AI and ML Fundamentals

0:00Now, AI is a broad term that is all the rage.

0:03It's used for a multitude of things.

0:05Some of those things are honestly just to sell more products,

0:09say it has AI in it, and then all of a sudden it's a new product,

0:12even if there's nothing artificial or intelligent as part of it.

0:16But in this skill, we're going to learn at the very least

0:18what Google means when they talk about AI and ML.

0:00So with Google Cloud, there's a lot of talk about artificial intelligence,

0:03usually referred to as AI and machine learning, which is usually just referred

0:08to as ML. So, ML and AI are going to be with this for a long time.

0:12And like I just said, it's, it's a AI, especially is just all the rage now

0:17because of things like chat GPT and, and like mid journey, all of these

0:24AI tools, which are part of the AI umbrella, if you will, are very much in the

0:29news and people are like starting to play with them. And you know, there's

0:33concerns about them, but there's also benefits with them.

0:35But when it comes to Google Cloud, you need to remember that AI is the all

0:40encompassing group of tools that live inside of it. So AI is the general

0:46umbrella term.

0:48And at least in the world of Google, when they talk about artificial

0:51intelligence, they basically mean this is kind of their definition. It's

0:55technology or tech that can mimic the cognitive function of human beings.

1:01So it's not just data analytics. That's something we covered in previous skills

1:06. AI mimics, again, it doesn't replace, but it mimics how humans

1:14use data and use data to do things like predicting and AI can do things like it

1:20can see things that the things that could see could be like images or video or

1:25text.

1:25So it sees it understands it responds, analyzes, predicts and recommends. Now,

1:33for some reason, Google really wants to point out that it is part of a system.

1:38It's not the system itself. So you need to build a system

1:42to use AI tools on your data, like just AI itself is not the finished system.

1:51AI is a tool that is used inside the system. And of course, with Google Cloud,

1:56it's going to be a bunch of different tools that are inside the Google Cloud

2:00world that are going to work together. And one of those tools could be, you

2:05know, an AI based tool to accomplish the tasks. And the real big thing to

2:10understand is that

2:12AI is going to have the ability to predict and recommend. Okay, so it's going

2:17to do more than just extrapolate data. Right, we have other ways of extrapol

2:22ating data. That's kind of how we've crunched numbers forever.

2:27So it's kind of like a step beyond. And sometimes it's really sometimes it's

2:32really subtle and hard to tell the difference. But there is a significant

2:36difference. And when it comes to AI versus ML, it's kind of like every thumb

2:41is a finger, but not every finger is a thumb. Hopefully that makes sense.

2:46Machine learning is a subset of artificial intelligence. Now, I have to admit,

2:51for a long time, I didn't think that thumbs were considered fingers.

2:54Which has nothing to do with Google Cloud. It was just, I thought that a thumb

2:58was not a finger, but apparently it is a thumb is a finger, but not all fingers

3:02are thumbs. So anyway, AI is the finger.

3:07Machine learning is the thumb. Now, when we talk about machine learning or ML,

3:11it's going to sound like I'm taught like what we just talked about with AI

3:15because, of course, it is AI. Machine learning is AI. And it's just that subset

3:21, right? Like the fingers and thumbs thing.

3:23It learns from data without explicit programming. And again, it's a subtle

3:30difference, but that's different from extrapolating.

3:36Like on a chart, like, oh, these are our sales over this time, this time, this

3:41time. If you see an upward trend in data, you can extrapolate where it's going

3:45to be next to your kind of thing. That's not what machine learning does.

3:49Machine learning takes all of the data and makes predictions that are more

3:54human-like.

3:54Like, well, because of this factor and that factor, when combined together, it

4:00can predict what might happen in any given situation. So not just like two data

4:06points, you know, how they correspond in a graph, it's different types of data.

4:12When put together, you can predict what the outcome will be or machine learning

4:17should is trying to predict what the outcome might be. And it uses really,

4:23really large sets of data.

4:25And then it uses previous results. So like, it will make predictions. And then

4:30if those predictions prove to be true or good predictions or bad predictions,

4:36it will then use that data to revise the next set of predictions that it makes.

4:42So that's what it means by learning, right? It's not like a conscious, you know

4:46, intelligence. It's just it takes the results. And then if those were good

4:51results, then it, you know, weights its internal

4:54processes in such a way so that they keep getting better. And if they're if the

4:58results or the predictions that makes are poor, then it will tweak it to avoid

5:03those types of predictions in the future. So that's how it learns. It just

5:09keeps doing it over and over and gets better by using the data of the outcome

5:14to be another input in the next iteration.

