| description |
|---|
This theorem basically provides a framework for learning algorithms and heuristics. There are no any heuristics or algorithms that are best at solving both the generic case and specific case of any problem. These generic algorithm are beaten by the performance of the specific ones in specific cases.
Say, there is a whole bunch of problems from the real world in a bowl. These problems are represented by a function and all these functions are learnable. And if we draw a uniformly from this bowl, and try to learn the functions, none of the learning algorithm will perform really good on all kinds of problems. Some algorithm are better at other in some specific use cases.
This theorem brings us to the world of speed, accuracy, bias, variance complexity world. Keep in mind that a model could be trained by multiple algorithms but still this holds true.