Compare AI Vs. Machine Learning

0:00Hello and welcome to the next scale in the AWS Certified AI Practitioner course

0:07.

0:07In this scale, we're going to compare AI with machine learning.

0:11Up to this point, we've talked about cloud versus local and all of the related

0:17AI core

0:18concepts and tools.

0:20Now we'll begin to compare AI versus machine learning and we'll do that by

0:26asking is AI

0:28the same as machine learning.

0:31And in this scale, we'll break everything down and explore AI in its different

0:35subdomains

0:37at a deeper level.

0:38But first, let's talk about everything we're going to cover in this scale.

0:42Number one, we're going to talk about AI versus machine learning and really ask

0:47what

0:48is the difference.

0:49Then we'll talk about real world AI and we're going to do that with use cases

0:54again.

0:55This time, we're going to compare it to the subdomains.

0:58Now we need to understand what is AI, generally speaking, what is machine

1:03learning and what

1:04is deep learning and what is gen AI.

1:07These are all important subdomains and we need to start to understand where

1:11they fit

1:12into in the real world of AI.

1:15Then we'll dive into deep learning and explain it simply.

1:18We've talked about it in the context of software modeled after the brain.

1:23This is also what we described as software 2.0 because it learns from data

1:29using neural

1:30networks.

1:31And next, we'll dive into how AI recognizes handwritten digits.

1:35There's a lot to talk about when we're talking about CV and what is that?

1:40Computer vision.

1:41If you remember, we said CNN, that's part of computer vision and we'll talk

1:46about that

1:46with some amazing visualizations.

1:48In number 5, we're going to talk about generative AI, also abbreviated as gen

1:55AI and how machines

1:57create text, images and more.

2:00And then we're going to really just tie it down with humans plus AI, which is

2:05prompting

2:06software 3.0 and the future.

2:09And then number 7, I'm just going to put it up here.

2:11This is the challenge just to make sure that we don't have any knowledge gaps

2:16and that

2:17we understand all of this as we start to get more granular and start to talk

2:21about deep

2:22learning more of the neurons or the deep layers and then computer vision.

2:28So the way I like to think of it as NLP or natural language processing is how

2:32we teach

2:33the computer to read and write.

2:37And then computer vision is how we teach the computer to see the world.

2:41So number 6, when we get into the humans plus AI, it's a great tool, I would

2:47say, to think

2:47about all of these deep learning concepts as if it were in your mind.

2:53Because we do have biological neural networks and you can think about computer

2:57vision, you

2:58can think about NLP.

3:01And I think it's a really great way to internalize these concepts.

3:05If it feels alien at all, always just remember there's a biological version

3:10inside of you.

3:11All right, so I'll see you in that next video,

3:13AI versus machine learning.

0:00Welcome back. We already know that AIs a field that creates systems that mimics

0:04human intelligence

0:06or you can think of it as software modeled after the brain. And these systems,

0:11much like human

0:12minds can perceive, think of cameras seeing like eyes, reason, chess playing AI

0:20or AlphaGo

0:20beating the best Go player, and learning when we say learning, we mean learning

0:25from data and

0:26rules. Also, they can solve problems and make decisions. Artificial

0:31intelligence is a field of

0:33computer science focused on building systems capable of performing tasks

0:37typically that require

0:39human cognitive abilities. And remember, we talked about software 1.0, 2.0, and

0:453.0.

0:45Those are all inside of this main term artificial intelligence. And I think we

0:51have a good understanding

0:52of what AI is. But first, let's expand on the AI use cases, because I'd like to

0:58make sure that

0:59we understand which type of use case is using which type of artificial

1:04intelligence. So let's start

1:06with computer vision. Computer vision is using AI to analyze images and videos.

1:12And you can think

1:13of video as a series of images. If you know what 24 FPS means, that means 24

1:20frames per second.

1:21And that's normally how many frames per second that we see when we're watching

1:25movies. So they're

1:2624 images. So if you think of it like that, computer vision is really just

1:30working with images. It

1:32can be one at a time or many in a series. So what kind of machine learning do

1:36you think this is?

1:38Let's break it down. It extracts patterns like edges, textures, and objects

1:43from pixels. It's

1:44using autonomous cars to detect lanes, signs, and pedestrians. And neural nets

1:50process video

1:52frame by frame to make driving decisions. Again, it's set frame by frame. But I

1:57think the takeaway

1:58here is that we said neural nets. So if you said that this was deep learning,

2:02then that's correct.

2:03No logic, just learning from the data. All right. So now how about facial

2:09recognition?

2:10Here, AI identifies and matches human faces. You've probably seen movies where

2:15they're trying to

2:16find somebody and they look at all the public camera feeds. So what kind of AI

2:21is this? Let's

2:22talk about how it works. It converts facial features into numerical factors

2:27called face prints. These

2:30are then compared using cosign similarity or nearest neighbors. These are again

2:35algorithms.