5:19And yeah, it can it can garner insight and advise on what to do again, using

5:24multiple different data sources. It's going to based on its previous iterations

5:29and its previous learning, if you will, give you answers.

5:33So machine learning crunches a ton of data generates data and then crunches

5:39that data as well. Now, we're going to talk a lot about machine learning

5:43because when it comes to data, which is pretty much what we're talking about,

5:47right?

5:47With Google Cloud and all of these digital transformations that we've been

5:50talking about, we're going to be crunching a lot of data.

5:53So machine learning is one of the main artificial intelligence tools that we're

5:58going to be talking about as we learn more about how to utilize Google Cloud.

6:02But there are other artificial intelligent things that are able to be done

6:08inside the Google Cloud world, so to speak.

6:11So when we think about other types of artificial intelligence technologies,

6:17apart from machine learning, or maybe to overuse our metaphor here, more thumbs

6:22, there are things like deep learning, which doesn't just make predictions, but

6:29it looks for patterns in data.

6:31So deep learning is kind of like tries to mimic how a human brain would

6:35identify patterns in different types of data, different data sets, different,

6:40you know, all sorts of data types.

6:42Deep learning is looking for the patterns that normally take a human brain to

6:47identify.

6:48There's also natural language processing, which is something that you're

6:51probably familiar with, like the large language models that are out there,

6:56things like chat, GP, there are other ways that you can use large language

7:00models.

7:00And one of those is natural language processing, which interestingly enough is

7:05something that Google offers in the Google workspace.

7:09Now remember Google workspace is software as a service.

7:13That is just the, you know, the umbrella of all the Google workspace stuff.

7:17But one of the things that Google can do is look inside your documents and then

7:22categorize them automatically for you by looking in there, seeing what type of

7:28information is in there.

7:29And then arranging it by again, looking deep into there, seeing the language,

7:35the text that is in there, and it actually reads it and understands what it

7:40says and then categorizes it accordingly.

7:43So that's something that is already implemented inside Google workspace.

7:46And then one that everybody is the most familiar with, at least usually is

7:52generative AI now generative AI.

7:54And that's why I have the picture, the robot here with a pencil. This is the

7:58type of AI that produces images or text or music.

8:02So things like mid journey, which will create images for you or, oh, what are

8:08the let's see, there's a Dali, I think that will do images.

8:12And there's a bunch of different AI tools that will generate images based on

8:17all the other image data that they've, you know, drawn in and used.

8:21There are some contention about if they had rights to that, but that's a whole

8:24nother. That's a whole nother skill.

8:26And then text, this is going to be the GPT chat, right, where it will create

8:32text based on all the language that it has consumed and compared to each other

8:39and, you know, remixed into a way that it can generate text that is usually

8:46very easy to read.

8:47It is important to remember that just because something is easy to read does

8:51not mean that it is reliable data, right?

8:54The generative AI generates language you can understand and it will often pull

9:01from sources that are reliable, but that is not its goal.

9:06Right? Generative AI for text, its goal is to make clear readable text.

9:13So remember, it doesn't care if it's lying to you and it often will.

9:19So whatever you feed it, it's going to pull from.

9:22And if you just give it ideas, it's going to pull from a vast array of

9:26information that may or may not be accurate.

9:29So I just wanted to throw that in there.

9:31And then lastly, music.

9:32This isn't one that we see quite as often, but AI is able to generate amazing

9:37music.

9:38A lot of times you'll see this as background music in things because of

9:42copyright issues, right?

9:44If it's generated by AI, it's not going to be violating the copyright in and of

9:48itself.

9:49Yes, there are still concerns about if it has the right to the music that it

9:53used to learn how to make music,

9:55but usually though that music that is it's created isn't copyrighted itself.

10:00So you'll see that in places where people don't want to pay for copyrighted

10:04music.

10:04So that's generative AI, which is one more thumb in the hand of AI.

0:00Now, we have already looked at business intelligence that takes data, which is

0:05combined into our

0:06data warehouse and using previous data, looking back on previous data, it can

0:12forecast or

0:14predict in a slight degree trends that are going to go forward.

0:19It extrapolates past data and gives us an idea of what we could expect in the

0:25future.

0:26But it's based entirely on historic trends and historic data that we've put in

0:31there,

0:32and then it's just following a trajectory.

0:35With artificial intelligence or machine learning, we take that and go a little

0:40bit deeper.

0:41And it's so subtle, the difference that it's easy to get confused.

0:45So here's an example of something that would be a business analytics projection

0:50, right?

0:51Let's say an airport or an airplane company is looking at all the flights that

0:57are traveling

0:59to India.

