-line and transfer these implementation. This paper provides mathematical foundation training with standard nodes gradually altering allow mapping unit network. procedure analogous taking limit as gain parameter goes infinity. It demonstrated that, if error trained small, small will cause error. result must implemented can first using traditional back propagation gradient descent, further progressively steeper functions. In theory, this process could require many repetitions. simulations, however, have successfully mapped true after modest number slope changes. important emphasize method only applicable situations learning appropriate. >