2:36And this is used often to unlock phones, right? It looks at your face and you

2:40can unlock your

2:41phone with the face. Airport security to find people that are not supposed to

2:46be in the airport

2:47and even surveillance systems. So what kind of AI is this? If you were thinking

2:52about face images

2:53and convolutional neural networks, neural nets that is correct, this is also

2:58deep learning.

2:58All right. So now how about raw detection? Here AI spots abnormal behaviors in

3:04transactional data.

3:06Here, we're using models like decision trees or random forests or others called

3:12XG boost and all

3:13of these learn spending patterns. Remember that we looked at a random forest

3:17earlier? That might

3:18ring a bell. They flag transactions that differ from normal patterns. That

3:22means automates, weird

3:24times or strange locations. For example, let's say that you bought a pizza, you

3:29're in New York City,

3:31and then 15 minutes later, someone in China bought a Tesla. Well, probably not

3:36the same person,

3:37right? You went from a pizza to a car in a completely different continent. So

3:41that might be

3:42something that gets flagged and they don't need images, just structured tabular

3:47data. So that's

3:48the clue. In this case, this is traditional machine learning, also software 1.0

3:53, because you are

3:55writing the logic, you are writing the code. All right. So now let's move to ID

3:59P or intelligent

4:00document processing. This is a little bit trickier. Let's try to think through

4:04it. AI extracts text

4:06and meaning from form and it's using something called OCR. This is optical

4:11character recognition.

4:13What it's doing is recognizing the characters and it's converting scanned

4:18images to text.

4:19In short, it transforms messy documents into clean, usable digital records. All

4:25right. So this

4:27is interesting because it's a mix of deep learning plus machine learning. Let

4:31me explain how this works.

4:33So the deep learning is for the image in text parsing, and then the machine

4:37learning is for

4:38classification and tagging. So the OCR uses deep learning to convert scanned

4:45images to text,

4:46then the machine learning part, which can also be rule based systems, labels,

4:52fields like dates,

4:53totals, and names. So what it's doing is combining both deep learning and

4:58machine learning. So here's

4:59the last one, robotics. What do you think? Well, the difference between

5:03computer vision and robotics

5:05is very similar because it's using computer vision to navigate the real world.

5:11And so this is much

5:12harder than chat GPT deep learning again is used for robotics. And so what's

5:17the takeaway? They all

5:19use some kind of AI, but now we know which are using which. And also we know

5:24that they could be

5:25mixed. Now let's talk about the AI system layers, also known as AI components.

5:33So let's start at

5:33the bottom here with a data layer. So here's where we have the data and the

5:37data scientist.

5:38This is where we collect, store, and organize structure and unstructured data.

5:44These are data

5:46structures. There's different kinds. And so these are the two that we're going

5:50to really focus on

5:51in this skill. So what does a data scientist do with the data? Well, you have

5:56to frame a problem

5:57and try to solve a problem, but using machine learning in this case, that's the

6:02framework layer.

6:03Do we want to use deep learning and in TensorFlow and PyTorch? Or are we going

6:07to use

6:08traditional machine learning and algorithms using scikit learn? So we choose

6:13the tools,

6:14libraries, and algorithms for the development of the model. And then we move

6:18into the model

6:19component. And here we train and define the model architecture. We do some

6:24tuning and loss

6:25functions to get the model to be performant. We want to increase the

6:29performance as much as

6:30possible. We really want to minimize error and maximize orments. So builds,

6:35trains, and tests

6:37the machine learning model. And finally, we have the application layer. And

6:41that's where users can

6:42access the model. So we deploy the train model for real time or scheduled use.

6:48So these are the

6:49ones that I'm checking that are important. Let me just put asterisks here and

6:52here. We're going to

6:53revisit these to real quick note on the application layer. There's two ways

6:57that we can handle this

6:59application layer that can be an API where two computers are talking to each

7:03other or UI user

7:05interface where someone is using a website or mobile phone or an interface

7:10directly for users.

7:12All right. So that's it for this video. And I'll see you in the next video

7:15where we answer the

7:16question, what is machine learning?

0:00Welcome back. In this video, we're going to talk about machine learning and how

0:03machines learn from data so that we can compare it to rule-based systems and

0:08how

0:09Rule-based AI is not machine learning. One is a rule-based system and the other

0:14one learns from data.

0:16All right, so we know that machine learning is a subset of AI that learns

0:20patterns from data instead of being

0:23Explicitly programmed it improves over time and it can make predictions. For

0:28instance, instead of following hard-coded rules

0:31Machine learning learns from examples. They use past data

0:35So we're talking about historical data to recognize patterns and make

0:40predictions

0:40Machine learning is a part of AI focused on learning from experience

0:45So we're going to dive into regression here in classification to further

0:49illustrate these differences and the process relies on statistics

0:53There is programming, but it's using statistics and to clarify this. We're not

0:59using logic to find these patterns

1:02We're using statistical programming. So first of all, what is regression

1:07regression predicts continuous numerical values?

1:11So real quick, what is a continuous value?