1:00In the past, they can look at the number of vegetarian meals that are requested

1:07on flights

1:07to India and see that they tend to be higher.

1:11The percentage of vegetarians in India is much higher than other destinations

1:16they might

1:17have. So based on the requests from previous flights to India, on flights to

1:24India, they

1:25could stock more vegetarian meals knowing that there's going to be a higher

1:30demand.

1:31And that's pretty simple.

1:32But you can see where that comes from.

1:33And it's extremely useful data, right?

1:35That is data that is invaluable.

1:37So you don't have extra meals going bad and unhappy clients and all that sort

1:41of thing.

1:42But that's just a trend, right?

1:43Going to the past where people have requested things going to a certain

1:46destination with

1:47machine learning, it looks at lots and lots and lots of data from multiple

1:55locations and

1:56different aspects and things that aren't necessarily easily related together

2:03and can not only look

2:05at past trends, but combine the various types of data into a prediction model

2:11that can say

2:12based on all of this data that is just wide and varied in its sources, we

2:19predict that

2:20this is what could be expected based on these conditions.

2:23And here's an example that I have.

2:25Let's say we are a seed company and we want to make sure that customers who buy

2:30our grass

2:31seeds get the best yield possible for every aspect of who they are, all their

2:37demographic

2:38information, they get the best.

2:39Now, we could look at simple things like where are you in the world and what

2:46seeds grow best

2:47there based on past experiences and go forward and predict what might work.

2:52But there are so many other variables that we could take into place and using a

2:57tool like

2:57machine learning, we could combine those in a way that is a lot more difficult

3:04than just

3:05extrapolating existing data.

3:07For example, what if we had just an enormous amount of data about a particular

3:14user and

3:15the location, things like the variety of seeds, how much sunlight the location

3:21gets, how much

3:21water is applied or not applied or occurs naturally, how deep the seeds are s

3:27owed, if

3:27they're covered like with straw or if they're not covered with straw, the zones

3:31where the

3:32individual planting will take place, how fresh seeds are, how dense the seeds

3:38are, meaning

3:39like how many seeds per square inch of dirt, what time of year it's being

3:44planted, the

3:45soil type that it's being planted on are their pesticides.

3:49If so, what types of pesticides, all of these variables that maybe individually

3:54we could

3:55say, okay, if there's pesticides, then it tends to grow better than if there's

4:00not.

4:01And depending on how much sunlight there's like a curve, we're too much or just

4:05the right

4:06amount.

4:07But ML or AI, machine learning or artificial intelligence can take all of these

4:12sources.

4:13And based on data from past information that we feed into it, it can make a

4:19prediction

4:20on what the best way to grow grass in any particular location at any particular

4:27time would be.

4:29And then when we get the information back on how well it growed, or how well it

4:34grow,

4:35like, I don't know, leaf per square inch and leaf width and stock height, I don

4:44't know,

4:45I don't know, grass terms.

4:46But we could take all of that information and feed it back into artificial

4:50intelligence

4:51so that the next prediction is even more accurate based on all of the other

4:56variables and the

4:57results from previous predictions.

5:00So a big difference is that there is a vastly larger amount of data that AI and

5:06ML crunches

5:07through.

5:08And it does it in a way that isn't as straightforward as just projecting based

5:14on past trends.

5:15It's far more than extrapolation.

5:17It's making predictions and estimations on what will happen in the future based

5:24on many,

5:25many variables.

5:27So data analysis and business intelligence does do a small version of what we

5:33're talking

5:34about with predictive algorithms in that it takes past data and it extrapolates

5:41or based

5:42on a trend can tell you what will probably happen.

5:45But when it comes to artificial intelligence and machine learning, it's

5:49multiplied orders

5:50of magnitude more using more and more and more data, creating data and then

5:55using that

5:56data to make even more data and more accurate predictions.

6:00So it's recursive, it's huge, it's a larger set of data and it's going to be

6:05more of an

6:07accurate prediction rather than just showing you what the expected trend would

6:13be.

0:00Now, you can probably guess machine learning is a very powerful tool that can

0:05benefit many

0:06aspects of many businesses, you know, using data, creating data, sorting

0:12through that

0:12data to make predictions.

0:14But it's not always the best tool for every job.

0:17You know, I mean, you don't need machine learning to serve out web pages.

0:20I mean, that doesn't even make sense.

0:22But there are some use cases and like four specific use cases that Google

0:27spells out

0:27that machine learning is really, really good at at at crunching data and coming

0:34up with

0:35good prediction models that are going to be better than standard data analytics

0:40or business

0:41intelligence.

