Computer Scientist Explains Machine Learning in 5 Levels of Difficulty

Hi, I'm Hilary Mason. I'm a computer scientist.

And today, I've been asked to explain machine learning

in five levels of increasing complexity.

Machine learning gives us the ability to learn things

about the world from large amounts of data

that we as human beings can't possibly study or appreciate.

So machine learning is when we teach computers

to learn patterns from looking at examples in data,

such that they can recognize those patterns

and apply them to new things that they haven't seen before.

[playful music]

Hi.

Hi.

I'm Hilary, what's your name?

I'm Brynn.

Do you know what machine learning means?

Have you heard that before?

No.

So machine learning is a way that we teach computers

to learn things about the world by looking at patterns

and looking at examples of things.

So can I show you an example

of how a machine might learn something?

Sure.

[Hilary] So is this a dog or a cat?

It's a dog.

And this one?

A cat.

And what makes a dog, a dog and a cat, a cat?

Well, dogs are very playful, I think, more than cats.

Cats lick themselves more than dogs, I think.

That's true.

Do you think, if we look at these pictures,

do you think maybe we could say,

Well, they both have pointy ears,

but the dogs have a different kind of body

and the cats like to stand up a little different.?

Do you think that makes sense?

Yeah. Yeah.

What about this one?

A dog.

A cat.

I think, a cat?

Because it's more skinny.

And also, its legs are like really tall

and its ears are a little pointy.

This one's a jackal. And it's actually a kind of dog.

But you made a good guess.

That's what machines do too. They make guesses.

Is this a cat or a dog?

[Brynn] None.

[Hilary] None. What is it?

It's humans.

And how did you know that it's not a cat or a dog?

Because cats and dogs...

Because they walk on their paws

and their ears are like right here, not right here,

and they don't wear watches.

And so, you did something pretty amazing there.

Because we asked the question, Is it a cat or a dog?

And you said, I disagree with your question. It's a human.

So machine learning is when we teach machines

to make guesses about what things are

based on looking at a lot of different examples.

And I build products that use machine learning

to learn about the world and make guesses

about things in the world.

When we try to teach machines to recognize things

like cats and dogs, it takes a lot of examples.

We have to show them tens of thousands

or even millions of examples

before they can get even close to as good at it as you are.

Do you have tests in school?

Yeah, I have.

After every unit, we have a review and then we have a test.

Are those like the practice problems

you do before the test?

Well, just like everything that's gonna be on the test

is on the review.

Which means that in the test,

you're not seeing any problems

that you don't know how to solve.

As long as you did all your practice, right?

Yeah.

So machines work the same way.

If you show them a lot of examples and give them practice,

they'll learn how to guess.

And then when you give them the test,

they should be able to do that.

So we looked at eight pictures

and you were able to answer really quickly.

But what would you do if I gave you 10 million examples?

Would you be able to do that so quickly?

No.

So one of the differences between people and machines

is that people might be a little better at this,

but can't look at 10 million different things.

So now that we've been talking about machine learning,

is this something you want to learn how to do?

Kind of.

Because I kind of want to become a spy.

And we used to do coding,

so I may be kind of good at it.

And machine learning is a great way to use

all those math skills, all those coding skills,

and would be a super cool tool for a spy.

[quirky music]

Hello.

Hi. Are you a student, Lucy?

Yes, I just finished ninth grade.

Congratulations.

Thank you. It's very exciting.

Have you ever heard of machine learning before?

I'm going to assume that it means humans being able

to teach machines or robots how to learn themselves?

That's right.

When we teach machines to learn from data,

to build a model from that data or a representation of that,

and then to make a prediction.

One of the places we often find machine learning

in the real world is in things like recommendation systems.

So do you have an artist that you really like?

Yeah, Melanie Martinez.

So I'm gonna look up Melanie Martinez.

And it says here, If you like Melanie Martinez,

one of the other songs you might like is by Au/Ra.

Do you know who that is?

I do not.

So let's listen to a hint of this song.

Okay.

[alternative pop music]

All right.

So why do you think Spotify might've recommended that song?

Well, I know that in Melanie Martinez's music,

she used a lot of the filtered voice

to make it sound very deep and low

and that song had that.

And that's actually a really interesting thing

to think about because that creepy vibe

is something that you can perceive and I can perceive,

but it's actually really hard to describe to a machine.

What do you think might go into that?

Pitch of the music.

If it's really low or if it's super high,

it could know that.

What can the machine understand?

It's a great question.

The machine can understand

whatever we tell it to understand.

So there might be a person thinking about things,

like the pitch or the pacing or the tone,

or sometimes machines can figure out

things about music or images or videos

that we don't tell it to discover,

but that it can learn

from looking at a lot of different examples.

Why do you think companies might use machine learning?

Well, I think things like Facebook or Instagram,

they probably use it to target ads.

