You can ignore 90% of the machine learning literature, treat the whole thing as a craft rather than a science and use OPAL/PA-II or an averaged perceptron (which are closely related with proven 50s and 60s machine learning technology, but are very well-understood these days), with regularization.
It doesn't necessarily take a PhD to realize that, but fortunately (for me) you'll need a PhD to convincingly sell that point to non-machine learning people.
You can ignore 90% of the machine learning literature, treat the whole thing as a craft rather than a science and use OPAL/PA-II or an averaged perceptron (which are closely related with proven 50s and 60s machine learning technology, but are very well-understood these days), with regularization.
It doesn't necessarily take a PhD to realize that, but fortunately (for me) you'll need a PhD to convincingly sell that point to non-machine learning people.