0
$\begingroup$

Hello I'm studying machine learning processes and I'm beside of a misunderstanding..

Is this right?

"Minimization is a process that minimize the error rate of Y (output of the feature) to be a valid limiter and this is followed by an optimization process that acts on the parameters to find the best ones for the best model to choose"

The only parameters I know in my head now are the FEATURES (X inputs). Which are the parameters to optimize? Technical parameters?

Thanks for help!

Improve this question
$\endgroup$
6
  • $\begingroup$ Where did this quote come from? It doesn't make much sense to me either. $\endgroup$
    – gung - Reinstate Monica
    Commented Jul 14, 2015 at 14:26
  • $\begingroup$ @gung Damn.. that was what I realized. Can you help me to undestand please? :) $\endgroup$
    – kamauz
    Commented Jul 14, 2015 at 14:33
  • $\begingroup$ Was this quotation originally rendered in a language other than English and then translated? $\endgroup$
    – Sycorax ♦
    Commented Jul 14, 2015 at 14:35
  • $\begingroup$ Think about a “multi-purpose box” waiting for input and producing the corre- sponding output depending on operations influenced by internal parameters. The information to be used for “customizing” the box has to be extracted from the given training examples. The basic idea is to fix the free parameters by demanding that the learned model works correctly on the examples in the training set. Now, we are all believers in the power of optimization: we start by defining an error measure to be minimized1, and we adopt an appropriate (automated) opti- mization process to determine optimal parameters. $\endgroup$
    – kamauz
    Commented Jul 14, 2015 at 14:42
  • $\begingroup$ what kind of parameters? features or other parameters? Excuse me but I guess I'm really confused $\endgroup$
    – kamauz
    Commented Jul 14, 2015 at 14:43

2 Answers 2

Reset to default
0
$\begingroup$

Let me explain that for you.

Every model has some unknowns that are called parameters. For example, if $X_1,\ldots, X_n$ are some features then a simple model can be like $Y=b_1X_1+\ldots+b_nX_n +e$ where $e$ is the error that is bounded (simply $var(e)<\infty$). Then the goal of optimization is to find (or estimate) a set of parameters, $b_1,\ldots,b_n$ that minimizes a certain criteria, like $\sum e^2$. If you have a set of measurements from the features, then using optimization techniques you can estimate parameters numerically.

PS: parameters can be coefficients, distribution unknown (like mean, var ...), risks and .... . Thus a parameter is a part of model and it is not something individual.

Improve this answer
$\endgroup$
2
  • $\begingroup$ Ok.. so minimization should minimize the "e" value to the maximum error rate accepted? $\endgroup$
    – kamauz
    Commented Jul 14, 2015 at 15:08
  • $\begingroup$ minimum error rate, yes. Note that $e=y-b_1X_1-\ldots-b_nX_n$ $\endgroup$
    – TPArrow
    Commented Jul 14, 2015 at 15:12
0
$\begingroup$

You can see your machine learning solution as a black box. The black box has a number of knobs which you can easily tune. The performance of your solution depend how these knobs are set. You can set them at random, you can set them manually or let an optimization algorithm to set them. In most cases you want the latter.

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.