Discussion of Lecture 3 "The Linear Model I"

[vikasatkin]

Question: If I have a nice feature of digits in mind, how do I decide if I should include it into the set of variables I feed into my learning algorithm or not?

Answer: If you add a random variable, you don't improve the result, but you pay for it, because you increase the amount of variables. In each particular machine learning task you know the size of your data set and there is a way to determine the number of variables you can safely feed into the learning algorithm, we will be talking about it later.

In this case (digits classification) if you invent features by hand, chances are you will exhaust your imagination much earlier.

If you want to play with the digits classification data, you can find it at the bottom of the page http://amlbook.com/support.html.
Lecture slides are available at http://amlbook.com/slides/.

[vikasatkin]

Question: On the slide 6/23 why do we have flat regions on the graph?

Answer: Chances are, that during this flat regions, we don't change the classification of any point. We may change the coefficients on some steps, but most of the time we are just looking at the points, which are classified correctly, so we even don't change any coefficients (and probably we count these steps as iterations too and, therefore, getting flat regions in the graph of and for hundreds of iterations).

[yaser]

Originally Posted by vikasatkin

Question: On the slide 6/23 why do we have flat regions on the graph?

Answer: Chances are, that during this flat regions, we don't change the classification of any point. We may change the coefficients on some steps, but most of the time we are just looking at the points, which are classified correctly, so we even don't change any coefficients (and probably we count these steps as iterations too and, therefore, getting flat regions in the graph of and for hundreds of iterations).

Indeed, this is how the plot in this slide was created. It would have looked different if we had counted only the iterations involving a misclassified point (hence a genuine update to the weight vector).