Sometimes, the ads you see are really uncanny.

And I think that's because they're based on so much data.

They know where you live. They know where your device is.

It's also important to realize that people in aggregate

are actually pretty predictable.

Like when we talk to each other,

we like to talk about the novel things,

like here, we're having this conversation.

We don't do this every day.

But we probably still eat breakfast.

We're gonna eat lunch. We're gonna eat dinner.

You probably are going to the same home

you go to most of the time.

And so, they're able to take that data

that we already give them and make predictions based on that

as to what ads they should show us.

So, you're saying I give them enough data as it is

about what I might be talking about or thinking about

that they can read my mind,

[Hilary laughs]

but just use the data that I've already given them.

And it almost seems like

they're watching us. That's right.

To do machine learning, we use something called algorithms.

Have you heard of algorithms before?

A set of steps or a process

carried out to complete something?

That's right.

So do you think that we've been able

to teach machines enough

so that they can do things that even we can't do?

And on the opposite side of that,

do you think there are things that we can do

that a machine might never be able to do?

So there are things that machines are really great at

that humans are actually not great at.

And imagine watching every video posted to TikTok every day.

So we just don't have enough time to do that

at the rate at which we can actually watch those videos.

But a machine can analyze all of them

and then make recommendations to us.

And then thinking about things that machines are bad at

and people are good at, people are really great

with only one or two examples of learning something new

and incorporating that into our model of the world

to make good decisions.

Whereas machines often need tens of thousands of examples,

and that's not even getting into things like good judgment

because we care about people,

because we can imagine a future that we want to live in

that doesn't exist today.

And that's something that is still uniquely human.

Machines are great at predicting

based on what they've seen in the past,

but they're not creative.

They're not going to invent.

They're not gonna, you know,

really change where we're gonna go.

That's up to us.

[serene music]

I'm Sunny.

[Hilary] And what are you majoring in?

I study Math and Computer Science.

So in your studies,

have you learned about machine learning?

Yeah, I have.

So to me, machine learning is essentially

exactly what it sounds like.

It's trying to teach a machine specifics about something

by inputting a lot of data points

and slowly, the machine will build up knowledge

about it over time.

For example, my Gmail program,

I assume that there would be a lot of, like,

machine learning models happening at once, right?

Absolutely.

And that's a great example because you have models

that are operating to do things like figure out

if a new email is spam or not.

So what would you think

about if you were looking at an email

and trying to decide if it went in one category or another?

I'd probably look at certain keywords.

Maybe if the recipient and the sender

had exchanged emails before

and generally, those fell into in the past.

So these are things we would call features.

And we go through a process where we do feature engineering,

where somebody looks at the example and says,

Okay, these are the things that I think might allow us

to statistically tell the difference

from something in one category versus another.

So for example, perhaps you don't speak Russian,

you start getting a lot of email in Russian.

Obviously, like the features that you just described

are features which a person would have had to think about.

Are there features

which, like, the machine itself could learn?

This is a great question

because it really gets to the difference

between some of our different tools

in our machine learning tool belt

in addressing problems like this.

So if we were to use a supervised learning classic

classification approach,

a person would need to think about those features

and creatively come up with them

in approach we call the kitchen sink approach,

which is just try everything you can possibly think of

and see what works.

Unsupervised learning, where we don't have labeled data

and we're trying to infer some structure out of the data

is you're projecting that data into a space

and looking for things like clusters.

And there's a bunch of really fun math

about how you do that, how you think about distance

and by distance, I mean that if we have two data points

in space, how do we decide if they're similar or not?

And how do the algorithms themselves usually differ

between unsupervised and supervised learning.

Supervised learning, we have our labels

and we're trying to figure out what statistically indicates

if something matches one label or another label.

Unsupervised learning,

we don't necessarily have those labels.

That's the thing we're trying to discover.

So reinforcement learning is another technique

that we use sometimes.

You can think about it like a turn in a game

and you can play, you know, millions and millions of trials

so that you're able to develop a system

that by experimenting with reinforcement learning

can eventually learn to play these games

pretty successfully.

Deep learning, which is essentially using neural networks

and very large amounts of data to eventually iterate

on a network structure that can make predictions.

With reinforcement learning versus deep learning,

it seems to me that reinforcement learning,

is it sort of like the kitchen sink approach

that you were talking about earlier,

where you're just kind of trying everything?

It is, but it also thrives in environments

where you have a decision point,

a pallet of actions to choose from.

And it actually comes historically

from trying to train a robot to navigate a room.

If it bonks into this chair, it can't go forward anymore.

And if it falls into that pit,

you know, it's not going to succeed.

But if it keeps exploring, it'll eventually get to the goal.

Oh, like roombas?

[Hilary] Yes.

[both laugh]

Oh, I didn't realize it was that deep, almost.

