Member-only story
Empirical Risk Minimizer — ERM
Machine Learning from First Principles: Blog Post 2
In the last post I described What is Pac Learning?
Today, I present to you the framework of a Machine Learning problem and the ERM.
Any ML problem would involve operating on some data. Feeding data to the machine and finding and discovering patterns is what we are after. But more so than that I want you to keep the following two things in mind:
1) If you perfectly knew the distribution from where the data came, you would not need Machine Learning
2) Machine Learning is not about predicting the class of a sample. It is about finding a prediction rule that would help you make that class prediction.
Even a simple if-else program can predict a class according to rules. The main idea behind ML is that you can find a hypothesis and apply it often to a distribution of data that you do not know fully and generalize it such that you are correct about your predictions more often.
Lets say that you are on an island where there are lots and lots of papayas. You want to come up with a hypothesis by which you can tell if a papaya is tasty or not by just looking or holding it. Of course you would need some experience to come with such an intuition. You would need to taste lots of papayas before you are…
