GWENDOLYN STRIPLING: discipline, like physics for models, we have labels.

Hello. And welcome to example. AI is a branch of Labeled data is data that

Introduction to Generative computer science that deals comes with a tag like a name,

AI. My name is Dr. with the creation of a type, or a number.

Gwendolyn Stripling. And I intelligence agents, which Unlabeled data is data that

am the artificial intelligence are systems that can reason, comes with no tag. This

technical curriculum and learn, and act graph is an example of the

developer here at Google autonomously. Essentially, problem that a supervised

Cloud. In this course, you AI has to do with the theory model might try to solve. For

learn to define generative AI, and methods to build example, let's say you are

explain how generative AI machines that think and act the owner of a restaurant?

works, describe generative AI like humans. In this You have historical data of

model types, and describe discipline, we have machine the bill amount and how

generative AI applications. learning, which is a subfield much different people tipped

Generative AI is a type of of AI. It is a program or based on order type and

artificial intelligence system that trains a model whether it was picked up or

technology that can produce from input data. That trained delivered. In supervised

various types of content, model can make useful learning, the model learns

including text, imagery, predictions from new or from past examples to

audio, and synthetic data. never before seen data predict future values, in this

But what is artificial drawn from the same one case tips. So here the model

intelligence? Well, since we used to train the model. uses the total bill amount to

are going to explore Machine learning gives the predict the future tip amount

generative artificial computer the ability to learn based on whether an order

intelligence, let's provide a without explicit was picked up or delivered.

bit of context. So two very programming. Two of the This is an example of the

common questions asked are most common classes of problem that an

what is artificial intelligence machine learning models unsupervised model might

and what is the difference are unsupervised and try to solve. So here you

between AI and machine supervised ML models. The want to look at tenure and

learning. One way to think key difference between the income and then group or

about it is that AI is a two is that, with supervised cluster employees to see

whether someone is on the learning, let's briefly explore neural network to generalize
to new examples. Now we
fast track. Unsupervised where deep learning fits as a
finally get to
problems are all about subset of machine learning
where generative AI fits into
discovery, about looking at methods. While machine
this AI discipline. Gen AI is a
the raw data and seeing if it learning is a broad field that subset of deep learning,
which means it uses artificial
naturally falls into groups. encompasses many different
Let's get a little deeper and techniques, deep learning is neural networks, can process
both labeled
show this graphically as a type of machine learning
and unlabeled data using
understanding these that uses artificial
supervised, unsupervised,
concepts are the foundation
neural networks, allowing and semi-supervised
for your understanding of them to process more methods. Large language
generative AI. In supervised complex patterns than models are also
machine learning. Artificial
learning, testing data values a subset of deep learning.
neural networks are inspired
or x are input into the model. by the human brain. They are Deep learning models, or
made up of many machine
The model outputs a
interconnected nodes or learning models in general,
prediction and compares
neurons that can learn to can be divided into two
that prediction to the perform tasks by processing types,
training data used to train data and making predictions. generative and
Deep learning models discriminative. A
the model. If the predicted
typically have many layers of discriminative model is a
test data values and actual neurons, which type of model that is used to
training data values are far allows them to learn more classify or predict labels for
apart, that's called error. And complex patterns than data points. Discriminative
traditional machine learning models are typically trained
the model tries to reduce on a data set of labeled data
models. And neural networks
this error until the predicted can use both labeled and points. And they learn the
unlabeled data. This is called relationship between the
and actual values are closer
semi-supervised learning. In features of the data points
together. This is a classic
semi-supervised learning, a and the labels. Once a
optimization problem. Now neural network is trained on discriminative model is
that we've explored the a small amount of labeled trained, it can be used to
data and a large amount predict the
difference between artificial
of unlabeled data. The label for new data points. A
intelligence
labeled data helps the neural generative model
network to learn the basic
and machine learning, and
concepts of the task while
supervised and unsupervised the unlabeled data helps the
generates new data
instances based on a learned
probability

distribution of existing data. the data and the label, or or inputs used for the
Thus generative models what you want to predict. formula. So the model
The bottom image shows output is a
generate new content. Take
this example here. The a generative AI model which function of all the inputs. If
discriminative model learns attempts to learn the y is the number,