Is there a situation which you'd want to use

a deep learning algorithm

over a reinforcement learning algorithm?

So typically, you would choose deep learning

if you have sufficient high quality data,

hopefully labeled in a useful way.

If you really are happy not to necessarily understand

or be able to interpret what your system is doing

or you're willing to invest

in another set of work afterwards to understand

what the system is doing once you've already trained it.

And this also comes down to the fact that some things

are actually really easy to solve with linear regression

or simple statistical approaches.

And some things are impossible.

What would be the outcome if you were to choose

the, quote-unquote, wrong approach?

You build a system that could actually be useless.

So years ago, I had a client that was a big telecom company

and they had a data scientist

who built a deep learning system to predict customer churn.

It actually was very accurate, but it wasn't useful

because nobody knew why the prediction was what it was.

So they could say, you know,

Sunny, you're likely to quit next month.

But they had no idea what to do about it.

And so, I think there are a bunch of failure modes.

Would that be an example of, like, linear regression

where the regression is accurate, but,

you know, for marketing purposes, it's like,

if you don't know why I'm quitting the service,

then how can we fix this?

Yeah.

This is actually a good example of a very real world

kind of machine learning problem where the solution to this

was to build an interpretable system

on top of the accurate predictions not to throw it away,

but to do a bunch more work to figure out the why.

How can we improve machine learning algorithms?

It's actually fairly new

that we're able to solve all of these problems

and start to build these products and apply it in businesses

and apply it, you know, everywhere.

And so, we're still developing good practices

and what it means to be a professional in machine learning.

We're really developing a notion of what good looks like.

[quirky music]

I'm in my first year of a PhD in Computer Science

and I'm studying natural language processing

and machine learning.

So would you mind telling me a bit about

what you've been working on or interested in lately?

I've been looking at understanding persuasion

in online text and the ways that we might be able to

automatically detect the intent behind that persuasion

or who it's targeted at

and what makes effective persuasive techniques.

So what are some of the techniques you're applying

to look at that debate data?

Something I'm interested in exploring

is how well it works to use deep learning

and sort of automatically extracted features from this text

versus using some of the more traditional techniques

that we have, things like lexicons

or some sort of template matching techniques

for extracting features from texts.

That's a question I'm just interested in, in general.

When do we really need deep learning

versus when can we use something

that's a little bit more interpretable,

something that's been around for a while?

Do you think there are going to be general principles

that guide those decisions?

Because right now, it's generally

up to the machine learning engineer to decide

what tools they want to apply.

I definitely think there is,

but I also, sort of, see it varying a lot

based on the use case,

something that, kind of, works out of the box

and maybe works a little bit more automatically

might be better.

And in other cases, you do, sort of, kind of,

you want a lot of fine grain control.

So is that where some of that frustration

around the lack of controllability

and interpretability comes from?

Yeah, if you're building a model

that just predicts the next thing

based off of everything it's seen from texts online,

then yeah, you're really gonna be replicating

whatever that distribution online is.

If you train a model off of language off the internet,

it sometimes says uncomfortable things

or inappropriate things and sometimes really biased things.

Have you ever run into this yourself?

And then how do you think about that problem

of potentially even measuring the bias

in a model that we've trained?

Yeah, it's a really tricky question.

As you said, these models are trained to, sort of, predict

the next sequence of words,

given a certain sequence of words.

So we could start with just, sort of, prompts

like the woman was versus the man was,

and, kind of, pull out common words

that are, sort of, more used

with one phrase versus the other.

So that's, sort of, a qualitative way of looking at it.

It's not ever kind of a guarantee of how the model

is gonna behave in one particular instance.

And I think that's what's really tricky

and that's why I, sort of, think it's really good

for creators of systems to just be honest

about, This is, sort of, what we have seen.

And so then, someone can make their own judgment about,

Is this gonna be too high risk

for, sort of, my particular use case?

I imagine in the last few years,

we've seen a lot of changes and improvements

in the capabilities of NLP systems.

So is there anything in that

that you're particularly excited about exploring further?

I'm really interested in, sort of, the creative potential

that we've started to see from NLP systems

with things like GPT-3

and other really powerful language models.

It's really easy to write long grammatical passages

thinking about the way that we can then harness, like,

the human ability to actually give meaning to those words

and, sort of, provide structure

and how we can combine those things with the, kind of like,

generative capabilities of these models now

is really interesting.

Yeah, I agree.

[quirky music]

So, hi Claudia. It's so great to see you.

It has been far too long.

You know, we first met 10, 11 years ago

and machine learning has changed a lot since then.

Tooling that we now have, the capacity,

and also, an elevation of the problem sets

that we're dealing with and how to frame the problem.

And I'm almost struggling to figure out

whether it's a blessing or curse that it has become

as accessible and as democratized and as easy to execute

and you just build another new company from scratch.

