ANDREW HOH: Airbnb's an online marketplace.

We have over 5 million different homes

in 81,000 cities, which equates to hundreds of millions

of photos, making it possibly the largest

collection of images of homes in the world today.

When a guest decides to select a home,

one of the biggest influences in their decision

is a diverse set of images.

But a lot of times, hosts will take

a lot of pictures of a single room

and forget to take pictures of the other rooms.

SHIJING YAO: They also have the option to add captions,

but a lot of the cases, they're totally off.

ANDREW HOH: No.

ALFREDO LUQUE: We were faced with the challenge

of identifying what's actually in these pictures

and present them properly on the site.

That's one area where machine learning excels.

The real challenge was one of scale.

We had upwards of half a billion images to get through.

It was going to take months to really go through all of these.

ANDREW HOH: Using TensorFlow, we were

able to speed up the process and deliver

a reasonable model within days.

Bighead is Airbnb's machine learning platform.

We had the idea of making it very agnostic to different ML

frameworks, and so we levered TensorFlow to train the model.

And then Bighead helps with the model lifecycle, the feature

management, and then TensorFlow Serving

to help serve the model results.

SHIJING YAO: Before you're thinking

about which tool to use, you're first thinking

about which model to use.

And we did research on this.

We find that ResNet 50 was one of the state-of-the-art

performing models.

We used that as the basic architecture.

ALFREDO LUQUE: We used TensorFlow's cross APIs

and serving and some of the distributed GPU computations.

This ultimately led to a pipeline

that we could deploy to go through hundreds of millions

of images very quickly.

ANDREW HOH: So the end goal is basically

using these classifications of images

to make sure that their first, initial set of photos

that they see aren't just a picture of the garage

and bathroom.

But it could be of the living room that's gorgeous,

and the bedroom, and the swimming pool.

Future applications could be to detect different objects

in homes.

And if users decide to search on the website

for specific amenity types, we can actually bubble that up

to the surface.

SHIJING YAO: If you like how Airbnb operates today,

the reason was because of machine

learning because machine learning is almost everywhere

in the company.

Search ranking, pricing, predictive booking.

ALFREDO LUQUE: We're passing into the hundreds of models,

so it's something I expect to keep growing.

ANDREW HOH: And with a lot of these new frameworks coming

out, we can make better experiences

for our guests and better business decisions, as well.

[MUSIC PLAYING]