Original subtitles

GWENDOLYN STRIPLING: Hello.

And welcome to Introduction to Generative AI.

My name is Dr. Gwendolyn Stripling.

And I am the artificial intelligence

technical curriculum developer here at Google Cloud.

In this course, you learn to define generative AI,

explain how generative AI works, describe generative AI model

types, and describe generative AI applications.

Generative AI is a type of artificial intelligence

technology that can produce various types of content,

including text, imagery, audio, and synthetic data.

But what is artificial intelligence?

Well, since we are going to explore

generative artificial intelligence,

let's provide a bit of context.

So two very common questions asked

are what is artificial intelligence

and what is the difference between AI and machine

learning.

One way to think about it is that AI is a discipline,

like physics for example.

AI is a branch of computer science

that deals with the creation of intelligence agents, which

are systems that can reason, and learn, and act autonomously.

Essentially, AI has to do with the theory and methods

to build machines that think and act like humans.

In this discipline, we have machine learning,

which is a subfield of AI.

It is a program or system that trains a model from input data.

That trained model can make useful predictions

from new or never before seen data

drawn from the same one used to train the model.

Machine learning gives the computer

the ability to learn without explicit programming.

Two of the most common classes of machine learning models

are unsupervised and supervised ML models.

The key difference between the two

is that, with supervised models, we have labels.

Labeled data is data that comes with a tag like a name, a type,

or a number.

Unlabeled data is data that comes with no tag.

This graph is an example of the problem

that a supervised model might try to solve.

For example, let's say you are the owner of a restaurant.

You have historical data of the bill amount

and how much different people tipped based on order type

and whether it was picked up or delivered.

In supervised learning, the model learns from past examples

to predict future values, in this case tips.

So here the model uses the total bill amount

to predict the future tip amount based on whether an order was

picked up or delivered.

This is an example of the problem

that an unsupervised model might try to solve.

So here you want to look at tenure and income

and then group or cluster employees

to see whether someone is on the fast track.

Unsupervised problems are all about discovery,

about looking at the raw data and seeing if it naturally

falls into groups.

Let's get a little deeper and show this graphically

as understanding these concepts are

the foundation for your understanding of generative AI.

In supervised learning, testing data values or x

are input into the model.

The model outputs a prediction and compares that prediction

to the training data used to train the model.

If the predicted test data values and actual training data

values are far apart, that's called error.

And the model tries to reduce this error

until the predicted and actual values are closer together.

This is a classic optimization problem.

Now that we've explored the difference

between artificial intelligence and machine learning,

and supervised and unsupervised learning,

let's briefly explore where deep learning

fits as a subset of machine learning methods.

While machine learning is a broad field that

encompasses many different techniques,

deep learning is a type of machine learning

that uses artificial neural networks,

allowing them to process more complex patterns than machine

learning.

Artificial neural networks are inspired by the human brain.

They are made up of many interconnected nodes or neurons

that can learn to perform tasks by processing data and making

predictions.

Deep learning models typically have many layers

of neurons, which allows them to learn

more complex patterns than traditional machine learning

models.

And neural networks can use both labeled and unlabeled data.

This is called semi-supervised learning.

In semi-supervised learning, a neural network

is trained on a small amount of labeled data

and a large amount of unlabeled data.

The labeled data helps the neural network

to learn the basic concepts of the task

while the unlabeled data helps the neural network

to generalize to new examples.

Now we finally get to where generative AI

fits into this AI discipline.

Gen AI is a subset of deep learning, which

means it uses artificial neural networks,

can process both labeled and unlabeled data using

supervised, unsupervised, and semi-supervised methods.

Large language models are also a subset of deep learning.

Deep learning models, or machine learning models in general,

can be divided into two types, generative and discriminative.

A discriminative model is a type of model

that is used to classify or predict labels for data points.

Discriminative models are typically

trained on a data set of labeled data points.

And they learn the relationship between the features

of the data points and the labels.

Once a discriminative model is trained,

it can be used to predict the label for new data points.

A generative model generates new data instances

based on a learned probability distribution of existing data.

Thus generative models generate new content.

Take this example here.

The discriminative model learns the conditional probability

distribution or the probability of y,

our output, given x, our input, that this is a dog

and classifies it as a dog and not a cat.

The generative model learns the joint probability distribution

or the probability of x and y and predicts

the conditional probability that this is a dog

and can then generate a picture of a dog.