1:13This is something with a decimal point and that could be housing prices

1:18temperature sales stock prices

1:20So our example is predicting house prices, but you could also use temperature

1:25or sales if you'd like

1:26So input features could be size

1:29Location or the age of the house

1:32So you would feed those input features to your model and then the model finds a

1:36relationship between the input and the output

1:39A common algorithm that we would use here is linear regression

1:43And that's why in this image you see a line going through those data points as

1:47we discussed before in

1:49In addition to linear regression, we can use decision trees and X cheap boost

1:53This is what we've been talking about when we're talking about machine learning

1:57Classification on the other hand predicts discrete labels or categories

2:02So the example here is it spam or not spam so the model learns from labeled

2:08examples and the outputs will be something like spam or not spam

2:12Hot dog or not hot dog and we explored this in a fun example with teachable

2:17machine

2:17So you should have a good idea of how this works

2:20Common algorithms include logistic regression support vector machines random

2:26forests and naive base

2:27Now I'm saying that regression and classification are machine learning

2:32But remember that machine learning is a form of AI in a subset of machine

2:36learning as deep learning

2:38So we were using deep learning for classification now that we have a really

2:43good idea of how machines learn from data

2:45Let's compare it to rule based systems. So what are rule based systems or what

2:52is rule based AI?

2:54We can say that it's classic AI and by that I mean no learning no data just

3:00rules

3:01So you manually program logic and that would be something like if X and Y then

3:08do

3:08Z and that is programming logic it follows flow charts or decision trees and

3:15Everything has to be coded up front. It doesn't improve with experience. It's

3:20just following rules

3:21So here the user inputs information and then the expert system does the rest

3:27Let's say that this non expert users

3:29Interacting with a UI or a user interface and then there's this rules engine in

3:35a knowledge base

3:36But all of that knowledge base was coming from an expert

3:41They literally had to add all of those rules and then once the rules engine

3:47Communicates to the knowledge base it flows back to the UI and back to the user

3:51and then provides of advice

3:53So let's take an example then draw is an expert system and it was designed in

3:59the 1960s for chemistry

4:02It analyzed mass spectrometry data to guess molecular structures and in this

4:09example

4:09It used over a thousand handcrafted rules and that means no learning just

4:15reasoning from fixed logic

4:17All right, so let's go through this dental expert system

4:21It engaged in scientific discovery by analyzing organic compounds and

4:26identifying their structure

4:27Which is pretty complex, but it did so using a collection of over a thousand

4:32rules and these rules apply to data using

4:36Logical reasoning and you can call that an inference engine in this case

4:40It was developed for mass spectrometry and it was used to generate the chemical

4:44structure

4:45Hypothesis from data and so the takeaway here is that dendral was rule based

4:51and used structured lab data

4:54Not machine learning to infer possible molecular structures

4:58So let's compare rule based AI with machine learning

5:02So we know that this one used rules and we can ask a question that it learned

5:07from data and the answer here is no

5:10And here we're learning patterns and did it learn from data?

5:14Yes, how about deep learning and prompted AI?

5:18So deep learning is using neural nets. I'll just use NN for that and

5:23Does it learn from data?

5:26Yes, even more so than machine learning. How about prompted AI?

5:31It uses large language models and does it learn from data? Yes

5:36Even more so than deep learning because there's no code. So this is software 1.

5:440

5:44This is software 2.0 and this is 2.0 because it learns from data

5:49You're not explicitly programming it and this is 3.0 because you're using

5:54prompts to do the programming

5:56There's no explicit code

5:58So the takeaway is that rule based is logic written by humans

6:02Machine learning is logic learned from data

6:06Deep learning is logic that is learned end to end from data and prompted AI

6:12You just give it goals and plain English and it accomplishes them

6:16So I would say that rule based AI is fixed machine learning is flexible

6:22Deep learning is powerful and prompted AI is creative and language driven

6:29All right, so that's it for this video and I'll see you in the next video when

6:33we talk about deep learning

0:00Welcome back. In this video, we're going to dive into deep learning.

0:04And we're going to start to think about human neurons or our biological neural

0:08network to help us understand artificial neurons.

0:12First of all, you might be asking, "What does a neuron even do?" So let's talk

0:16about that.

0:17So let's say that this is a neuron and it's taking in an input.

0:21And we're going to say that there's weight to this input.

0:25Meaning, how important is this piece of information, this input?

0:30And we'll give this a threshold of 0.5.

0:34So if this input is more important or has a higher value than 0.5, let's say

0:40that the weight is 0.6.

0:43That means that it's going to go into the neuron, and when it reaches this

0:46neuron, it's going to fire forward.

0:49But let's say that instead of 0.6, it's 0.2.

0:54Well, below the threshold.

0:56That means when it gets into the neuron, this input, since it's below the

1:00threshold, it just won't fire.

1:02And so what happens in a neural network is these weights get adjusted over time

1:06.

1:07And that's how it learns.

1:09Understanding the neuron is the first part.

1:12So deep learning uses artificial neurons which are computational units.

1:17And remember, all they do is decide whether or not something goes through or

1:21does not.

