|
I just had a reviewer blind-side me with a question/point about empirical risk minimization and I feel kind of silly for it. I've been doing a little bit of reading and searching and have hit a brick wall of sorts. Basically I have two questions for the group: 1) Can anyone point me to a good recent resource with state-of-the-art ERM results? 2) Is ERM something that is actually used (i.e. people have developed algorithms for) or is it something that is just theorized about? Any help would be greatly appreciated. |
|
Most machine learning algorithms in practice are variations of structural risk minimization, which is empirical risk minimization with added regularization. Even for models that are not explicitly formulated as such one can usually find some empirical/structural risk that is minimized either directly or approximately by the model. So for you question (1) I'd be really hard-pressed to point you to one example of state-of-the-art ERM models, but look at SVMs, neural networks, online linear machine learning, bayesian methods, etc, and you will almost always find some notion of risk that is minimized. For (2), yes, people actually use ERM and SRM, both directly and indirectly. It could be that I didn't understand your question, also. |