0:42Now, the first thing it's going to be really, really good at is replacing a

0:47rule based

0:48system.

0:49And what I mean by that is with traditional business intelligence, you might

0:54have a way

0:54that it could predict some things and it's going to just be a long list of if

0:59then, right?

1:01If our grass seed is fresh, then it will require less water.

1:07If it has water in excess, then we should use a different seed thing and and

1:14these branching

1:15if then statements have a really difficult time scaling, especially when

1:21multiple multiple

1:23aspects of the data are going to influence more than just one thing.

1:28It's tough to if then your way through an entire set of data when they're not

1:33all directly

1:34related in the same way.

1:36And that's where machine learning doesn't just do if then it does, you know, it

1:42's internal

1:43algorithms learning which aspects of things are weighed more and which of the

1:49multiple,

1:50you know, which if affects the multiple events of other data and it comes out

1:55with this prediction

1:56model that isn't just a flow chart, right?

2:00You can't flow chart your way through machine learning.

2:03It just does a much better job of using disparate and data that is tough to

2:11identify relationships

2:13between.

2:14It does that in a way that the traditional if then decision trees just can't do

2:19.

2:19So replacing those older rule based systems is really where it shines.

2:24It can also automate processes.

2:26Now automation is something that we do all over it whether it's in the cloud or

2:30in our

2:30own data centers or, you know, when we do DevOps or when we write bash scripts,

2:36if you're

2:36a Linux person, right?

2:38All of these automation things are not new, but machine learning can do them in

2:43new ways

2:44and provide automation where we could never automate things before.

2:49A perfect example is Google's cloud vision, which is one of their machine

2:54learning tools

2:55that can examine an image, which is unstructured data.

2:59I mean, it's like the very definition of unstructured data.

3:03Just look at this image and it can identify and sort and tag and rate based on

3:08different

3:09criteria, things that would normally take a human being to look at that image

3:14and say,

3:15this is a dog.

3:16This dog has brown fur.

3:19The fur is curly.

3:20Therefore it is probably a poodle, right?

3:23And the machine learning using again, Google's cloud vision is a way to add

3:28some structure

3:29to unstructured data.

3:31But the cool part is it can do it automatically because we could hire somebody

3:34to sit down

3:35and tag all of our images with all of the various things that it contains and

3:39then,

3:40you know, we could search on that and we could search our images.

3:42But using something like Google's cloud vision, we can automate the process,

3:47dump a full folder

3:48of images in and it will look at them, look at them using its beaty little

3:53robot eyes

3:53and determine what they are and sort them.

3:57And in a way that gets better and better, right?

3:58When you rate those, the results and say, hey, you, you were spot on here.

4:03This is not a dog.

4:04This is a goat.

4:07And then maybe it will realize after enough corrections like, oh, goat eyes

4:11look different

4:12than dog eyes.

4:14And so then it will add that to its way of identifying things in images.

4:20Again, that's just an example.

4:21I pulled out of the top of my head, but automating things gets a lot easier and

4:24you can do things

4:25like text to speech, the Google text to speech API.

4:30And this is more than just gimmicky, like writing text and having it say it,

4:34you can

4:34have your English written text and then have it spoken in multiple languages.

4:40Now we've all heard the, you know, the AI voices and stuff, but you can really

4:45have an

4:45effective way of automating translations using something like text to speech or

4:50speech to

4:51text or text to text.

4:54And you can automate translations in a way that isn't just the traditional look

4:59up a

4:59dictionary word and, you know, substitute it.

5:03I mean, we've all had really bad luck with online translation tools, but as

5:08machine learning

5:09gets better and better at translating with those large language models, it's

5:14going to

5:14allow us to translate back and forth into various languages automatically.

5:19Again, that automation and more reliably as time goes on.

5:24So yeah, the automation process is the concept of automation isn't new, but the

5:29things we

5:29can automate sure are and that is just awesome.

5:33Three, they specifically talk about understanding unstructured data and the

5:38cloud vision is a

5:38big part of that, right, where you look at an unstructured image and unstruct

5:42ured data

5:43of image, and then you add some semi structure to it and make it semi

5:46structured data with

5:47tags and identifiers and all of that sort of thing.

5:51But it's not just images, like we could do something like feed it a default

5:57inbox, you

5:57know, like say a company has, you know, help at widget code.com or whatever.

6:03And every incoming email can then be read, read with the AI, the machine

6:09learning, speedy

6:10little robot eyes and determine, okay, based on the content, who should this

6:16email go to?