1:22That's it.

1:24And they're designed to simulate the way the human neurons process information.

1:28And that's how human neurons do it.

1:30They either pass it to the next neuron or they don't.

1:33So these artificial neurons are organized in layers to analyze and learn from

1:38data.

1:39So the deep part is that there's many, many layers of these neurons.

1:43You can think of this as a neural network.

1:46So you have the input going here into these layers.

1:50And it says, how important is this input?

1:53Let's say this is like horsepower miles per gallon and cylinders.

2:00And it's some sort of a neural network to understand if a car is fuel efficient

2:03or something like that.

2:05So this is not important.

2:07This is important.

2:09This is not important.

2:10And so there are going to be many hidden layers like many of these.

2:14That's why it's called deep.

2:15And it's going from the input and then eventually you'll get this output that

2:19will tell you something that you want to know about the input, so the

2:23relationship.

2:24And it does this through these neural networks.

2:27Now that's a very simplified version.

2:29But the reason I explain it like this is that we need to understand what the

2:32neurons are doing.

2:34And they're acting like a filter or a gate.

2:37You know, just think of, just think of Gandalf.

2:40Thou shall not pass.

2:42Or thou shall pass.

2:44Think of a neuron as Gandalf.

2:46All right.

2:47Deep learning finds complex patterns and data more than traditional machine

2:51learning methods.

2:53And you can really think of deep learning as the newer version of machine

2:56learning.

2:57So it's more powerful.

2:58It uses multiple layers to process information.

3:02And that's why it's called deep.

3:04And these layers, we're thinking about networks of these artificial neurons.

3:09It needs large amounts of data to work well.

3:12And it requires a GPU.

3:14This is a powerful processor to handle the calculations.

3:18And the good news is that we can use GPUs for free and co-lab.

3:23We're also going to use AWS, of course, but as I said, this is an industry

3:28standard in machine learning and deep learning.

3:31So some examples of deep learning include computer vision, which we've talked

3:36about.

3:37It identifies objects or dividing images into parts.

3:41And then this is like teaching the computer to see.

3:45And this is teaching the computer to read and write.

3:49And that's natural language processing.

3:51And these are tasks like translating languages or analyzing feelings and texts

3:56that's called sentiment analysis.

3:59What's important to take away from deep learning is that they require lots of

4:04data and powerful GPUs.

4:06I think that's the most important part and how these neurons are working.

4:11They're basically ganned off. They're allowing you to pass or not allowing you

4:15to pass depending on some threshold.

4:18Alright, so now here we have a neural network and we can examine how it works.

4:25So nodes, these are small units that are interconnected.

4:29These nodes are the neurons.

4:31So you have the input going into these hidden layers and you can have not just

4:35four, but you can have a bunch of these.

4:37The more you have, the deeper your neural network would be.

4:41And as you can see, they're interconnected.

4:43These nodes are organized in layers.

4:46And when the neural network sees a lot of data, what it's doing is that it

4:50identifies patterns and changes the connections between the nodes or these

4:56neurons.

4:57How does it change that?

4:58By changing the weight. And what is weight?

5:01How important is this connection to this one?

5:04So if the connection is not important, well, it's not going to pass to the next

5:09neuron.

5:10And so it'll change the weight of that feature and it'll give another feature,

5:16which is more important, more weight.

5:18So it can do that over and over and over in order to classify digits.

5:23And we're going to do that together using Google Colab.

5:27It can get to the point where it says this curvature is important, but this

5:31straight line is not important for this input.

5:36And that's how it makes this prediction.

5:38And so what's cool here is that these nodes are talking to each other.

5:42That's what happens in our brains as well.

5:44And they do so by passing or not passing data to the next layer.

5:50So that's that Gandalf action that we're talking about.

5:53And the math and the parameter tuning behind deep learning is beyond the level

5:57of this course.

5:59But the good news here is that we don't have to talk about this in this course

6:04because we're using AWS and it abstracts a lot of this.

6:08And that's what makes this such a good entry level course to get started with

6:14building and architecting AI applications on AWS.

6:19And finally, neural networks can have billions of nodes or neurons.

6:24So the takeaway here is that the data flows through these neurons and these

6:29connections between the neurons change as it learns.

6:33And that's about it for this video.

6:35And I'll see you in the next video where we explore a deep learning example,

6:40the MNIST data set.

6:42And that's going to be the handwritten digits.

6:45See you there.

0:00In this video, we're going to explore two examples. Here we have the MNIST

0:05dataset of hand-written

0:06numbers. This is a classic example. And we're also going to look at the fashion

0:11dataset, also an

0:12MNIST dataset. And this way, we can compare and contrast between the two. We're

0:17going to look at

0:18it from a high level and try to understand conceptually what's happening by

0:23looking at the neural network

0:24that we've been discussing, these artificial neurons. And then we'll go into

0:29Google Colab and look at

0:30different fashion items and try to understand how it's done programmatically.

