emerged as an effective and promising approach

to many machine learning tasks including

computer vision, language understanding,

or classification in general.

In this video series, we are going

to introduce a new learning framework

to you called neural structured learning, which

enables neural nets to learn with structured signals

for improving model quality and robustness.

I am Da-Cheng, and I'm going to be your guide.

You do not need to know a lot to get started,

and we will be coding with Python language.

Don't worry you have never used it.

It's simple to understand, and you will

be up and running in no time.

So let's get started with a simple example.

Consider you are creating a neural network to classify

an image into a cat or a dog.

Like this figure shows, the image

is fed into the neural net, activating neurons layer

by layer, forming several activation

path that determine this image to be

classified as a cat or a dog.

Seems pretty straightforward, isn't it?

What if I tell you there are other similar images related

to this input image?

That is, there there's actually a structure,

for example, a graph representing the similarity

among all these images.

And as you can see, all these images are English bulldogs.

So is it possible that we can make a neural net learn better

with the whole structure in addition

to just using one image?

And the answer is yes through neural

structured learning framework.

Neural structured learning jointly

optimizes simple features and the structured signals

existed among samples in order to learn a better neural net.

Specifically, we now have two types

of input for a neural net.

The first input is the features of a training sample,

for example, the pixels of an image.

And a second input is the structure,

for example, the graph representing

the similarity among samples.

Both the features and the structure

will be fed into a neural net for training.

You may now have a question.

We know in a neural net the input features

are used to activate the neurons layer by layer

for making a classification.

But how do we use the structure to help a neuron net learn?

The structure is used to regularize the training

of a neural network.

Don't worry if you are not familiar with this concept.

We are going to provide more details to you

about this whole training process.

Are you ready?

First, each training sample is augmented

to include its neighbor information from a given

structure, specifically the neighbor information here

refers to the features of a neighbor.

So we get a new training batch where

both the original training samples and their neighbors

are included.

Next, both the training sample and its neighbors

are fed into the neural net.

After the training sample is fed into the neural net,

its features activate different neurons layer by layer,

forming and embedding representation for this sample.

If you're not familiar with the concept of embedding layers,

just think of it as a new representation

formed by the second to last layer of a neural net.

The neighbor is processed in the same way.

So we will have an embedding representation for the neighbor

as well.

Then the difference between the samples embedding

and its neighbors embedding is calculated and added

into the final loss as a regularization term.

So what is the intuition here?

By adding this regularization term,

the neural network learns to keep

the similarity between a sample and its neighbor.

In other words, the neural net learns

to maintain the local structure of a sample

and its neighborhood.

By leveraging these structured signals,

neural nets can learn with last label data

and also be more robust.

We also provide several hands-on tutorial

to guide you step by step how to use the neural structured

learning framework.

In a next video of these series, we

will take what you have learned and apply that

to a language understanding problem,

classifying the topic of a document.

You will find a tutorial for that in the description below

as well as more information on getting

started with neural structured learning.