You are on page 1of 2

Now, common to all of these tasks

is that a, we can learn to solve them,


but b, they would be very hard to solve
in a traditional engineering way where
we specify rules or solutions directly to this problem.
How do I specify the solution to the problem of whether you
like this movie or not?
This is much easier to learn from experience.
Let's be a little bit more concrete.
So the machine learning tasks are really
provided in terms of examples.
It's very easy to show examples of how
you would want the method to behave,
examples of correct behavior.
I can take an image.
I can provide an annotated category
for that image, what I want the method to predict.
Another image, corresponding category, and so on.
And that will generate a large set of examples that
illustrate what the task is.
This is typically very easy to do.
OK.
I don't have to specify the solution
to be able to illustrate the task implicitly
through examples.
The power of machine learning really
is now in terms of automating the process
of finding the solution to a task that's
illustrated by example.
OK.
So what we do is we hypothesize a set of possible mappings--
a large set that takes an image and [? produces ?] a category
as an answer.
What that mapping is is specified by parameters.

A different value for these parameters


will specify a different mapping.
And I would like to find, out of the possible parameters,
the one that agrees with my set of examples-- the best.

We will automate this process of finding these parameters,


and also specifying what the mappings are.
So this is what the machine learning algorithms do for you.

You can apply the same approach in different contexts.


For example, in machine translation,
you can easily specify what the task is
by giving an English sentence and a corresponding Spanish
sentence, and have a large number of them
to illustrate what the task is.
Then, specify a complex mapping that
can take a sentence as an input, has some adjustable parameters
that specify different possible mappings, each parameter
value providing a possible answer that may or may not
agree with the correct one.
And we search for those parameter
values in an automated way in order
to find the best mapping for this task.
Google Translate already operates in this mode.

You might also like