Clever Geek Handbook

Effective Machine Learning Queries

I am reading ML text which has a question:

If we have a supervisor who can provide us with a shortcut to any x , where should we choose x for training with fewer requests?

So I interpret this as:

We ask you to help speed up the learning process by selecting test items that provide the most discriminatory information.

I would suggest it would be an iterative algorithm: the selection of subsequent test items depends on which items were preselected. It's not clear what the question is:

  • Whether subsequent items are selected will also depend on the results of previous queries.

So my question here is, is there some kind of information-theoretic unique answer to the above? Or are there multiple approaches that can have the same efficiency / speed?

** Footnote: this is not homework - this is my independent learning

+3


source to share


2 answers


There are several approaches. In fact, this parameter is a machine learning task called Active Learning . There are many methods for selecting x based on different strategies - see Section 3 of this overview document .



+1


source


In general, most active learning methods (also known as Query learning or "Best Experimental Design") will use all observations, including the results or results of previous queries. Since you want to know the relationship between x and the result, both pieces of information are helpful in determining how to choose the optimal value for x .



0


source






More articles:


All Articles