A.I.で癒される｜A.I.の時代 (Healed through A.I. | The Age of A.I.)

字幕表動画を再生する

You know how they say
there are two certainties in life, right?
Death and taxes.
Can't we get rid of one of those?
See, 100 years ago, life expectancy was only 45,
can you believe that?
Then by the 1950s, it was up to 65,
and today, it's almost 80.
Tomorrow, who knows? Right?
Healthcare has made huge progress.
We've eradicated epidemics that used to kill millions,
but life is fragile.
People still get sick, or pass away
for reasons that maybe
should be, someday curable.
What if we could improve diagnosis?
Innovate to predict illness instead of just react to it?
In this episode,
we'll see how machine learning is combating
one of the leading causes of blindness,
and enabling a son with a neurological disease
to communicate with his family.
AI is changing the way we think
about mind and body,
life and death,
and what we value most,
our human experience.
[fanfare music playing]
[announcer] ...and our other co-captain, Number 8! Tim Shaw!
[crowd cheering]
[John Shaw] We've lived with his football dream,
All the way back to sixth grade
when his coach said,
"This kid is gonna go a long way."
From that point on,
Tim was doing pushups in his bedroom at night,
Tim was the first one at practice.
Tim took it seriously.
[crowd screaming and cheering]
[whistle blows]
[announcer] Number 8, Tim Shaw!
[crowd cheering]
I don't know what they're doing out there,
and I don't know who they comin' to!
[Robert Downey Jr.] For as long as he can remember,
Tim Shaw dreamed of three letters...
N-F-L.
[whistle blows]
He was a natural from the beginning.
As a kid, he was fast and athletic.
He grew into 235 pounds of pure muscle,
and at 23, he was drafted to the pros.
His dream was real.
He was playing professional football.
[reporter] Hello, I'm with Tim Shaw.
You get to start this season right. What's it feel like?
It's that amazing pre-game electricity,
the butterflies are there, and I'm ready to hit somebody. You might wanna look out.
Hey, Titans fans, it's Tim Shaw here,
linebacker and special teams animal.
He loves what he does.
He says, "They pay me to hit people!"
[crowd cheering]
I'm here to bring you some truth,
a little bit of truth,
and so we'll call it T-Shaw's truth,
'cause it's not all the way true, but it's my truth.
[Tim Shaw speaking]
[Tim from 2015 interview] In 2012,
my body started to do things it hadn't done before.
My muscles were twitching, I was stumbling,
or I was not making a play I would have always made.
I just wasn't the same athlete,
I wasn't the same football player that I'd always been.
[Tim speaking]
[Downey] The three letters that had defined Tim's life up to that point
were not the three letters that the doctor told him that day.
A-L-S.
[Tim speaking]
Okay...
[Downey] A-L-S, which stands for "amyotrophic lateral sclerosis,"
is also known as Lou Gehrig's Disease.
It causes the death of neurons controlling voluntary muscles.
[Sharon Shaw] He can't even scratch his head...
Better yet?
...none of those physical things
that were so easy for him before.
He has to think about every step he takes.
So Tim's food comes in this little container.
We're gonna mix it with water.
[Tim speaking]
[Downey] As the disease progresses,
muscles weaken.
Simple everyday actions, like walking, talking, and eating,
take tremendous effort.
Tim used to call me on the phone in the night,
and he had voice recognition,
and he would speak to the phone,
and say, "Call Dad."
His phone didn't recognize the word "Dad."
So, he had said to me...
[voice breaking] "Dad, I've changed your name.
I'm calling... I now call you "Yo-yo."
So he would say into his phone, "Call Yo-yo."
[Sharon] Tim has stopped a lot of his communication.
He just doesn't talk as much as he used to,
and I, I miss that.
I miss it.
-What do you think about my red beard? -No opinion.
[snorts] That means he likes it,
just doesn't wanna say on camera.
Now, my favorite was when you had the handlebar moustache.
[Downey] Language, the ability to communicate with one another.
It's something that makes us uniquely human,
making communication an impactful application for AI.
[Sharon] Yeah, that'll be fun.
[Julie Cattiau] My name is Julie.
I'm a product manager here at Google.
For the past year or so, I've been working on Project Euphonia.
Project Euphonia has two different goals.
One is to improve speech recognition
for people who have a variety of medical conditions.
The second goal is to give people their voice back,
which means actually recreating the way they used to sound
before they were diagnosed.
If you think about communication,
it starts with understanding someone,
and then being understood,
and for a lot of people,
their voice is like their identity.
[Downey] In the US alone, roughly one in ten people
suffer acquired speech impairments,
which can be caused by anything from ALS,
to strokes, to Parkinson's, to brain injuries.