6:17And then it can automatically be forwarded to technical support or sales or

6:24whatever the

6:25email is asking, machine learning can read it and route it so the customer can

6:30just send

6:31one email to the default email address.

6:34And even if you have internal turnaround or, you know, somebody's not working

6:39that day,

6:40we don't have to worry about, you know, what email address people are emailing

6:45to because

6:46machine learning can route it internally when it goes to just one default email

6:51address.

6:52So understanding unstructured data and then doing something with it is

6:56something machine

6:57learning is really, really good at.

6:58And then this is something that is both a blessing and a curse.

7:03Personalization looks at a user.

7:05So when I go to YouTube and I watch a video, when I'm done watching that video,

7:11there is

7:11a recommendation for what video I might like next, right?

7:15If you're a YouTuber, it's the algorithm.

7:18Well, the algorithm is machine learning.

7:20Looking at your likes and dislikes based on, you know, how recently you liked a

7:25thing,

7:26if you were interested, how much of a thing you watched before you clicked away

7:30, using

7:30all of those various, sometimes hard to connect data points, it will determine

7:36what exact

7:37video you would probably want to watch next.

7:40And so personalized recommendations on things like, yeah, YouTube and Netflix

7:45does the same

7:46thing.

7:47Spotify will, once you play your playlist, it'll be like, hmm, so you liked

7:51everything

7:51in that playlist.

7:52I bet you'd also like this artist and it will start playing, you know,

7:56something else based

7:58on what it learned about your listening habits.

8:01So personalization can feel creepy, but also it can be really nice because you

8:06could actually

8:07really enjoy the things that it is curating for you using the things that it

8:13learned

8:13about what you have done previously.

8:16And again, it's not just a trend of you liked this artist, you will now listen

8:20to this artist.

8:22It's like, well, you liked this artist during the winter on sunny days in your

8:30area when

8:31it was three days past a full moon.

8:36I don't know, I'm making up data points now, but all these various data points

8:39can be

8:39worked together in a way to predict not just find a trend, but predict what is

8:44going to

8:45be ideal for your particular moment in time right then.

8:51So yeah, there are some real great use cases, even if it's not the perfect tool

8:54for everything,

8:55there are some great use cases for machine learning, regardless of what your

9:00business

9:01is, you know, it might be something you don't ever recommend things for your

9:06clients, but

9:07it would sure be nice to automate some of those things.

9:09You know, when people email, it'd be great to have it look at that unstructured

9:13email

9:13and route it to the proper place.

9:15So there are lots of things that can do and certainly there are going to be

9:19aspects of

9:19your cloud computing experience that are going to benefit from adding machine

9:24learning into

9:25your quiver to switch metaphors to an arrow into your toolbox so that you can

9:32get the

9:33best use out of the data that you're collecting and storing in your data

9:37warehouse.

0:00We know machine learning needs a lot more data in order to make good

0:05predictions on whatever it is you're looking for to make predictions on.

0:10But not only does it need a lot of data, it's very important that it gets high

0:15quality data as well.

0:17Now simply put the reason that we need high quality data is because if you put

0:22garbage into the system, you're going to get garbage out of the system.

0:26But what that means is we don't want any unrelated data.

0:31Like we don't want data introduced into the machine learning algorithm that is

0:36not going to be related at all.

0:39So it's going to try to pull information or it's going to try to connect

0:44information that you feed into it that doesn't have anything to do with the

0:49actual data set that you want it to be working with.

0:53So if it's unrelated, it's only going to give it some false pretenses to form

0:58its predictions on.

1:00So we don't want unrelated data in the data set that it uses.

1:04Also, we don't want biased data.

1:07And this is a lot easier to do than you might think.

1:10I mean, create biased data.

1:12We think that we aren't biased, but everybody is biased.

1:17So we have to consciously go into how we create our data sets and where we pull

1:21our data from and how we pull that data in a way to make sure that it's as

1:26unbiased as possible.

1:28Like for example, let's say you're using an app that only has metric gathering

1:36for iOS and not for Android.

1:39Well, then all of your data is going to be biased for Apple users and not

1:45Android users.

1:46Or maybe it only gathers data for mobile clients and you also have clients that

1:53use desktop applications, in which case it's only going to be pulling data from

1:58mobile clients.

2:00And so it's going to be biased for people who are accessing your application

2:04using that mobile application.

2:06So we have to make sure that the data we're getting is unbiased and that it's

2:10related to whatever outcome we're hoping to determine.

2:14So again, garbage in garbage out.

2:16We want high quality data.

2:18Now, it's easy to say, hey, we of course want high quality data.

