00:00with use cases from predicting the stock

00:02market to curing canter

00:03machine learning and ai are already

00:05revolutionizing our world as we know it

00:07a term i've heard that really

00:08encapsulates a machine learning process

00:10is a machine learning is a science of

00:12getting computers to act without being

00:14explicitly programmed

00:15in this video we're going to cover the

00:17components that make up machine learning

00:19break down the different models and

00:20algorithms and review common use cases

00:23in cyber security and i t

00:25before we go any further this is a

00:26perfect moment to hit like on the video

00:28to give me a boost in the youtube

00:30it will also help you to identify videos

00:32that interest you by reinforcing the

00:33machine learning algorithm

00:35please also subscribe if you're

00:36interested in seeing more tech and cyber

00:39now let's dive into machine learning

00:42let's start by clarifying the ai versus

00:44machine learning terminology

00:46artificial intelligence or ai is a

00:48general term we use when we talk about

00:50machines making decisions

00:52this could be as simple as making

00:53recommendations on your next song on

00:56to predicting stock market prices based

00:58on massive amounts of data

01:00machine learning is one approach or

01:03that is rooted in statistical algorithms

01:05to make data-driven decisions

01:08deep learning is a further subset of

01:10machine learning and it uses neural

01:11networks to use reason

01:13in determining whether or not machine

01:14learning output was correct

01:16and adjusted if necessary at its core

01:19machine learning are algorithms that

01:21parse data they learn from data and then

01:23apply what they've learned to make

01:26more specifically algorithms take data

01:29along with training that it has received

01:31to provide an output or decision

01:33over time the training process improves

01:36the overall quality of each

01:37output each piece of information that is

01:40learned feeds into the machine learning

01:41model in a continuous cycle of

01:43output and retraining there are three

01:46important components to a machine

01:48and they include data algorithm and

01:52by definition training consists of

01:54learning some new piece of information

01:56whether that be by someone teaching you

01:59that is why an important part of the

02:00training process involves how

02:02information is going to be learned by

02:04the different models

02:05and there are different models based on

02:06the use case four of the most common

02:09include supervised unsupervised

02:13and reinforcement training in supervised

02:16learning the machine is given

02:18some prior knowledge what is good or bad

02:20what works what doesn't

02:22a good example this would be image

02:23classification where we train the

02:25machine to detect a dog from a series of

02:28so that it can learn what a dog is

02:30likewise if we wanted the machine to

02:32a cat we need to train the machine with

02:35a series of pictures of cats so that it

02:38a dog versus a cat from an image with

02:41unsupervised training the machine has no

02:44and it's essentially starting from a

02:45blank state it still has

02:47data to work from but it has no trained

02:49concept of what good or bad is what a

02:52this kind of learning is used frequently

02:54in algorithms like regression analysis

02:56where you're taking massive amounts of

02:58past data to predict

02:59a future value unsupervised learning is

03:01good for finding hidden gems in data

03:03where we don't know exactly what we're

03:05looking for but we want to find

03:06anomalies or relationships that we

03:09semi-supervised is a combination of

03:11these two learning approaches because

03:12you're giving the machine some data to

03:14but ultimately it has to make its own

03:16decisions a good example of this is

03:18facial recognition on video

03:20where the machine needs a known picture

03:23to match a face to a video unsupervised

03:27lastly reinforcement training is about

03:28having a machine learn to make a

03:30sequence of decisions based on trial and

03:32reward or penalties reinforce good or

03:36on its ultimate goal of solving some

03:39autonomous cards are a good example of

03:40reinforcement training because a

03:42programmer cannot account for every

03:44possible scenario on the road and it

03:46must leave it to the machine to make

03:47decisions as it learns them

03:49this gives algorithms the flexibility to

03:51make decisions on the fly based on

03:52feedback like proximity sensors

03:54or objects on the road at this point

03:57it's important to note that how the

03:58machine learns the data

04:00is an important part of the training

04:01process that makes up the overall

04:03machine learning model

04:04but models themselves are not to be

04:06confused with algorithms

04:08as we mentioned previously as you train

04:10an algorithm with data it becomes a

04:13so to put it another way models equal

04:16times algorithm plus data the more data

04:19and time that runs the algorithm

04:21the more fine-tuned it becomes which

04:22lead to better decisions

04:24and just like there are different

04:25learning methods for different use cases

04:27so too are there different algorithms

04:29based on the need we're going to review

04:31five of the most common algorithms here

04:33classification regression recommendation

04:37dimensionality reduction and clustering

04:40let's start off with a classification

04:42algorithm which as its name indicates

04:44is focused on classifying or

04:47in a cyber security context this type of

04:49algorithm is one of the most common

04:51because the machine is classifying what

04:54clean or malicious this kind of

04:56algorithm is common in email servers to

04:58spam messages part of the learning

05:00process would be to train the algorithm

05:02based on examples of

05:03actual spam messages for example having

05:06unknown email addresses

05:07having too many exclamation marks or

05:09sentences that just don't sound natural

05:11the training process would also include

05:13giving the algorithm examples of real

05:15emails for example emails that are

05:17personally addressed clear language etc

05:20classification algorithms are also being

