View Single Post
  #17  
Old 01-13-2013, 02:15 PM
magdon's Avatar
magdon magdon is offline
RPI
 
Join Date: Aug 2009
Location: Troy, NY, USA.
Posts: 595
Default Re: Exercises and Problems

It is true that white noise in the time domain is uniform in the frequency domain. But I do not quite follow how that relates to the prediction task, which in your example can be formulated as follows:

A target function f(x) is white noise. You are given f(x_n) on some data points x_n and asked to learn how to predict f on a test data point x_test. Your best prediction is 0 and your error will be distributed according to a Gaussian, no matter what your training data were, which is by definition of white noise. This would be your prediction even before you saw any training data, so you have not improved your prediction (i.e. you have not been able to learn) from the training data.

What bounds like Hoeffding tell you is that you will find out from your training data that your error will be large.

The same happens in classification where the function is a random \pm1. The data does not help you to predict outside the data, but at least you will learn from your data that you cannot predict outside the data. It is better than learning nothing .

Quote:
Originally Posted by cygnids View Post
With respect to the discussion on Prob 1.10 thread here, it is noted that learning is not possible when the target function is a random function. On that note, what crosses my mind is how does one reconcile that statement with the thought that a purely random function could have more characterizable features after transformation to other domains? For eg., white noise maps to constant in spectral domain. Perhaps naively, if I had the a set of images with a reasonable proportion of fully noisy images, and choose to apply the ML apparatus in the frequency domain, I could learn a system to classify noisy vs not-noisy? Right?

In learning, it is important to distinguish between how well you can predict outside the data, and how well you think you can predict outside the data. We only have access to the latter and the first task of the theory is to ensure that we have not fooled ourselves, so that the latter matches the former.

It's very likely I'm picking on a point out of context here?
__________________
Have faith in probability
Reply With Quote