2:21All of our data is high quality.

2:23We're a high quality business.

2:25But what does high quality data actually mean?

2:28It doesn't mean that you're a good company.

2:30And so any data you gather is good data.

2:33No, high quality data has qualities that make it more useful and more accurate

2:39and beneficial and more likely to produce predictions that are going to be more

2:44likely to produce predictions.

2:44That are going to be useful for you in your business endeavors or for your

2:48customer experience, etc, etc.

2:50Google has identified six different ways that we can use to measure our data,

2:56like to make sure that it's high quality.

2:59We want to make sure that our data is complete.

3:02We want to check it for completeness because if we only have data for our North

3:08American customers, that doesn't give us a full picture of our client base.

3:13It's incomplete.

3:14Now, location could be part of a complete data set.

3:17But if our data only comes from a subset of our users, our data is not complete

3:23and we're not going to get accurate and useful predictions from the machine

3:27learning model.

3:28So we need to make sure that our data is complete.

3:31We need to make sure that our data is unique.

3:35And by unique, I mean, we don't want to have duplicates in there, right?

3:39We want to make sure that every record in our set or every piece of information

3:44is unique.

3:45If you have multiple entries of the same data, it's going to be weighed more

3:51heavily than the other pieces of data, even though it didn't happen twice, for

3:55example, right?

3:57If every time there's a customer that orders something and you add a new record

4:02as they are a new client named Sean Powers, well, your data is going to turn

4:07out the fact that it's not going to be a new client.

4:08And it's going to turn out the fact that you really need to market heavily to

4:12people named Sean Powers because my goodness, are you attracting people named

4:16Sean Powers?

4:17And it's going to use that incorrect or that low quality data to make poor

4:22business decisions on your behalf because, yeah, you're not attracting a bunch

4:28of Sean Powers is, you just have duplicate data that doesn't make it unique and

4:35makes it far less useful.

4:37And it will corrupt the algorithm as it's predicting information that you would

4:41like a timeliness. You want to make sure that you have information that is

4:46timely.

4:47Now, historical data is good as long as it is marked. So you don't want to have

4:53incorrect data in there.

4:55You want to make sure that the data you're feeding it is the most recent and

5:00the most up to date that you have.

5:03Because if you are basing decisions on something the way the world was back

5:08before the pandemic, it's probably going to be different than the way your data

5:13is represented now. So timeliness is important.

5:17And I wanted to mention that that doesn't mean historical data is bad. It just

5:29needs to have its place in your algorithm, right, and needs to, you need to

5:30make sure that things are time stamped and you need to have as current or up to

5:32date information as possible so that the predictions and the, the algorithms

5:37that are applied to your data represent the current time that we are living in

5:43now.

5:43Validity, it's important. It is vitally important that you have valid data.

5:49If you have, you know, international clients or data records that have dates on

5:55them, if some of those dates are month they year and some of them are day,

6:01month year because you have, you know, some clients from the US or some

6:07businesses in the US and some, you know, overseas.

6:10You need to make sure that you have your data entered properly because one you

6:15're going to have errors in your data right there is no 19th month of the year,

6:19but you want to make sure that the consistency in the validity I guess

6:24consistency is further down.

6:25You want to make sure that you have valid data kept in a way that is easy to

6:31make sure so make sure that your ranges like if you have ranges there in the

6:37expected range, you know, like if your data shows that Sean Powers purchased 7.

6:453 million ink pens on one day when normally he purchases one ink pen every three

6:53months.

6:54That's probably not valid data, right? I don't generally go through 7.3 million

6:59ink pens in a three month period. So you need to make sure that your data is

7:03checked for validity, you know, make sure that there's not errors that are

7:07getting introduced into your data collection processes.

7:10You need to make sure that it is accurate. Obviously, you want to make sure

7:14that I really do buy one pen every three months and you don't just have guesses

7:17going in there like, "Oh, Sean, well, when he comes in, what does he do? I don

7:22't know. I think he normally buys a pen.

7:25What if I never buy a pen? What if I'm the guy that buys mechanical pencils?"

7:28You know, you want to make sure that the data you're collecting and giving to

7:33the machine learning model is accurate data.

7:36And then lastly, consistency. Now, this goes alongside validity. I know I

7:41mentioned it a couple of, but consistency is like, if you have a chickpea, you

7:48know, the legume, right, a chickpea, and you call it a chickpea and you label

7:54it a chickpea in your inventory.

7:56And then somebody else orders garbanzo beans and they put garbanzo beans in the

8:02inventory alongside of your chickpea entry. Well, garbanzo beans and chickpeas

8:07are the same exact thing.