Solving it is a big challenge,
which is why Julie partnered with a big thinker to help.
[Downey] Dimitri is a world-class research scientist and inventor.
He's worked at IBM, Princeton, and now Google,
and holds over 150 patents.
Accomplishments aside,
communication is very personal to him.
Dimitri has a pretty strong Russian accent,
and also he learned English when he was already deaf,
so he never heard himself speak English.
Oh, you do? Oh, okay.
[Downey] Technology can't yet help him hear his own voice.
He uses AI-powered Live Transcribe
to help him communicate.
[Cattiau] Okay, that's awesome.
So we partnered up with Dimitri to train a recognizer
that did a much better job at recognizing his voice.
The model that you're using right now for recognition,
what data did you train it on?
[Downey] So, how does speech recognition work?
First, the sound of our voice is converted into a waveform,
which is really just a picture of the sound.
Waveforms are then matched to transcriptions,
or "labels" for each word.
These maps exist for most words in the English language.
This is where machine learning takes over.
Using millions of voice samples,
a deep learning model is trained
to map input sounds to output words.
Then the algorithm uses rules, such as grammar and syntax,
to predict each word in a sentence.
This is how AI can tell the difference
between "there," "their," and "they're."
[Cattiau] The speech recognition model that Google uses
works very well for people
who have a voice that sounds similar
to the examples that were used to train this model.
In 90% of cases, it will recognize what you want to say.
[Downey] Dimitri's not in that 90%.
For someone like him, it doesn't work at all.
So he created a model based on a sample of one.
[Downey] But making a new unique model
with unique data for every new and unique person
is slow and inefficient.
Tim calls his dad "Yo-yo."
Others with ALS may call their dads something else.
Can we build one machine
that recognizes many different people,
and how can we do it fast?
[Cattiau] So this data doesn't really exist.
We have to actually collect it.
So we started this partnership with ALS TDI in Boston.
They helped us collect voice samples
from people who have ALS.
This is for you, T. Shaw.
[all] One, two, three!
[all cheering]
I hereby accept your ALS ice bucket challenge.
[yelping softly]
[Downey] When the ice bucket challenge went viral,
millions joined the fight, and raised over $220 million for ALS research.
There really is a straight line from the ice bucket challenge
to the Euphonia Project.
ALS Therapy Development Institute is an organization
that's dedicated to finding treatments and cures for ALS.
We are life-focused.
How can we use technologies we have
to help these people right away?
Yeah, they're actually noisier.
That's a good point.
I met Tim a few years ago
shortly after he had been diagnosed.
Very difficult to go public,
but it was made very clear to me
that the time was right.
He was trying to understand what to expect in his life,
but he was also trying to figure out,
"All right, what part can I play?"
All the ice bucket challenges and the awareness
have really inspired me also.
If we can just step back,
and say, "Where can I shine a light?"
or "Where can I give a hand?"
When the ice bucket challenge happened,
we had this huge influx of resources of cash,
and that gave us the ability
to reach out to people with ALS who are in our programs
to share their data with us.
That's what got us the big enough data sets
to really attract Google.
[Downey] Fernando didn't initially set out
to make speech recognition work better,
but in the process of better understanding the disease,
he built a huge database of ALS voices,
which may help Tim and many others.
[John] It automatically uploaded it.
[Tim] Oh.
How many have you done, Tim?
2066?
[Fernando Vieira] Tim, he wants to find every way that he can help.
It's inspiring to see his level of enthusiasm,
and his willingness to record lots and lots of voice samples.
[Downey] To turn all this data into real help,
Fernando partnered with one of the people
who started the Euphonia Project, Michael Brenner...
-Hey, Fernando. -Hey, how are you doing?
[Downey] ...a Google research scientist
and Harvard-trained mathematician
who's using machine learning
to solve scientific Hail Marys, like this one.
Tim Shaw has recorded almost 2,000 utterances,
and so we decided to apply our technology
to see if we could build a recognizer that understood him.
[Tim speaking]
The goal, right, for Tim, is to get it so that it works
outside of the things that he recorded.
The problem is that we have no idea
how big of a set that this will work on.
[Brenner] Dimitri had recorded upwards of 15,000 sentences,
which is just an incredible amount of data.
We couldn't possibly expect anyone else
to record so many sentences,
so we know that we have to be able to do this
with much less recordings from a person.
So it's not clear it will work.
[Tim speaking]
-That didn't work at all. -Not at all.
He said, "I go the opposite way,"
and it says, "I know that was."
[Brenner] When it doesn't recognize,
we jiggle around the parameters of the speech recognizer,
then we give it another sentence,
and the idea is that you'll get it to understand.