0:35And I'll guide you

0:36through each part of the code. You don't have to write the code. It's just to

0:39get you to understand

0:41what it does at each stage in order to make these predictions. When we're

0:45looking at this example

0:47for the neural network, the first layer detects lines and edges. And what that

0:52means is that each

0:52layer is learning from the features or these inputs. We can call them input

0:57features. So the

0:58first layer detects lines and edges and the deeper layers recognize digit

1:03shapes. And then the final

1:05layer chooses the most likely digit. So when we're thinking about hand-written

1:10numbers, these are

1:11going to be the input data or the input features. And the process begins with

1:16an input layer consisting

1:18of pixel values. So each one of these images turns into a pixel. And then the

1:24data is analyzed through

1:26hidden layers that detects lines and curves. As you can see, zero has a lot of

1:31curves. And one

1:32is just pure lines. Finally, the output layer determines the most likely number

1:37. And that's

1:38this output here. Each layer recognizes a distinct pattern within the data. And

1:44pattern examples

1:45include vertical lines for the numbers one, four, seven, and a curved bottom

1:52for the numbers six,

1:54eight, and zero. Now this works is that the neural network learns to identify

1:59all of these

2:00patterns. Now let's take a look at the MNIST data set for fashion in Google Col

2:06ab. If you want to

2:07learn more about the MNIST database of the hand-written digits, here is the

2:12Wikipedia page. And I'll

2:14share this link below. And we're going to look at a similar MNIST data set but

2:19fashion. And here's

2:20a good view of all the items. You have shirts, pants, shoes. And this is a

2:25representation of the

2:27data set in a vector space. And we'll talk more about that as we progress

2:30through the course. But

2:32now let's take a look at the Google Colab here. So this first cell, we're going

2:37to shift enter,

2:38or you can hit the play button. What we're doing is importing TensorFlow. That

2:42's the framework we're

2:43going to use for our model. And then matplotlib to visualize and NumPy. This is

2:48something that

2:49we often use in machine learning and deep learning. And here we're going to

2:52load the data from the

2:54fashion MNIST from TensorFlow. Let me run this. And there we have our data

2:59makes it really easy.

3:01And then we can look at the classes. So I know that it's T-shirt, tops, trouser

3:06, pullover, dress,

3:07coat, sandals, shirts, sneakers, bags, and ankle boot. So if we run this, now

3:12we can see that this

3:13is the classes that we're working with. One, two, three, four, five, six, seven

3:17, eight, nine, 10.

3:19So 10 classes. And now let's take a look at our images. And we can kind of

3:24search through eight

3:25images. And that's what we were doing here. Number of rows and columns just to

3:28look at the

3:29images and see what we have. And now we have our model. So here we're going to

3:33create a train

3:34test split, meaning we're going to have training data. This is where the model

3:38learns the patterns.

3:40We have the test data. And this is where the model is measured or tested to see

3:47how well it does.

3:48If it does really well on the test, then that means that it has a good chance

3:52of making predictions.

3:53Let's go through each one of these. So let me go ahead and run this. And now we

3:58're going to

3:58convert the labels to one hot encoding. This is pre processing. We have to

4:04change the data.

4:05And that's what we did here. We did normalizing whenever you change the image

4:10to numbers,

4:11there's some pre processing that you have to do. And now here is the model. So

4:15what's really cool

4:16here is that we're using a sequential model, meaning we're doing one layer at a

4:21time. So here is the

4:22input layer. And 28 by 28 means each image for each garment is 28 by 28 pixels.

4:30Here we have our

4:32first layer in what this number is for neurons. So we have one layer for

4:36neurons, another layer

4:38of four neurons, right? So this is our neural network. It's pretty cool. So we

4:43have two layers

4:44of four neurons. So we have eight neurons in this neural network. But 10 is the

4:49number of classes

4:51that we had. So the output is going to be 10. Let's go ahead and run this. So

4:55all this does.

4:56Okay, so this is this code. I wrote it a while ago. We don't need this anymore,

5:01because it's changed. The code has changed. So they updated it. So you don't

5:06need the input

5:06shape anymore. But I'm glad I had that there because I was able to explain that

5:10there was 28

5:11pixels by 28 pixels. Okay, so this has to do with optimization and the loss

5:16function. We're

5:17not going to talk about this right now. But what we're going to do here is

5:20train the model. So

5:22let me, okay, so right now I'm just going to take off callbacks. Well, actually

5:26, let me run this.

5:28And then I can run this again. That is not important for right now. But what is

5:32important,

5:32it's this part. This is the first epoch of 25. And this is where it's learning

5:38the patterns

5:38of all of the different classes that we looked at. This is the machine learning

5:43part in action.

5:45So what's happening is going through each one of these neural network layers,

5:49and it's adjusting those weights. And that's how the neurons are talking to

5:52each other.

