PLA computing speed

[jlaurentum]

Got it. for the inner product (in R) I was using sum(w*x). I changed that to w%*%x and the speed changed DRAMATICALLY. Misterios arcanos del R!

[chiraz]

Originally Posted by Elroch

Good point about sapply!

In answer to your question, I have observed empirically that the way you choose the misclassified point has quite a large effect on the convergence. As an attempt to improve on randomisation, one could certainly attempt to use the information in the value of w.x when sign(w.x) is wrong. But should we prefer large modulus of w.x or small?

About choice of misclassified point, the homework specifically asks to choose it arbitrarily. Aside from that, I'm not sure if a more informed choice would lead to faster convergence..

[chiraz]

I also want to add that I am using Matlab and the whole experiment always takes only a few seconds (both for N=10 and N=100).

[Elroch]

Originally Posted by jlaurentum

Got it. for the inner product (in R) I was using sum(w*x). I changed that to w%*%x and the speed changed DRAMATICALLY. Misterios arcanos del R!

This is curious. I have found in benchmark tests that with 3 dimensional vectors w and x, the times for sum(w*x) and w%*%x are indistinguishable. Could you post a fragment of test code where you can detect a significant difference?

[chiraz, yes, it should be noted that attempts to improve convergence are a separate matter from the homework]

[IsidroHidalgo]

I select randomly a missclassified point, but I don't think that this improves speed so much, but the use of inner product maybe (using 10000 points to calculate the probability of error between f(x) and g(x) can lead to important improvement)

[brian.holt]

I'm glad to hear it took some time for your iterations. I got to a point where I thought my code wasn't working, so I kept stopping it. Needless to say, my scores on the last 3 questions reflected my lack of confidence in my code.

[clarkfitzg]

Choosing a random misclassified point versus the first misclassified point made a significant difference for me when N = 100.

Check this thread.