the conditional probability patterns on content so that it like predicted sales, it is not
distribution or the can generate new content. A gen AI. If y is a sentence,
good way to distinguish
probability of y, our output, like define sales, it is
given x, our what is gen AI and what is generative as the question
not is shown in this
input, that this is a dog and would elicit a text response.
illustration. It is not gen AI
classifies it as The response would be
when the
based
a dog and not a cat. The
output, or y, or label is a
generative model learns the on all the massive large data
number or a class, for
the model was
joint probability distribution
example spam or not spam,
or the probability of already trained on. To
or a probability. It is gen AI
summarize at a high level,
x and y and predicts the when the output is
conditional probability the traditional, classical
natural language, like speech
supervised and unsupervised
that this is a dog and can or text, an image or
then generate learning process takes
audio, for example.
training code and
a picture of a dog. So to Visualizing this
summarize, mathematically label data to build a model.
Depending on the
generative models can would look like this. If you
generate new data instances haven't seen use case or problem, the
model can give
while discriminative models this for a while, the y is equal
discriminate between to f of you a prediction. It can
different classify something
x equation calculates the
kinds of data instances. The dependent output of a or cluster something. We use
top image shows this example to show
process given different
a traditional machine inputs. The y stands for you how much more robust
learning model which the gen AI process is. The
the model output. The f
gen AI process can take
attempts to learn the embodies the function
relationship between training code, label data,
used in the calculation. And
and unlabeled data
the x represents the input
of all data types and build a from the multiple sources to the data it was trained on.
foundation model. The across the internet and As previously mentioned, a
foundation model can
build foundation language generative language model
then generate new content. models we can use simply can take what it has learned
For example, text, code,
by asking a question, from the examples it's been
images, audio, video, et whether typing it into shown and create
cetera. We've come a long
a prompt or verbally talking something entirely new
away from
into the prompt itself. So based on that information.
traditional programming to when you ask it Large language models are
neural networks
what's a cat, it can give you one type of generative AI
to generative models. In everything it since they generate novel
traditional
has learned about a cat. Now combinations of text in the
programming, we used to we come to our form of natural
have to hard code the rules
formal definition. What is sounding language. A
for distinguishing a cat-- the generative AI? Gen AI is a generative image
type, animal; legs, type of
model takes an image as
four; ears, two; fur, yes; artificial intelligence that input and can output text,
likes yarn and catnip. In the creates new content
another image, or video. For
wave of
based on what it has learned example, under
neural networks, we could from existing content. The
the output text, you can get
give the network process of learning
visual
pictures of cats and dogs and from existing content is
question answering while
ask is this a cat and called training and
under output image, an
it would predict a cat. In the results in the creation of a
image completion is
generative statistical model
generated. And under output
wave, we as users can when given a prompt. AI video,
generate our own uses the model to predict
animation is generated. A
content, whether it be text, what an expected response generative language
images, audio, might be and this generates
model takes text as input
video, et cetera, for example new content. Essentially, it and can output more text, an
models like PaLM or learns
image, audio, or decisions.
Pathways Language Model, the underlying structure of For example, under
or LaMDA, Language Model the data and
the output text, question
for Dialogue Applications, can then generate new answering is generated. And
ingest very, very large data samples that are similar under output image,
a video is generated. We've from the use of text difficult to understand.
stated that generative transformers. Transformers They can also make
produced
language models learn about the model more likely to
patterns and language a 2018 revolution in natural generate incorrect
language processing. At a
through training data, then, or misleading information. A
high level, a
given some text, they prompt is a
transformer model consists
predict what comes next. short piece of text that is
of an
Thus generative language given to the large
models encoder and decoder. The
language model as input.
encoder encodes
are pattern matching And it can be used to control
systems. They learn about the input sequence and
the output of the model in a
patterns passes it to
variety of ways. Prompt
based on the data you the decoder, which learns design is the
provide. Here is an example. how to decode
process of creating a prompt
Based on things it's learned
the representation for a that will
from its training data, it relevant task. In
generate the desired output
offers predictions of how transformers, hallucinations
from a large language model.
to complete this sentence, are words or phrases that are As previously mentioned,
I'm making a sandwich with generated
gen AI depends a lot on the
peanut butter and jelly. Here by the model that are often training data that
is the same nonsensical or
you have fed into it. And it
example using Bard, which is grammatically incorrect. analyzes the patterns
trained on a Hallucinations can be caused
and structures of the input
massive amount of text data by a number of factors, data and thus learns. But
and is able to including the model is not with access to a browser