05:21a used lot in the sim space where the

05:24is taking in data from devices on your

05:26network learning what is normal behavior

05:30outliers unlike traditional sims without

05:32machine learning where rules are static

05:34here it would be learning and getting

05:37as it learns more data from your network

05:40clustering algorithm works in a similar

05:41way but for a different purpose

05:43with clustering we're focused on finding

05:45similarities between data points and

05:47grouping or clustering them together the

05:49main difference between classification

05:52is a cluster identify similarities

05:54between objects whereas

05:56classification uses predefined classes

05:59in other words clustering automatically

06:00finds and groups similarities whereas

06:02classification puts data into these

06:04pre-configured classes

06:06say for example you have an e-commerce

06:08site and you want to classify user

06:10traffic for marketing purposes

06:12based on their cookie or traffic

06:13information you can classify them into

06:15groups or clusters like new customers

06:17high-income earners low-income earners

06:20once a user has been classified it's up

06:23to you to decide what you want to do

06:25like pass it on to a recommendation

06:26algorithm to make recommendations based

06:28on that income or demographics

06:30clustering has many different use cases

06:32and it's common to use in combination

06:34with other algorithms using our previous

06:37once we have our users grouped together

06:39we can use a next algorithm to make

06:41calculated recommendation

06:42as its name indicates recommendation

06:44algorithms are focused on making

06:46recommendations based on past data

06:48the logic is simple based on past data

06:51the recommendation algorithm finds

06:53trends that people who bought or

06:55viewed x will also buy or view why

06:58algorithms like this are a multi-billion

07:00dollar industry and used by youtube

07:02amazon and facebook to make calculated

07:04recommendations based on who you are

07:06and what the recommendation algorithm

07:07thinks will get you engaged

07:09the recent documentary on netflix called

07:12talks in depth about this kind of

07:14recommendation algorithm and the dangers

07:15that come with it for a society at large

07:18next we have regression algorithm which

07:20is focused on predicting values based on

07:23or put another way the knowledge about

07:25existing data is utilized to predict new

07:28for example say you had details of every

07:30house sale over the last 10 years

07:32this includes square footage zip code

07:34number of beds sale price and so on

07:36the past data can be used as the input

07:39into the regression algorithm

07:41which is then used to predict future

07:42price values and trends

07:44and this kind of algorithm is very

07:46popular in medicine stock market and

07:47real estate but the reality is that

07:49every industry can benefit from this

07:52credit card companies and banks have

07:53been using this for fraud detection for

07:55a long time where your historical

07:58as the data that is learned when

07:59something out of the norm is detected

08:01an alert goes off that a purchase may be

08:03fraudulent regression algorithm is great

08:06for predicting based on past data that

08:07we have identified as important

08:09but what if we don't know what is

08:11important what if we have a lot of data

08:13points and we're not really sure

08:14what's significant or we want to

08:18dimensionality reduction is exactly for

08:20this kind of use case

08:21it finds outliers or significant factors

08:24from a very large data set

08:26let's take the stock market again as an

08:28example if you've ever looked at a

08:29fundamental breakdown of the stock

08:31you'll see that there are a lot of data

08:32points that can have a significant

08:34impact into the performance of the price

08:37dimensionality reduction can take all of

08:39these various data points across an

08:42over a specified amount of time and help

08:45identify important data points

08:47the same concept can be used in thread

08:50a 12 sudden increase in cpu usage may

08:53just be another windows update on a pc

08:55but when combined with other data points

08:57like the fact that the user also went to

08:59an uncategorized url

09:00has an old version of firefox an

09:02unpatched version of windows

09:04it could also point to something more

09:06serious dimensionality reduction is

09:08designed to find these patterns of

09:09suspicious behavior in a mountain of

09:12in a way that simply cannot be matched

09:14by human hands or static sim rules

09:16as we wrap up the algorithm discussion

09:18it's important to note that algorithms

09:20are not mutually exclusive

09:22it's common to have two or more

09:23algorithms combined or ensemble to

09:25produce a final verdict

09:27now let's put it all together by looking

09:29at a high level machine learning flow

09:31data comes in as the input into the

09:34machine learning model

09:35depending on the decision that needs to

09:36be made about the data a model is

09:39as we saw previously a model is made up

09:42times algorithm plus data the more data

09:45that runs through the algorithm

09:46the more a model is ultimately trained

09:49our data is now outputted as a decision

09:51which is sent off to some other

09:52application which then decides what it

09:54wants to do with that decision

09:56taking one of the most complex subjects

09:58in modern time and bringing it down in

10:00less than 10 minutes is almost

10:01impossible to do it justice

10:03but the idea here is for you to grasp

10:04the overall concept of the different

10:06components within machine learning to

10:08how it all comes together for our next

10:11video we're going to take a deep dive

10:12into machine learning use cases and

10:14examples of how it's already shaping the

10:17of it and cyber security so that does it

10:20for this video you guys and i hope you

10:22found it informative

10:23please subscribe if you want to see more

10:24videos about it and cyber security

10:27and if you haven't already please take a

10:28moment to hit like on the video as it

10:30greatly helps the youtube algorithm

10:32until next time this is andy with the

10:34cso perspective stay safe