P[|error(h) - beta| > epsilon] < delta

What you showed in class is that:

P[error(h(ERM)) - beta > epsilon] < 2|H|e^{-m*epsilon^2/2}

Now you are given delta and m, and you want to find epsilon such that the RHS is less than delta, which leads to: epsilon> sqrt{2ln(2H/delta)/m}.

So as long as you take epsilon like that the inequality will hold. In particular it will hold for epsilon = sqrt{2ln(2H/delta)/m}.

Hope this is clearer.

Amir

How do we get "2ln… Close" from that? What am I miss? ]]>

See section 2.4.2 of the scribe for lecture 2, which proves this result, and gives the sample size needed for learning with epsilon,delta.

Best,

Amir ]]>

At the last lecture pdf there is a generalization bound of ERM. As I remmember, the bounds we saw were for detemine the needed sample size. How do we get that bound? Where is its proof/explaination?

Yhanx!

]]>