5:53And then here you can see that we have a validation accuracy and in general

5:58accuracy. What this means

5:59is this is the performance of the model. So you can see that it's getting

6:02better and better and

6:03better. And here is being tested on unseen data. It's kind of like taking a

6:09test and

6:11memorizing the answers. You don't want to do that because then when you see new

6:15answers,

6:15you've never seen you'll do poorly. So that's why you want to compare the

6:19accuracy with the

6:20validation accuracy. The accuracy of this model is pretty good. It's already at

6:2470%. And it's also

6:26the same for validation accuracy. So when we get to the 25th epoch, we should

6:31be at a pretty good

6:32percentage. And this model will make predictions. For example, we give it a

6:37picture of, let's say,

6:38sandals or a shirt, it should accurately predict that. All right, so it got to

6:44the very end and it

6:45got to almost 74% and 71% for validation accuracy. That's pretty good. And the

6:52takeaway here is that

6:53you don't have to learn what the code does exactly. And my other courses, like

6:59introduction to machine

7:00learning and introduction to deep learning, we really code all of this together

7:05. So you understand

7:06every line of code. If you want to do that, then I've listed those courses as a

7:10resource.

7:11The reason we're doing this is because in AWS, it abstracts a lot of this. That

7:16means that

7:17you can just give it images and you could tell it what you want it to do. And

7:22it'll do it.

7:23I'm going to give you all of the foundational concepts that you need to not

7:27only pass the AWS

7:30certified AI practitioner, but also to build awesome AI applications and tools

7:37using AWS

7:39services. So I'll leave this link below this video so that you can run this

7:43code if you like.

7:44And you can take this code and put it into a large language model. It'll

7:48explain what normalization

7:49is, what one hot encoding is, if you want to learn more, you can do that. Or

7:54you can just take

7:54one of my courses, I'll leave it up to you. All right, so that's it for this

7:58video. And I'll

7:59I'll see you in the next video on generative AI.

0:00Welcome back. Generative AI is probably one of the most exciting parts of AI.

0:06So let's take a look at the different parts of the architecture and also some

0:11of the fun stuff

0:12like Dolly that creates images. When I think about generative AI, I think about

0:17it not

0:18analyzing but instead creating like creating text, images, sound, movies, and

0:25even code.

0:26We discussed earlier vibe coding and that's where people that don't have any

0:31programming background

0:32can talk to the different language models to generate new code. And I've seen

0:38complete entire

0:39worlds where people are collaborating in a game environment flying multiple

0:44different aircraft.

0:45And I'm talking like thousands of people and that was built with code but no

0:50coding experiences.

0:51So generative AI has a definite magical quality to it. And it's powered by deep

0:57learning architectures

0:58like the transformer. And in the challenge, we're going to use generative AI to

1:04learn about the

1:05transformer. But also I'm going to show you a couple of visualizations that

1:09really show how

1:11the transformer architecture works. It's super fascinating. Just like humans

1:17have attention,

1:18this transformer architecture uses attention to create weight for each word.

1:25And it completely

1:26understands a whole entire sentence at once, unlike other architectures which

1:32go token by token.

1:33And they can't remember the context for very long. And we know that foundation

1:38models are trained

1:39on large scaled or unlabeled or mixed data. And they can be adaptable to

1:44specific tasks via fine

1:46tuning or prompting. And so prompting is one example of creating vibe coding.

1:53So what's happening

1:54behind the scenes? So you have all of this data which can be unlabeled or mixed

2:01data types,

2:02videos, text, images, or structured or unstructured data. And here the pre-

2:09trained, this is where

2:10there's fine tuning that can go on in the foundational model. And then that

2:14means that you can adapt

2:16it to text generation summarizing image generation, which we're going to look

2:21at now and chat GPT and

2:24question and answering. So here's an example of generative AI that's pretty

2:29well known,

2:30which is Dolly. And as you can see, this is an amazing image. And we create

2:34these with these

2:35prompts. For example, here is one that we can grab and then go to chats GPT and

2:42then paste it in

2:43here. And that I kind of like a little bit better than this one, to be honest,

2:48wow, that's pretty

2:48cool. So now here I created another prompt. Let's go ahead and run this and see

2:53what happens.

2:54A hyper realistic, highly detailed photograph of a lifelike T-Rex, smoothly

3:01roller skating on a

3:02sunny day in Central Park, New York City. And I use this dictate. So we just

3:07performed a speech

3:09to text. And now we're going to get this image. So I like this line here, other

3:13dinosaurs in the

3:14background with people running and terra dactyls in the sky. This should be

3:19pretty interesting.

3:20And that's pretty cool. So I'm definitely going to download this and add it to

3:24the skill.

3:25Now let's talk about the transformer models. And this is the very architecture

3:30that started at all.