So to summarize, generative models

can generate new data instances while discriminative models

discriminate between different kinds of data instances.

The top image shows a traditional machine

learning model which attempts to learn

the relationship between the data and the label,

or what you want to predict.

The bottom image shows a generative AI model

which attempts to learn patterns on content so that it

can generate new content.

A good way to distinguish what is gen AI and what is not

is shown in this illustration.

It is not gen AI when the output, or y, or label is

a number or a class, for example spam or not spam,

or a probability.

It is gen AI when the output is natural language, like speech

or text, an image or audio, for example.

Visualizing this mathematically would look like this.

If you haven't seen this for a while,

the y is equal to f of x equation calculates

the dependent output of a process given different inputs.

The y stands for the model output.

The f embodies the function used in the calculation.

And the x represents the input or inputs used for the formula.

So the model output is a function of all the inputs.

If the y is the number, like predicted sales,

it is not gen AI.

If y is a sentence, like define sales,

it is generative as the question would elicit a text response.

The response would be based on all the massive large data

the model was already trained on.

To summarize at a high level, the traditional, classical

supervised and unsupervised learning process

takes training code and label data to build a model.

Depending on the use case or problem,

the model can give you a prediction.

It can classify something or cluster something.

We use this example to show you how much more robust

the gen AI process is.

The gen AI process can take training code, label data,

and unlabeled data of all data types

and build a foundation model.

The foundation model can then generate new content.

For example, text, code, images, audio, video, et cetera.

We've come a long away from traditional programming

to neural networks to generative models.

In traditional programming, we used

to have to hard code the rules for distinguishing a cat--

the type, animal; legs, four; ears, two; fur, yes;

likes yarn and catnip.

In the wave of neural networks, we

could give the network pictures of cats and dogs

and ask is this a cat and it would predict a cat.

In the generative wave, we as users

can generate our own content, whether it

be text, images, audio, video, et cetera, for example

models like PaLM or Pathways Language Model,

or LAMBDA, Language Model for Dialogue Applications,

ingest very, very large data from the multiple sources

across the internet and build foundation language

models we can use simply by asking a question,

whether typing it into a prompt or verbally

talking into the prompt itself.

So when you ask it what's a cat, it

can give you everything it has learned about a cat.

Now we come to our formal definition.

What is generative AI?

Gen AI is a type of artificial intelligence

that creates new content based on what it has

learned from existing content.

The process of learning from existing content

is called training and results in the creation

of a statistical model when given a prompt.

AI uses the model to predict what an expected response might

be and this generates new content.

Essentially, it learns the underlying structure

of the data and can then generate

new samples that are similar to the data it was trained on.

As previously mentioned, a generative language model

can take what it has learned from the examples it's

been shown and create something entirely new

based on that information.

Large language models are one type of generative AI

since they generate novel combinations of text

in the form of natural sounding language.

A generative image model takes an image

as input and can output text, another image, or video.

For example, under the output text,

you can get visual question answering

while under output image, an image completion is generated.

And under output video, animation is generated.

A generative language model takes text as input

and can output more text, an image, audio, or decisions.

For example, under the output text,

question answering is generated.

And under output image, a video is generated.

We've stated that generative language models learn

about patterns and language through training data,

then, given some text, they predict what comes next.

Thus generative language models are pattern matching systems.

They learn about patterns based on the data you provide.

Here is an example.

Based on things it's learned from its training data,

it offers predictions of how to complete this sentence,

I'm making a sandwich with peanut butter and jelly.

Here is the same example using Bard,

which is trained on a massive amount of text data

and is able to communicate and generate

humanlike text in response to a wide range of prompts

and questions.

Here is another example.

The meaning of life is--

and Bart gives you a contextual answer

and then shows the highest probability response.

The power of generative AI comes from the use of transformers.

Transformers produced a 2018 revolution

in natural language processing.

At a high level, a transformer model

consists of an encoder and decoder.

The encoder encodes the input sequence

and passes it to the decoder, which

learns how to decode the representation

for a relevant task.

In transformers, hallucinations are words or phrases

that are generated by the model that

are often nonsensical or grammatically incorrect.

Hallucinations can be caused by a number of factors,

including the model is not trained on enough data,

or the model is trained on noisy or dirty data,

or the model is not given enough context,

or the model is not given enough constraints.

Hallucinations can be a problem for transformers

because they can make the output text difficult to understand.

They can also make the model more

likely to generate incorrect or misleading information.