communicate and generate trained on enough data, or based prompt, you, the user,
humanlike text in response the model is trained can generate your own
content. We've shown
to a wide range of prompts on noisy or dirty data, or the
illustrations of the
and questions. Here is model is not
another example. The types of input based upon
given enough context, or the
meaning of life is-- and Bart data. Here are the
model is not
gives you
associated model types.
given enough constraints.
a contextual answer and Text-to-text. Text-to-text
Hallucinations can be a
then shows the highest models take
problem for transformers
probability response. The a natural language input and
because they can make the
power of generative AI produces a text output.
output
comes These models are trained
to learn the mapping taking some sort of action. from text descriptions. Let's
between a pair of text, e.g. For example, a say you have
for example, translation
text-to-task model could be a use case where you need
from one language to trained to navigate a to gather sentiments
another. Text-to-image. Text-
web UI or make changes to a about how your customers
to-image models are trained
doc through the GUI. A are feeling about your
on a large set of images, foundation model is a
product or service. You can
each captioned with a
large AI model pre-trained use the classification
short text description. on a vast quantity of data
task sentiment analysis task
Diffusion is one method
designed to be adapted or model for just that purpose.
used to achieve this. Text-to- fine tuned to a wide range And what if you needed to
video and text-to-3D. Text-
of downstream tasks, such perform occupancy
to-video models aim to
as sentiment analysis, analytics? There is a task
generate a video model
image captioning, and
representation from text
object recognition. for your use case. Shown
input. The input text can be
Foundation models here are gen
anything
have the potential to AI applications. Let's look at
from a single sentence to a
revolutionize many an example
full script. And the output is
a video that industries, including health of code generation shown in
care, finance, the second block
corresponds to the input
text. Similarly, text-to-3D and customer service. They under code at the top. In this
can be used to example, I've input a
models generate three
dimensional objects that detect fraud and provide code file conversion
personalized customer problem, converting from
correspond to a user's text
support. Vertex AI offers a Python to JSON. I use Bard.
description. For example,
And I insert into the
this can be used model garden that includes
foundation models. The prompt box the following. I
in games or other 3D worlds.
language foundation have a Pandas DataFrame
Text-to-task. Text-to-task
with
models are trained models include PaLM API for
chat and text. The vision two columns, one with the
to perform a defined task or
foundation models file name and one with the
action based on text input.
hour
This task can be a includes stable diffusion,
which has been shown to in which it is generated. I'm
wide range of actions such as
trying to convert
answering a question, be effective at generating
high quality images this into a JSON file in the
performing a search, making
format shown onscreen.
a prediction, or
Bard returns the steps I need
to do this and the code library of pre-trained models. test and experiment with
snippet. And here my output There is a tool for Google's large language

is in a JSON format. It gets fine tuning models. There is a models and gen AI tools. To
better. I happen to be using tool for deploying make prototyping quick
Google's
models to production. And and more accessible,
free, browser-based Jupyter there is a community developers can integrate
Notebook, known as Colab.
forum for developers to PaLM API with Maker suite
And I simply export the
share ideas and collaborate. and use it to access the
Python code to Google's Generative AI App
API using a graphical user
Colab. To summarize, Bart
Builder lets you create gen AI interface. The suite includes
code generation can help apps without a number of
you debug your
having to write any code. different tools such as a
lines of source code, explain Gen AI App Builder has a model training tool, a model
your code
drag and drop interface that deployment tool, and a
to you line by line, craft SQL makes it easy to model monitoring tool. The
queries model training tool helps
design and build apps. It has
for your database, translate a visual developers train ML models
code from one on their data using
editor that makes it easy to
language to another, and create different algorithms. The
generate documentation model deployment tool helps
and edit app content. It has a
and tutorials for source code. built-in developers deploy ML
Generative AI Studio lets you models to production with a
search engine that allows
number of
quickly explore and users to search for
customize gen AI models that different deployment
information within the app.
you can options. The model
And it has a
monitoring
leverage in your applications
conversational AI Engine that
on Google Cloud. Generative tool helps developers
helps users to
AI Studio helps monitor the performance
interact with the app using
developers create and of their ML models in
natural language. You can
deploy Gen AI models by production using a
create your own digital
providing a
dashboard and a number of
assistants, custom search
variety of tools and different metrics. Thank you
engines, knowledge bases,
resources that make it easy for watching
training applications, and
to get started. For example, our course, Introducti
much more. PaLM API lets
there's a
you
on to Generative AI.