And so, what's been, kind of, your reflection on that?

Well, you're absolutely right that the attention

machine learning gets has grown dramatically.

20 years ago, going to gatherings

and telling people what I was working on

and seeing the blank face or the like,

Where's the turn? and walk away.

Like, Oh, no.

The accessibility of the tooling,

like, we can now do in, like, five lines of code

something that would have taken 500 lines

of very mathematical, messy, gnarly code

even, you know, five years ago.

And that's not an exaggeration.

And there are tools that mean that pretty much anyone

can pick this up and start playing with it

and start to build with it.

And that is also really exciting.

In contrast, what I'm struggling with,

the friend of mine who asked me

to look at some health care data for him.

And despite the capabilities that we're having

in all of the, kind of, bigger societal problems

alongside with data collection engineering,

all the gnarly stuff,

that is actually not the machine learning itself,

it's the rest of it where certain data isn't available.

And to me, it's staggering how difficult it is

to get it off the ground and actually use.

And part of the challenge of it

is not the mathematics of building models,

but the challenge is making sure that the data

is sufficiently representative, potentially high quality.

And how transparent do I need to build it

for it to be adopted at some point?

What types of biases in the data collection,

and then also in the usage?

We now call it the bias, but we're still struggling

with the society not really living up to its expectations

and then machine learning, bringing it to the forefront.

Right.

And so, to say that another way,

when you're collecting data from the real world

and then building machine learning systems

that automate decisions based on that data,

all of the biases and problems

that are already in the real world then can be magnified

through that machine learning system.

And so, it can make many of these problems much worse.

Feeling increasingly challenged

that my skillset of being very good at programming

has become somewhat secondary.

And it's feeling...

[both laugh]

It's really the bigger picture understanding

of Who would be using that?

How transparent do I need to build it

for it to be adopted at some point?

What types of biases in the data collection

and then also in the usage?

I think, in certain areas, we have societal expectations

as to what is fair and what isn't.

And so, it's not just the provenance of that data,

but it's, sort of, deeply understanding,

Why does it look the way it looks?

Why was it collected this way?

What are the limitations of it?

We need to think about that

in entire process, how we document that process.

This is an issue in companies

where somebody might create something

that even their peers can't recreate.

What have you seen in terms of which industries,

where they stand, like who is adopting now?

Who is ready to utilize it?

Where would you maybe wish that didn't even try?

[Hilary laughs]

These are great questions.

So things like actuarial science, operations research,

where they actually are not using machine learning

as much as you might think.

And then you have other sorts of companies

or on the FinTech side, or even the ad tech side of things

where they perhaps are using machine learning

to the point of even absurdity.

So I spent about eight years working in ad tech.

And the motivation was really

because it was such an amazingly exciting playground

to push that technology

that used to largely live in academia, really,

out in the world and see, kind of, what it can achieve.

It has created such a hunger for data

that now everything is being collected.

I'm curious, when are we going to

make a foray into things like agriculture

about smart production of the things we eat?

You see and hear these interesting stories,

but I feel like we're not ready yet

to put that into a economically viable situation.

So when we think about the next five to 10 years,

the things that are really still holding us back

are these uneven applications of resources to problems

because the problems that get attention

are the high value ones

in terms of how much money you can make

or the things that are fashionable enough

that you can publish a paper on it.

So what do you think is holding us back?

I fully agree on the steps you pointed out

and the processes.

I think there is a chicken and egg problem,

like your former example,

that these areas that need to wait for data,

the value of the data collection

is then also slightly less apparent.

And so, it gets delayed further

and you'll see that happening.

But what my experience has been,

there's unfortunately, I feel a drifting apart

between academia

and the uses of AI,

but I am somewhat frustrated with a generation of students

who have standard data sets that they never think about

what the model needs to be used for,

that they never have to think about

how the data was collected.

So with all of these challenges ahead of us,

how optimistic are you about this world

that I deeply believe we can create

and the steps towards it?

I am incredibly optimistic and not...

Perhaps it's a personality flaw, but I can't help but look

at the potential of the technology to reduce harm,

to give us information that help us make better decisions.

And to think that we would choose

to address the big problems ahead of us.

I don't think we have a hope of addressing them

without figuring out the role

that machine learning will play.

And to think that we would then choose not to do that

is just unthinkable.

Despite that the rightfully raised concerns

about the challenges ahead,

but I think they also make us as a society better.

They challenge us to be a lot clearer

of what fairness means to all of us.

So with all of the setbacks,

I think we have exciting years to come.

And I am looking forward to a world where a lot more of that

is used for the right purposes.

[gentle upbeat music]

I hope you learned something about machine learning.

There has never been a better time to study machine learning

because you're now able to build products

that have tremendous potential and impact

across any industry or area that you might be excited about.