And welcome to the Tensorflow developer Summit.

I'm Laurence Maroney, and I'm here at the Tensorflow Cafe.

And it's my great honor to speak with Mustapha is Spear.

You just did a talk at the summit on estimators and hiking and lots of different things.

And, you know, I was part of the rehearsals for his talk, and I could chat with him for hours about this stuff.

But we've only got a few minutes, so it's good that you have a couple of So I just like to ask you about some of the highlights of your hiking scenario.

Could you tell us about that first?

So you know, I want to make it an example so that practitioner can understand or feel.

Can all of these AP ice develop in the journey off Mission?

And I do like that.

You think of a machine learning as a journey, not a destination.

Yeah, it said journey, never ending journey.

And it's an exciting journey.

And the focus of your talk was really about estimators and about premade estimators.

Pyramid estimators care us have to combat this through these things.

I think the main lion can be summarized as experiment in your ideas with a couple of lines of code.

So how can you do experiment with your ideas with owning a couple of lines of code?

I like that.

So one of the things you also shared was just how much estimators air actually used here.

Yeah, it's, you know, a person.

I'm surprised to see how Wow, it's so popular here.

And it's my person have some of the projects to use these estimators and learn from them.

And the feedback we got from all these different use cases, it really happens like they see their pain or they're trying to do this.

But now this partitioning is not going well.

So that's why the training so slow.

So how can we fix it?

So we saw all those pains or like I worked on the team.

We were using missionary before to improve the product.

And that team, I realized experimenting with features is the one of the most important thing, right?

So that's why we all the patrols experience under they P s right, because one of the things in your talk is you were talking about feature columns.

Proper use of feature columns could have produced.

Could you give us some examples?

So, Yeah, that's the thing I mentioned in that project when I wasn't there.

Like, we always have ideas about features.

Oh, how can I use this feature?

For example, this hiking example.

Let's think about it.

Oh, I 10 use the zip got off, user if they provide in that web site, right, So I can use that, like, let's make it easy.

So this kind of stuff, uh, you were right.

A cz Well, as machine learning, we always think of like you have trees decisions as well.

You introduce this concept of radiant boosted trees?

Yeah, which I want T shirt with that.

My other estimator is a grating and boosted tree.

Could you tell us about the way you got the right of caste?

And based on a survey, take a survey like really popular centuries and, for example, is stable.

It's not random.

Like it was like one bunting with learning.

Every time you trained that mother, you may get a different mother because of the random initialization, but the city is that always provided the same.

So you can see they're a bunch of good stuff.

Rebuild heart, create this pyramid estimator so that it should not be difficult for people if they want to experiment with no net or trees.

But if you don't have remained estimator unit to switch to another pipeline, you want to create another input function like this stuff.

So we are solving it.

Sounds good.

I'm looking forward to trying them out.

It's one of the things that's it's on my very long to do this, but like you say, it just works.

Hopefully, I think you use that phrase several times.

I also said It's maybe, I mean, you need the experiment so you don't know experimentation.

You don't know which one really fits well into your problems.

There's a court from Norfolk you are saying very famous people.

Yeah is saying mission learning changes are like how we look at the engineering because normally you build something based on the bottom of the rules and you provide that a solution.

But in this case, it's it's becoming more natural science.

So you need the experiment.

So understand unit, analyze the output based on the statistical thinking, so it is different than like, Hey, this will work for you and I'm getting this thing and it will work for You know, we cannot say, you know, expand.

Right.

Okay, so becoming more more scientific experiment.

Evaluate the results.

Former hypothesis and then test life offices in the experiment again switching gears for a second.

When the things that you also talking about that really intrigued, man, I still need to get my head around his transfer.

Well, you could take, like, the learning from one graph and effectively chocolate changing to transfer to another.

This surprising that that simple on works.

So you can't like this is it just works.

Yeah, just look like Giant Network.

So the first time I was really surprised to I'm copying all these rivals from here to there and the verge.

How come So then I realized that network learns transformation fitting layer by layer transformation.

And we are leveraging that those transformations in my other necklace and my other because of retraining on top of here has an advantage of I can't decide have been they.

How can it uses the transformation on whether it helps or not?

So it can.

It has a room to decide that Z really cool It's just the possibilities of, like some experts in one area being able to train a model and experts in another area being able to train a different model and chopping changes transfer.

I don't give a big shot too, huh?

So they make it one line like that would be the one line you are accessing one of the best mother trained in Google.

This that's so powerful, it only seems to get So if I want to take advantage of this and I want things to just work for me, what recommendation would hopefully recommendation?

Would you give me to get started?

Transfer, learning, transfer, learning, Grady, Inclusive trees, estimators all of these things.

So we're putting examples under called model Garden official models.

Examples did you get so you can start with those examples because an example as data on some training parameters.

So then you start with that example, then you can put your own data into slow by slow.

So I think starting with an example is investigating.

Let's take a look at those in the models garden.

Yes, I like that.

But outside is I think they call official Mother's way.

Should change.

Yeah, go guard so much nicer.

You could have some great infused trees in your garden.

Yes, and it's not out yet.

Example.

But we had anything example.

Sounds good.

But thank you so much.

My staff.

This is a lot of things.

I want to have more hours to do this, but maybe do more video on.

Thanks, everybody, for watching this episode.

If you have any questions for me, if you have any questions for Mustapha, please just leave them in the comments below.

Don't forget to hit that subscribe button.

Thank you.