8:09So you are going to have inconsistent data when you don't really have, you know

8:15, a bin of garbanzo beans plus another bin of chickpeas. You have one bin of

8:21that particular food.

8:23You've just named it in two different ways and so your data is inconsistent. So

8:27make sure that your data is consistent as well.

8:30And consistency is not just for garbanzo beans, right? Make sure that when you

8:34collect somebody's name that you collected in the same way. You know, maybe

8:39there is a client called Sean Powers, and maybe there's another client called

8:42Sean M Powers.

8:44That's the same client and you need to make sure that you are collecting

8:49information about every aspect of your business or your organization in a way

8:54that it's consistent because the machine learning model isn't going to be able

9:01to suss out the difference.

9:01It's going to give you the garbage that you give it. If you feed a garbage, it

9:05's going to give you garbage out. So again, high quality data is vital and it's

9:11not just about quantity. When it comes to machine learning, it's about quality

9:17too.

0:00Responsible AI is going to mean something different depending on who you ask.

0:05However, Google has made it quite clear that they take the principles of how AI

0:11is used very seriously.

0:13In fact, they've set up a set of principles that they go by and since they have

0:20all the systems that we use,

0:22any systems that you create using Google's AI tools are going to have to follow

0:27the principles that they've laid out

0:29as they have designed their AI system.

0:31So Google's principles of AI are kind of a two part.

0:35What it should do and then what Google will not allow to happen with AI tools

0:40under their purview.

0:42So first of all, on the left here, we have the things that AI should do and

0:46that is AI should be socially beneficial.

0:49So whatever we do with AI should benefit society in some sense.

0:54Now, that doesn't mean that a company being more profitable is against the

0:59benefits of society.

1:01But in general, what they're saying is that AI should benefit society, business

1:06is being a part of that society.

1:08AI should avoid creating or reinforcing unfair biases.

1:14So we should not be using AI to use somebody's differences in a way that isol

1:22ates them.

1:23So we're avoiding the reinforcement of biases.

1:27It should be built and tested for safety to make sure that it's functioning how

1:31we expect it to in a safe way.

1:33Be accountable to people.

1:37This basically means that it has to be beneficial to people, but it has to also

1:43be something that people are in control of.

1:46We're not going to allow a feedback loop of AI to run itself.

1:54It has to be accountable and answerable to human beings.

1:58It needs to incorporate privacy design principles.

2:02It's easy sometimes to think, well, since it's not a person that's looking at

2:06this data, it doesn't matter if it's private data.

2:09It's just an AI.

2:11So it doesn't matter if AI sees it because it's not another person, so it's

2:17still private, but that's not the case.

2:19AI has to be designed in a way that incorporates privacy so that a user's

2:24private information isn't used in a way to manipulate them.

2:28It has to uphold high standards of scientific excellence, meaning that we're

2:35not just going to throw junk code into fill the gaps.

2:37It has to be done properly, and it has to be made available for uses that match

2:42the principles outlined above.

2:45So the use cases for AI are use cases that will fit into these things that we

2:53have discussed.

2:54It has to fit into that worldview, that way of approaching things.

3:00Now, Google then specifically says things that their AI tools will not be able

3:06to or allowed to do.

3:08So Google will not use AI when it's likely to cause harm.

3:12It's not going to allow systems to be put into place that will harm other

3:18people to make weapons or tech that injure people, like actually injure people.

3:23It's not going to be used for that.

3:24It's not going to be used in violation of international norms.

3:28So just because something might be standard here, wherever here is for you,

3:34whether that's, you know, whatever part of the world you're in, doesn't mean

3:38that that is going to be applicable elsewhere.

3:41So it's going to be used in a way that is applicable according to international

3:47norms and not just norms, but it's not going to be used in a way that violates

3:52the principles of law and human rights around the world.

3:56So again, this is how Google has kind of put drawn the line in the sand with

4:01how they are going to use their extremely powerful AI tools and how they're

4:07going to allow others to build their AI tools into their systems.

4:11These are the principles of AI that Google has committed to.

4:16Now, just because Google has committed to it doesn't mean that that is the law

4:21of the land.

4:22If you're using Google's infrastructure, these are the rules that are going to

4:26be followed.

4:27But AI in general, there's no like three laws of robotics, you know, in Isaac

4:32Asimov's world where every single AI system on the planet is going to adhere to

4:37these rules Google does encourage other organizations to adopt their principles

4:43because they feel that they have written something that is encompassing and

4:48fair and wise in the way that we are using.

4:52AI as it continues to advance day by day second by second millisecond by

4:56millisecond.

