Hi I'm Artemy Kolchinsky. In these

tutorials we'll talk about what machine

learning is examples of machine learning

problems and some of the reasons why

it's become so prevalent in the news

let's just jump in and talk about what

machine learning is first you might be

curious how it differs from another term

you've heard which is artificial

intelligence I would say that artificial

intelligence is the general science of

creating intelligent automated systems

this could be things like chess playing

software software for controlling robots

or industrial processes and many many

other examples machine learning is a

subset of artificial intelligence and

it's particularly aimed at designing

software which can automatically learn

to perform intelligent things from

usually data that's provided to it and

from this point of view in some ways is

actually very closely related to

statistics because statistics is also

concerned about making inferences from

data and using them to predict other

things now to get a kind of schematic

idea of the difference between these two

we might consider two different

artificial intelligence systems one of

them is what's called the traffic

avoidance the traffic collision

avoidance system T caste and it's

basically a system that was developed

many decades ago to help planes avoid

each other and avoid mid-air collisions

this is essentially long list of if

statements and conditions that says when

to trigger alarms when to tell one plane

to go up and another plane to go down

and so on and these are hard coded rules

that were designed to help avoid

collisions this is not a machine

learning problem because it's

essentially some hard coded program now

I'm an example of what I would say is a

machine learning problem is the

following given some data make a

statistical model of something like

personal annual income depending on

things like age gender years in school

geographical zip code and so on and you

these what would be called independent

variables to make a prediction about

let's say how much money a person makes

in a year now if you've taken a

statistics class you might think that

this is just could be solved with a

simple linear regression and it turns

out that that's true that's one way to

approach the problem and simple linear

regression is a simple case of machine

learning it's using data to build a

model that then can be used to make

predictions about other things if you

want a more cinematic example of the

difference between AI and machine

learning you can think of the computer

in 2001 a Space Odyssey Hal 9000 which

was seemed like he was essentially

programmed to carry out a single-minded

task of the space mission and that would

be a classic AI example compare that to

the computer in the 1980s classic war

games which was trained to learn to play

global thermonuclear war and then could

learn to play things like tic-tac-toe

and that would be an example of a

machine learning a AI okay moving on I'm

gonna talk about two major areas of

artifice of machine learning supervised

learning and unsupervised learning in

supervised learning there's essentially

a right answer to a question and we

train the algorithm using but showing it

what the right answers are examples of

this might be the following I give you

an image and I ask you to classify it as

either cat or dog there's clearly a

right answer for most images or I give

you an audio recording and I ask you to

parse the text that was actually said in

that audio recording or finally I might

give you some kind of stock market

historical data and I ask you to predict

what the stock market will be now the

stock market will reach a certain level

and so we can be right or wrong in this

tutorial we'll mainly talk about

supervised learning however you should

also know about this other area of

machine learning called unsupervised

learning and in unsupervised learning

there's no right answer per se but what

we wish to do is find meaningful

patterns in the data this might be

something like take some very high

demand

dataset and represent it in a low

dimensional space it might be something

like take a dataset and cluster it into

groups so that's another example you see

on the screen where you can we visually

can see that there's a scatter plot and

there's some grouping there what is the

right grouping well there might not be

one right grouping but we still would

like to consider things it also includes

things like outlier detection anomaly

detection where it's not exactly clear

what the right answer is but it's still

useful to have some kind of pattern

finding algorithm in addition to

supervised learning and unsupervised

learning there's many many other

mixtures and variations and alternatives

to these such as transfer learning

semi-supervised learning and so on we

won't have time to get into most of

these but you should know that this the

domain areas of machine learning are

very large but supervised learning is

probably the one that's had the most

successes recently and it's the one

we'll talk most about okay so how does

supervised learning actually work in

supervised learning we're given what's

called a training data set I'm gonna use

the example of taking images and mapping

them to a label dog or cat depending on

what's shown in the image our training

data set would then consist of a whole

bunch of images and for each of those a

label is this a dog or is this a cat now

we take that training data set and we

take what's called a statistical model

or generally some kind of machine

learning model and we take a learning

algorithm and we'll talk about what

happens in this process when they're

combined but when they're combined we

say that we train the model and the

outcome of this is a trained model okay

so this train model has learned

something about the association between

images and the label dog or cat and once

we've trained it we can actually give

this train model an image it has never

seen before and it's going to make a

prediction is this a dog or a cat and if

we've done a good job then it should be

right about its prediction and we'll

talk about that

more later okay so just one thing that's

really important to remember is that

we've talked about some examples like a

classic one learning whether images or

dogs or cats but there's many many many

problems in the world that can be seen

as supervised learning problems for

example playing a video game

automatically learning how to play a

video game is an example of a supervised

learning problem we can think of the

supervised learning problem is learning

and mapping from the current state of

the video game screen to what should my

next move be should it be left right

jump or fire for example and one way

that such models are trained in a term

that you will hear is reinforcement

learning reinforcement learning is a

type of supervised learning in which

essentially the algorithm performs many

run throughs through some kind of

sequence of events and at the end of

these sequence of events it gets a

positive or negative reward and it

actually changes how it runs through the

sequence of events so as to maximize the

reward it's getting this will come up a

little bit later in the context of go

which we'll talk about in a later

tutorial another example of a supervised

learning problem is machine translation

so I may give you a sentence in English

and I might want your algorithm to learn

how to translate this to a sentence in

Spanish traditionally this has been a

very difficult task for supervised

learning and machine learning algorithms

but as we'll also talk about later

there's been actually a great deal of

success in this recently and one thing

you might hear in the context of machine

translation and similar tasks is

sequence learning sequence learning just

means that the thing that you're mapping

from or to is a long sequence of

something like for example it could be a

long sequence of words as in language

and you will also hear the terms

recurrent neural networks and things

like

short-term memory networks LST M's in

conjunction with sequence learning so

these are current neural network sirs

the reason that there's been a lot of

advances in language translation

recently actually finally the last

example I'll talk about is something

like a recommendation system so we might

want to take the set of songs that

you've listened to on some kind of music

listening up and now suggest a new song

that we think you'll like based on what

you've listened to you probably are all

familiar with these examples and they

actually can work surprisingly well in

recommending things you actually end up

liking so that's another example of a

supervised learning problem okay

so one of the last things I want you to

remember and take away from this

tutorial is that there's many many

algorithms that are doing supervised

learning here on the screen you see just

a small subset of these in recent times

you might have heard a lot about neural

networks or deep learning which are

actually an example of neural networks

you should remember that neural networks

are just one example of a machine

learning algorithm momentum among many

others they have like the other

algorithms their strengths and

weaknesses but we'll also talk later

about some of the reasons why they might

be so prevalent recently now in the next

tutorial I'll talk about a spatial or

geometrical way to understand how

machine learning actually does what it

does