Perhaps my perception of time is augmented by the faster machine speeds these days, but I was wondering if there was a form of machine learning that takes longer but will yield drastically better results on large datasets with lots of noise. I am assuming here that faster convergence somehow has a relationship with the likelihood of becoming stuck in a local extrema. I notice little fluctuation after 100-500 epochs at which point I simply have to restart. I am currently using a feed-forward neural network for both regression and classification.
I suppose genetic algorithms seem to be the most time intensive types of "brute force" machine learning. I was also thinking that other types of neural networks could be modified (such as their momentum or learning rate to increase its range over the function). Obviously, I have tried adjusting both of these but this has not solved my problem.