HENG-TZE CHENG: Wide and Deep learning

combines the power of memorization

and generalization, and it does that

by jointly training widening your models

and deepening your networks.

We're sharing a research paper about it

and also the implementation with an easy-to-use API

in TensorFlow, which is an open source library for a machine

intelligence.

So you might wonder what Wide and Deep learning is good for.

Wide and Deep learning is useful for generic large scale

regression and classification problems with sparse inputs,

things like recommendation systems, search,

and ranking problems.

Now imagine you wanted to build a search engine for food.

Given a query, you want to recommend the items

that your users will like the most.

Using widening your models, you can actually

use a wide set of cross-product features transformations

to memorize specific feature combinations.

An example would be when the users say

the query, "fried chicken," your model

might memorize that chicken and waffles

is more relevant than chicken fried rice.

But one limitation is that it's actually

hard to generalize to previously unseen combinations

without manual feature engineering.

So instead, using deep neuronetworks,

you can now generalize better through lower dimension

embeddings.

For example, your model might learn to recommend burgers

given the query, "fried chicken," because they

are similar types of food.

However, sometimes memorizing specific combinations

as rules and exceptions is very important.

When people ask for iced decaf latte,

you don't really want to overgeneralize and give them

hot latte no matter how close they

are in the embedding space.

So by jointly training Wide and Deep models,

we actually allow them to complement

each other's strengths and weaknesses.

MUSTAFA IPSIR: To help developers get started,

we released Wide and Deep as part of the TF Learn API.

TF Learn is a high level machine learning library

on top of TensorFlow.

The API helps users focus on the important questions

like, how will you design your features,

and what is your model structure?

You can create a Wide and Deep classifier

with just a few lines of code.

Then you specify the features you

use in the widening model and the deep neuronetworks,

and we handle the joint training under the hood.

There are different needs and requirements

from Deep learning networks and Wide [INAUDIBLE] models.

We found a way to balance this.

We provide a simple feature engineering interface

that lets you specify embeddings,

crosses, and bucketization easily.

For example, to learn the relationship

between a specific query and a specific item,

you can define across columns with a single line of code.

Similarly, to learn generalization,

you can define an embedding column

with a single line of code.

HENG-TZE CHENG: So to get started,

we encourage developers to check out our blog posts

in the description, which links to our tutorials, code samples,

and our research paper.

We really hope more and more people

will find these useful in their work

and explore the possibilities of Wide and Deep Learning with us.