Follow US:

Practice English Speaking&Listening with: Neural Structured Learning - Part 3: Training with synthesized graphs

Normal
(0)
Difficulty: 0

[MUSIC PLAYING]

ARJUN GOPALAN: Hi, I'm Arjun Gopalan,

and welcome to episode 3 of Neural Structured Learning.

In the previous episode, we learned about natural graphs

and how they can be used in neural structured learning.

While natural graphs are common, there

are many machine learning tasks where the input data does not

form a natural graph.

For instance, if you recall the document classification task

from the previous episode, we used citations

to form a natural graph.

In the absence of citations, we wouldn't have had a graph.

Similarly, if you're doing simple image classification

or text classification where the input data contains

just raw images or text, then we may not

have a natural graph in either case.

In this episode, we'll discuss how

we can apply neutral structured learning to such tasks.

So what do we do if we don't have a natural graph

to begin with?

The title of the video might have given this away.

But the main idea is to build or synthesize

a graph from the input data.

Building a graph can be done in many ways.

But in this video, we'll use the notion

of similarity between instances to build a graph.

In order to define a similarity metric,

we need to convert raw instances,

whether they are documents, text, or images,

to corresponding embeddings or dense representations.

We can do this using pretrained embedding models, such as those

on TensorFlow Hub.

Once we convert raw instances to their embeddings,

we can use a similarity function,

such as the cosine similarity, to compare

how similar pairs of embeddings are.

If the similarity score is greater than the threshold,

then we add a corresponding edge in the resulting graph.

Repeating this process to cover the entire data set

builds a graph.

And once we have a graph, using neutral structured learning

is straightforward, as we saw in the previous episode.

Let's illustrate this workflow for the task of sentiment

classification using the IMDb data set.

This data set contains movie reviews.

And the task is to classify them as good or bad.

Let's see what the code looks like to build

a neural structured learning model for this task.

Here again, we use Keras for illustration.

But neutral structured learning also supports estimators.

The first step is to load the IMDb data set.

For simplicity, we use a version of it that is part of Keras.

Once that is done, we want to convert

the raw text in the movie reviews to embeddings.

We use swivel embeddings in this example.

But any other embedding model, such as Bert, Word2Vec,

or an embedding model of your choice

may also be used instead.

Once we have created the embeddings,

we can build a graph using those embeddings.

Neural structured learning provides an API

called build graph to do so.

Notice that it accepts the similarity threshold

as one of its arguments.

This allows you to control the threshold

below which edges are dropped from the resulting graph.

In this example, we use a threshold of 0.8.

Once we have the graph, we define the features of interest

for our model and combine these features

with the graph using the partNeighbours API

in neural structured learning.

In this example, we use a maximum of three neighbors

to augment our training data.

Now that we have the augmented training data,

the next step is to create a graph regularized model.

This part is similar to what we did in the previous episode.

First, we define a base model, which

can be any type of Keras model, whether it's

a sequential model, a functional API-based model,

or a subclass model.

It can also have an arbitrary architecture.

Then we define a graph regularization configuration

object, which allows you to specify

various hyperparameters.

In this example, we use three neighbors

for graph regularization.

Once this configuration object is created,

you can draft the base model with the graph regularization

wrapper class.

This will create a new graph Keras

model whose training includes a graph regularization term.

What's left is then just compiling, training,

and evaluating the graph regularized model.

This example is also available as a colab-based tutorial

on our website.

You can find that link in the description below.

In summary, we looked at how to build a graph regularized

model when the input data does not form a natural graph.

This technique can be applied to all kinds of input data,

such as text, images, and videos.

Now, graph building is not the only approach

to handle input data that does not form a natural graph.

In fact, in the next video, you will

learn about another aspect of neural structured

learning called adversarial learning, which

can be very useful to improve a model's robustness

to adversarial attacks.

That's it for this video.

There's more information in the description below.

And before you get to the next video,

don't forget to hit the Subscribe button.

Thank you.

[MUSIC PLAYING]

The Description of Neural Structured Learning - Part 3: Training with synthesized graphs