ksr1492@cec1.wustl.edu (Kevin Scott Ruland) (02/14/90)
I heard that Wasserman had tried training feedforward nets by backprop with a random (cauchy, I think) vector added to the weights. I saw a single page report from a proceedings that reported Wasserman had tried this with some success but failed to list numerical results. I had tried this training on a 3-4-1 net to do the 3-d xor problem with some good convergance results (approx. 95% of all nets trained in this way converged compared to <15% when trained without the added noise). If anyone has done some of this, or knows of some references please drop me a line. kevin kevin@rodin.wustl.edu