4.8. The Cost Function Choice
It will not come as a surprise that the least squares cost function in (4.6) is not the unique choice available to the user. Depending on the specific problem, other cost functions can lead to better results. Let us look, for example, at the least squares cost function more carefully. Since all errors in the output nodes are first squared and summed up, large error values influence the learning process much more than the small errors. Thus, if the dynamic ranges of the desired outputs are not all of the same order, the least squares criterion will result in weights that have “learned” via a process of unfair provision of information. Furthermore, in [Witt 00] it is shown that for a class of problems, the gradient ...
Get Pattern Recognition, 4th Edition now with the O’Reilly learning platform.
O’Reilly members experience books, live events, courses curated by job role, and more from O’Reilly and nearly 200 top publishers.