3:32When chat GPT took off, that T and GPT stands for the transformer. And the

3:38transformer architecture,

3:40when they created the white paper, people really didn't take it serious. They

3:46didn't know that it

3:47was going to take off. And that's what's so cool about the white paper. And we

3:51'll talk about that

3:52white paper in the challenge so that you can create a podcast in that challenge

3:56so that you can learn

3:57more about the transformer architecture using generative AI. And what's cool

4:02about this

4:03transformer architecture is that it leverages hardware. And it wasn't until we

4:09had sufficient

4:10compute power that we saw the power of the transformer. So next time that you

4:15're totally amazed by

4:17chat GPT, that's because of the attention mechanism in the transformer. Let's

4:22check it out. The reason

4:24that the transformer changed everything is because the model pays attention to

4:29every word in a sentence

4:31and not just one at a time. And this is how chat GPT writes such good text and

4:37can understand

4:39context. So it understands the entire sentence at once instead of step by step,

4:45which is what

4:45RNNs did. And these RNNs or recurrent neural networks are the older models that

4:53processed

4:53words one at a time and they forgot earlier information. Then comes the

4:58transformer. And they

4:59look at the whole sentence at once. And this is called self attention. And we

5:04have it right here.

5:05And another cool attribute of the self attention mechanism is that it enables

5:11faster, more

5:12efficient training and generation. And it uses that attention mechanism to

5:17weigh which words in a

5:18sentence matter most. So here we have an input and it goes into an embedding.

5:24And this is how we

5:25convert text into a format that we can feed into the model. And then we have

5:31self attention. And

5:32this self attention then encodes the information, which is basically numbers,

5:38and then passes it to

5:39the decoder. And then it has the self attention again, and decodes it and then

5:44produces an output.

5:46And softmax is going to give you a list of words. So what it is is a

5:51probability distribution

5:53that says which one is the most probable, right? So that self attention

5:58mechanism is what gives the

5:59weight to these different words. Let's take a look at a visualization so that

6:04we could try to

6:05understand this a little bit better. Okay, so here we have the transformer

6:10explainer from polar club,

6:11which this allows us to go into great detail. I was just talking about the soft

6:17max probability

6:18distribution. And that's right here. So what it does is that it gives each word

6:23a different

6:24percentage. So here's softmax. And you can see that each word has a different

6:28percentage. And so

6:29it picks visualize. So data visualization empowers users to visualize. That's

6:35what's happening behind

6:37the scenes. So if I click on generate, you can see what's happening with the

6:41attention mechanism.

6:42And then you can see that it shows a comma because data visualization empowers

6:47users to visualize,

6:49comma, organize, comma, and visualize. So you see, it's literally taking you

6:56through and adding

6:58these embeddings. If you want to learn more about embeddings, it creates tokens

7:03, which are these

7:04numbers, and then it has positional encoding. And then that's what goes into

7:09the multi head

7:11self attention mechanism. And we're not going to get too far into this, but you

7:15can see the

7:15intention mechanism, which then takes all of these words and puts them into

7:20this probability

7:21distribution. And then it tells you that the next most probable word is going

7:25to be visualize.

7:26I will definitely link this so that you can play with it below and check this

7:30out. You can,

7:31it'll explain what it is more about embeddings, transformer blocks, and the

7:36output probabilities,

7:37the softmax, more visualizations, and tokens and tokenization. We're going to

7:42get into all of this

7:43stuff in a later scale. But I did want to give you this resource for now. All

7:48right, now on to

7:49diffusion models, diffusion models, generate images by starting with pure noise

7:55and gradually

7:56turning that noise into meaningful images. So there's two key steps. So the

8:02training. And this is where

8:03we take a real image and then slowly add noise until we get a box that

8:08resembles something like

8:11this. And the reason they do this is that the model learns how the noise

8:16changed the image. And

8:18then it can do this amazing thing by flipping it in reverse. So then it starts

8:24with noise here. And

8:26then it ends up with this image here. So the model reverses the noise step by

8:32step. And then we get

8:33that bunny image that was generated from this prompt using a diffusion model.

8:39And I'm getting

8:40all of these amazing visualizations from the Polo Club of data science. And if

8:45you scroll down,

8:47you'll see that this is where I got the transformer explainer. I'll definitely

8:51share this with you

8:52as well, because we also have the diffusion explainer. And this is showing you

8:57step by step how it goes

8:58from noise to an image. And this is the part where it reverses. This is not the

9:04training part. If you

9:05want to learn more, this goes into great detail. And this is a wonderful

9:10resource. So what are

9:11multi model models? So good examples are cloth three from anthropic, grok from

9:18X AI,

9:20Gemini from Google DeepMind. And they can understand more than one kind of

9:25input. So we're talking about

9:27a prompt, maybe an image or audio. And then you have this prompt generate a

9:33video with the picture,

9:34make the bunny speak in the included audio file. And then this multi model

9:40model will generate a

9:41video, including all of the different types of data. And it can generate words,

9:46pictures,

9:47audio. And in this case, video. And what's really exciting is that you can

9:51speak to the model,

9:53show it a photo, ask it to respond. And it will respond with text, voice,

9:58drawings, or videos.

9:59And one good example of a multi model model is lava, which is a large language

10:05and vision

10:06assistant. And you can check that out from old llama. So how do human brains

10:11compare to AI?

