The problem as always with machine learning methods is figuring out how to construct your machine correctly and what data to feed it. Mostly this is about good

The problem as always with machine learning methods is figuring out how to construct your machine correctly and what data to feed it. Mostly this is about good data collection, not about what machine you use. I am somewhat agnostic on whether you train using a linear model, SVM, neural network, etc. - all of these are likely to be analytically useful, provided I can give them an interesting and well-contsructed input set. But if the data I'm feeding into such a system is too noisy, doesn't sample deeply enough to catalog the extent of variation, or we don't have good descriptors for correctly labeling the dataset, it's useless to try to train a machine, we'll just have GIGO.