4:57So again, this is what we're talking about when we're talking about responsible

5:02AI, the responsible use of artificial intelligence in our systems and our

5:07applications and in our cloud computing.

5:10So again, Google is committed to the rules that we just looked at, but it's

5:14important to note that not necessarily every other organization is going to be

5:19committed to the exact same set of rules.

5:22Even if they have their own set of guidelines that they're going by, they're

5:26not necessarily going to match Google, even though Google has encouraged others

5:30to adopt their policies.

5:31Responsible AI also means that it's important to debug and improve the machine

5:37learning models.

5:38We need to be able to look at them and see if something is going wrong because

5:44if something is going wrong, it will continue to go more wrong because remember

5:50machine learning builds.

5:52It's on its past predictions.

5:54So if something is broken and needs to be debugged and isn't, it's going to get

5:59worse and worse and worse.

6:02And this need to debug machine learning models is going to make sense as we go

6:07further down this list about what responsible AI really means.

6:12We need to be able to adjust and tweak and fix and debug and modify these

6:17machine learning models that do a lot of heavy lifting.

6:21But we need to be able to understand how they're working so that we can debug

6:25them.

6:26It's vital that we continue to watch our data that is going into the machine

6:32learning models and detect biases and drift and gaps in the data that we have.

6:38So we need to make sure that we are following all of those things we looked at

6:42in the previous video.

6:43You know that our data is consistent and that it's accurate and all of those

6:47things.

6:47Otherwise, the predictions are going to drift again based on bad data rather

6:53than a bug in the machine learning model.

6:56The predictions will start to drift if our data is losing its consistency or

7:03loses its quality.

7:05So we need to detect and continually monitor the data that we're feeding into

7:10it.

7:10AI in order to be what Google calls responsible AI must have human

7:15understandable explanations of the model.

7:19We need humans to understand what the machine learning models are actually

7:26doing.

7:27It's OK if we can't do all of the things ourselves, especially in the amount of

7:34time that the machine learning models can like crunch through that data so fast

7:38.

7:38But we need to understand what is happening because if we don't understand what

7:42is happening, there's no way that we're going to be able to fix it or steer it

7:47when things start to drift or go awry or start to give us unreliable data.

7:52So we need to understand what's happening underneath the hood, so to speak.

7:56And Google has a specific term that they use for that.

8:00And every one of their AI tools comes with something that is called explainable

8:06AI.

8:07It's just it's a way of looking at the tools so that a human being can

8:11understand what the AI is actually doing.

8:15Because while we may have some doom and gloom notion about what the future of

8:24AI looks like when it comes to using Google's tools,

8:28we want it to be something that benefits everyone, us and our users.

8:33And in order to do that, we have to understand what it is actually doing.

8:38So while indeed it can be a touchy subject, I really appreciate that these

8:42concerns and these design standards and principles are put forward and laid

8:48right out so that we can understand the thought process that is going into the

8:54AI tools that we are integrating into our systems.

8:56And then we can think about how we're using those AI tools in the systems that

9:02we in turn our building.

0:00All right, just in case we got mired in the mud on answering some of these fill

0:03in the blanks,

0:04let's go through them together. While blank blank includes a wide variety of

0:08technologies

0:09in the skill, we looked mainly at one specific subset of that larger concept.

0:14Of course, that

0:14is going to be artificial intelligence. All right, AI includes tools like

0:25machine learning, deep

0:27learning and the subset of AI, which creates images, audio and video, or blank

0:34AI, which is

0:35generative. Let's see, while data analytics and business intelligence looks

0:42backward and

0:43data to make extrapolations on future potential outcomes, well, that was a

0:47mouthful. Machine learning

0:49uses a wider array of data sets to make blank one word plural. That's going to

0:57be predictions.

1:02All right, machine learning has several ideal use cases, including replacing

1:11rule based

1:11if then systems, automation, personalization, and the ability to understand

1:16blank data.

1:18The nice thing about it is that it can understand unstructured

1:24unstructured data. Correct. In order for data to be considered high quality, it

1:30must be

1:31complete, unique, timely, valid, accurate, and be recorded in the same way

1:35every time, which is

1:36called consistency. And lastly, Google follows itself created set of principles

1:47when developing

1:47and utilizing AI. All of their tools include blank blank, which makes the

1:53function of the

1:54particular tool human understandable. This one you might not have remembered,

1:58but they refer to it as

2:00explainable AI. So explainable AI. And there we go. Hopefully you got them all

2:08right as well.

2:09I hope this has been informative for you, and I'd like to thank you for viewing

2:12.