10:13I find it really useful when learning about AI to think about my biological

10:18neural network,

10:19because that's a form of reverse engineering. All of the software was created

10:25from us. So we can use

10:27their software and their algorithms to try to understand it by thinking about

10:32our biological

10:32version. So we can say when we're thinking about classic AI, that would be

10:39clear rules. If this

10:40happens, do that. If you're driving a car, you're probably going to see a red

10:44light. And that means

10:46stop. But when you're thinking like machine learning, this would be like

10:50learning that touching a hot

10:51pan hurts. So probably don't touch it again. Now when you're thinking about

10:56deep learning,

10:57this might be more subtle, like recognizing a pet that's in a different

11:02lighting. So you're

11:04in a darker surrounding at a park, but you can still recognize that that is

11:08your pet. And then

11:09finally generative AI, we can use what we've learned, like writing a poem

11:14painting a picture

11:16or coding from scratch. This allows us to create new things. So all of this is

11:21really not as foreign

11:23as it might seem. It really is a very intimate connection with biological

11:28neural networks and

11:29software modeled after the brain. Finally, we have some key machine learning

11:34terms to remember.

11:36So again, is a generative adversarial network. And it creates new synthetic

11:42data by learning to

11:44mimic the training data. You can think of it as a forger or a twin maker. And

11:50this is good for

11:51creating fake faces or synthetic art. And this is an example of that. If you

11:56look down here,

11:57it's a style again. This is definitely a fake face, because the website is

12:01called this person

12:03does not exist. So every time I reload this, you're going to get a different

12:07person. This is interesting,

12:08because this is a very young person, but they seem to have facial hair here. So

12:12you can kind of tell

12:13that's you don't usually see those giveaways. And these are great examples of a

12:17generative

12:18adversarial network. Next, we have GPT. And we talked about that because we

12:24looked at the

12:25transformer architecture, and we can think about chat GPT. Next, we have BERT.

12:31This is

12:31bi-directional encoder representation from Transformers. And this is very

12:36similar to chat GPT. But I

12:39would think of it more of a reading comprehension expert. It's really good for

12:44search classification

12:45and summaries. And then we have WaveNet. I like to think of WaveNet as a sound

12:50designer,

12:51because it's good for voice synthesis and music. And here we have ResNet, which

12:57I like to think of

12:57as a vision specialist. And it's a deep neural network built for recognizing

13:03images,

13:04detecting objects, and facial recognition. Next, we have XG Boost. And I like

13:09to think of this for

13:11tabular data prediction, because it's a fast and efficient version of gradient

13:16boosting,

13:16which is used for structured data. So data inside of tables. And then RNN, we

13:22talked about this

13:23briefly, which is a recurrent neural network. We were talking about the

13:28transformer model,

13:30and how good it is at understanding the context of sentences. We said that the

13:35older model,

13:36the RNN, wasn't as good, because it processes sequences like text or time-sene

13:42arest data,

13:43but it did so one at a time. And it's also useful in speech and forecasting,

13:49but it's not as good as

13:50the GPT because of the transformer architecture. And last but not least, we

13:56have the support

13:58vector machine, which is a classical algorithm. And you can think of this as a

14:02decision maker.

14:04It's good for binary classification. And it's mostly associated with early

14:09machine learning,

14:10used to classify data or make predictions. All right, so that's it for this

14:15video. And I will

14:16see you in the next video, which is the challenge.

0:00Congrats on making it to the end of the skill and the challenge. This has been

0:04a

0:04fun skill for me. I hope you enjoyed it as well. There's so many mind-blowing

0:09concepts and AI tools and versions of artificial intelligence in this skill

0:15one after the other. So let's go ahead and switch gears. Right now we're gonna

0:20check out Google LM and we're going to check out the transformer architecture

0:25the white paper and try to create a podcast and it's something that's very

0:30useful for learning. Let me go ahead and break it down now. Alright so here we

0:35are notebook.google.com and all you need to do is search for attention is all

0:40you

0:40need and download this PDF and try to figure out how to make a podcast using

0:46notebook LM. It should be pretty straightforward and then after you do this

0:51I'll see you in the solution video in case you get stuck and there'll be some

0:55quiz questions to answer in the rest of the challenge. See you in the solution

0:00Alright, congrats on completing the challenge.

0:02And now let's go ahead and create a new notebook.

0:06I'm going to upload that attention is all you need PDF, white paper.

0:11And there it is.

0:12And you can do a bunch of stuff.

0:15You can actually type here and give it some more instructions or ask it

0:20questions.

0:21Or you can customize this audio overview.

0:25So you want to shorter.

0:26Let's do a shorter one.

0:28So you can add things here like you can say focus on a specific source or maybe

0:32target audience like

0:34explain it to me like I'm five years old.

0:37You can do all kinds of stuff.

0:38So let's just leave it like that.

0:40I'm just going to say shorter and then generate.

0:42And it'll take a few minutes and then you'll have a podcast and you can

0:46download it and you can

0:47share it with people.

0:48And this can be a useful educational assistant as well.

0:52Okay, until next time, I hope this has been informative and I'd like to thank

0:57you for viewing.