A prompt is a short piece of text

that is given to the large language model as input.

And it can be used to control the output of the model

in a variety of ways.

Prompt design is the process of creating

a prompt that will generate the desired output

from a large language model.

As previously mentioned, gen AI depends a lot

on the training data that you have fed into it.

And it analyzes the patterns and structures of the input data

and thus learns.

But with access to a browser based prompt, you, the user,

can generate your own content.

We've shown illustrations of the types of input based upon data.

Here are the associated model types.

Text-to-text.

Text-to-text models take a natural language input

and produces a text output.

These models are trained to learn the mapping

between a pair of text, e.g.

for example, translation from one language to another.

Text-to-image.

Text-to-image models are trained on a large set of images,

each captioned with a short text description.

Diffusion is one method used to achieve this.

Text-to-video and text-to-3D.

Text-to-video models aim to generate a video representation

from text input.

The input text can be anything from a single sentence

to a full script.

And the output is a video that corresponds to the input text.

Similarly, text-to-3D models generate

three dimensional objects that correspond to a user's text

description.

For example, this can be used in games or other 3D worlds.

Text-to-task.

Text-to-task models are trained to perform a defined task

or action based on text input.

This task can be a wide range of actions

such as answering a question, performing a search,

making a prediction, or taking some sort of action.

For example, a text-to-task model

could be trained to navigate a web UI or make changes to a doc

through the GUI.

A foundation model is a large AI model pre-trained

on a vast quantity of data designed to be adapted or fine

tuned to a wide range of downstream tasks,

such as sentiment analysis, image captioning, and object

recognition.

Foundation models have the potential

to revolutionize many industries, including

health care, finance, and customer service.

They can be used to detect fraud and provide

personalized customer support.

Vertex AI offers a model garden that

includes foundation models.

The language foundation models include

PaLM API for chat and text.

The vision foundation models includes stable diffusion,

which has been shown to be effective at generating

high quality images from text descriptions.

Let's say you have a use case where

you need to gather sentiments about how your customers are

feeling about your product or service.

You can use the classification task sentiment analysis task

model for just that purpose.

And what if you needed to perform occupancy analytics?

There is a task model for your use case.

Shown here are gen AI applications.

Let's look at an example of code generation

shown in the second block under code at the top.

In this example, I've input a code file conversion problem,

converting from Python to JSON.

I use Bard.

And I insert into the prompt box the following.

I have a Pandas DataFrame with two columns, one with the file

name and one with the hour in which it is generated.

I'm trying to convert this into a JSON file

in the format shown onscreen.

Bard returns the steps I need to do this and the code snippet.

And here my output is in a JSON format.

It gets better.

I happen to be using Google's free, browser-based Jupyter

Notebook, known as Colab.

And I simply export the Python code to Google's Colab.

To summarize, Bart code generation

can help you debug your lines of source code,

explain your code to you line by line,

craft SQL queries for your database,

translate code from one language to another,

and generate documentation and tutorials for source code.

Generative AI Studio lets you quickly explore and customize

gen AI models that you can leverage in your applications

on Google Cloud.

Generative AI Studio helps developers create and deploy

Gen AI models by providing a variety of tools and resources

that make it easy to get started.

For example, there's a library of pre-trained models.

There is a tool for fine tuning models.

There is a tool for deploying models to production.

And there is a community forum for developers

to share ideas and collaborate.

Generative AI App Builder lets you

create gen AI apps without having to write any code.

Gen AI App Builder has a drag and drop interface

that makes it easy to design and build apps.

It has a visual editor that makes

it easy to create and edit app content.

It has a built-in search engine that

allows users to search for information within the app.

And it has a conversational AI Engine

that helps users to interact with the app using

natural language.

You can create your own digital assistants, custom search

engines, knowledge bases, training applications,

and much more.

PaLM API lets you test and experiment

with Google's large language models and gen AI tools.

To make prototyping quick and more accessible,

developers can integrate PaLM API with Maker suite

and use it to access the API using a graphical user

interface.

The suite includes a number of different tools such as a model

training tool, a model deployment tool, and a model

monitoring tool.

The model training tool helps developers train ML models

on their data using different algorithms.

The model deployment tool helps developers deploy ML models

to production with a number of different deployment options.

The model monitoring tool helps developers

monitor the performance of their ML models

in production using a dashboard and a number

of different metrics.

Thank you for watching our course, Introduction

to Generative AI.