video_id
stringclasses
7 values
text
stringlengths
2
29.3k
rLlZpnT02ZU
And if you've heard of KL, you've
rLlZpnT02ZU
probably heard of entropy.
rLlZpnT02ZU
And that's what-- it's basically minus the entropy.
rLlZpnT02ZU
And that's a quantity that just depends on theta star.
rLlZpnT02ZU
But it's just the number.
rLlZpnT02ZU
I could compute this number if I told
rLlZpnT02ZU
you this is n theta star 1.
rLlZpnT02ZU
You could compute this.
rLlZpnT02ZU
So now I'm going to try to minimize
rLlZpnT02ZU
the estimate of this function.
rLlZpnT02ZU
And minimizing a function or a function plus a constant
rLlZpnT02ZU
is the same thing.
rLlZpnT02ZU
I'm just shifting the function here or here,
rLlZpnT02ZU
but it's the same minimizer.
rLlZpnT02ZU
OK.
rLlZpnT02ZU
So the function that maps theta to KL of P theta star
rLlZpnT02ZU
to P theta is of the form constant minus this expectation
rLlZpnT02ZU
of a log of P theta.
rLlZpnT02ZU
Everybody agrees?
rLlZpnT02ZU
Are there any questions about this?
rLlZpnT02ZU
Are there any remarks, including I
rLlZpnT02ZU
have no idea what's happening right now?
rLlZpnT02ZU
OK.
rLlZpnT02ZU
We're good?
rLlZpnT02ZU
Yeah.
rLlZpnT02ZU
AUDIENCE: So when you're actually employing this method,
rLlZpnT02ZU
how do you know which theta to use as theta star and which
rLlZpnT02ZU
isn't?
rLlZpnT02ZU
PHILIPPE RIGOLLET: So this is not a method just yet, right?
rLlZpnT02ZU
I'm just describing to you what the KL divergence
rLlZpnT02ZU
between two distributions is.
rLlZpnT02ZU
If you really wanted to compute it,
rLlZpnT02ZU
you would need to know what P theta star is
rLlZpnT02ZU
and what P theta is.
rLlZpnT02ZU
AUDIENCE: Right.
rLlZpnT02ZU
PHILIPPE RIGOLLET: And so here, I'm just saying at some point,
rLlZpnT02ZU
we still-- so here, you see--
rLlZpnT02ZU
so now let's move onto one step.
rLlZpnT02ZU
I don't know expectation of theta star.
rLlZpnT02ZU
But I have data that comes from distribution P theta star.
rLlZpnT02ZU
So the expectation by the law of large numbers
rLlZpnT02ZU
should be close to the average.
rLlZpnT02ZU
And so what I'm doing is I'm replacing any--
rLlZpnT02ZU
I can actually-- this is a very standard estimation method.
rLlZpnT02ZU
You write something as an expectation with respect
rLlZpnT02ZU
to the data-generating process of some function.
rLlZpnT02ZU
And then you replace this by the average of this function.
rLlZpnT02ZU
And the law of large numbers tells me
rLlZpnT02ZU
that those two quantities should actually be close.
rLlZpnT02ZU
Now, it doesn't mean that's going to be the end of the day,
rLlZpnT02ZU
right.
rLlZpnT02ZU
When we did Xn bar, that was the end of the day.
rLlZpnT02ZU
We had an expectation.
rLlZpnT02ZU
We replaced it by an average.
rLlZpnT02ZU
And then we were gone.
rLlZpnT02ZU
But here, we still have to do something,
rLlZpnT02ZU
because this is not telling me what theta is.
rLlZpnT02ZU
Now I still have to minimize this average.
rLlZpnT02ZU
So this is now my candidate estimator for KL, KL hat.
rLlZpnT02ZU
And that's the one where I said, well, it's
rLlZpnT02ZU
going to be of the form of constant.
rLlZpnT02ZU
And this constant, I don't know.
rLlZpnT02ZU
You're right.
rLlZpnT02ZU
I have no idea what this constant is.
rLlZpnT02ZU
It depends on P theta star.
rLlZpnT02ZU
But then I have minus something that I can completely compute.
rLlZpnT02ZU
If you give me data and theta, I can compute this entire thing.
rLlZpnT02ZU
And now what I claim is that the minimizer of f or f plus--
rLlZpnT02ZU
f of X or f of X plus 4 are the same thing,
rLlZpnT02ZU
or say 4 plus f of X. I'm just shifting
rLlZpnT02ZU
the plot of my function up and down,
rLlZpnT02ZU
but the minimizer stays exactly where it is.
rLlZpnT02ZU
If I have a function--
rLlZpnT02ZU
so now I have a function of theta.
rLlZpnT02ZU
This is KL hat of P theta star, P theta.
rLlZpnT02ZU
And it's of the form-- it's a function like this.
rLlZpnT02ZU
I don't know where this function is.
rLlZpnT02ZU
It might very well be this function or this function.
rLlZpnT02ZU
Every time it's a translation on the y-axis of all these guys.
rLlZpnT02ZU
And the value that I translated by depends on theta star.
rLlZpnT02ZU
I don't know what it is.
rLlZpnT02ZU
But what I claim is that the minimizer is always this guy,
rLlZpnT02ZU
regardless of what the value is.
rLlZpnT02ZU
OK?
rLlZpnT02ZU
So when I say constant, it's a constant with respect to theta.
rLlZpnT02ZU
It's an unknown constant.
rLlZpnT02ZU
But it's with respect to theta, so without loss of generality,
rLlZpnT02ZU
I can assume that this constant is 0 for my purposes,
rLlZpnT02ZU
or 25 if you prefer.
rLlZpnT02ZU
All right.
rLlZpnT02ZU
So we'll just keep going on this property next time.
rLlZpnT02ZU
And we'll see how from here we can move on to--
rLlZpnT02ZU
the likelihood is actually going to come out of this formula.
rLlZpnT02ZU
Thanks.