[Tim's recording] Can we go to the beach?
-Yes! Got it. -Got it.
That's so cool. Okay, let's try another.
[Downey] If Tim Shaw gets his voice back,
he may no longer feel that he is defined,
or constrained, by three letters,
but that's a big "if."
While Michael and team Euphonia work away,
let's take a moment and imagine what else is possible
in the realm of the senses.
Speech.
Hearing.
Sight.
Can AI predict blindness?
[truck horn beeps]
Or even prevent it?
[Downey] Santhi does not have an easy life.
It's made more difficult because she has diabetes,
which is affecting her vision.
[Downey] If Santhi doesn't get medical help soon, she may go blind.
[Dr. Jessica Mega] Complications of diabetes
include heart disease,
kidney disease,
but one of the really important complications
is diabetic retinopathy.
The reason it's so important is that
it's one of the lead causes of blindness worldwide.
This is particularly true in India.
[giving instructions]
In the early stages, it's symptomless,
but that's when it's treatable,
so you want to screen them early on,
before they actually lose vision.
In the early stages,
if a doctor is examining the eye,
or you take a picture of the back of the eye,
you will see lots of those bleeding spots in the retina.
Today, the doctors are not enough to do the screening.
We are very limited ophthalmologists,
so there should be other ways
where you can screen the diabetic patients
for diabetic complications.
[Downey] In the US,
there are about 74 eye doctors for every million people.
In India, there are only 11.
So just keeping up with the sheer number of patients,
let alone giving them the attention and care they need,
is overwhelming, if not impossible.
[Dr. R. Kim] We probably see about 2,000 to 2,500 patients
every single day.
[Mega] The interesting thing with diabetic retinopathy
is there are ways to screen and get ahead of the problem.
The challenge is that not enough patients undergo screening.
[Downey] Like Tim Shaw's ALS speech recognizer,
this problem is also about data,
or lack of it.
To prevent more people from experiencing vision loss,
Dr. Kim wanted to get ahead of the problem.
So there's a hemorrhage.
All these are exudates.
[Downey] Dr. Kim called up a team at Google.
Made up of doctors and engineers,
they're exploring ways to use machine learning
to solve some of the world's leading healthcare problems.
So we started with could we train an AI model
that can somehow help read these images,
that can decrease the number of doctors required
to do this task.
So this is the normal view.
When you start looking more deeply,
then this can be a microaneurysm, right?
-This one here? -[man] Could be.
[Downey] The team uses the same kind of machine learning
that allows us to organize our photos
or tag friends on social media,
image recognition.
First, models are trained
using tagged images of things like cats or dogs.
After looking at thousands of examples,
the algorithm learns to identify new images
without any human help.
For the retinopathy project,
over 100,000 eye scans
were graded by eye doctors
who rated each eye scan on a scale from one to five,
from healthy to diseased.
These images were then used
to train a machine learning algorithm.
Over time, the AI learned to predict
which eyes showed signs of disease.
[Dr. Lily Peng] This is the assistant's view
where the model's predictions
are actually projected on the original image,
and it's picking up the pathologies very nicely.
[Downey] To get help implementing the technology,
Lily's team reached out to Verily,
the life sciences unit at Alphabet.
[Mega] So, how was India?
[Peng] Oh, amazing!
[Mega] Verily came out of Google X,
and we sit at the intersection of technology, life science, and healthcare.
What we try to do is think about big problems
that are affecting many patients,
and how can we bring the best tools
and best technologies
to get ahead of the problems.
The technical pieces are so important, and so is the methodology.
How do you capture the right image,
and how does the algorithm work,
and how do you deploy these tools not only here,
but in rural conditions?
If we can speed up this diagnosis process
and augment the clinical care,
then we can prevent blindness.
[Downey] There aren't many other bigger problems that affect more patients.
Diabetes affects 400 million worldwide,
70 million in India alone,
which is why Jessica and Lily's teams
began testing AI-enabled eye scanners there,
in its most rural areas,
like Dr. Kim's Aravind Eye Clinics.
-Is the camera on? -Now it's on.
Yeah.
So once the camera is up,
we need to check network connectivity.
[Sunny Virmani] The patient comes in.
They get pictures of the back of the eye.
One for the left eye, and right eye.
The images are uploaded to this algorithm,
and once the algorithm performs its analysis,
it sends the results back to the system,
along with a referral recommendation.
It's good. It's up and running.
Because the algorithm works in real time,
you can get a real-time answer to a doctor,
and that real-time answer comes back to the patient.
[Kim] Once you have the algorithm,
it's like taking your weight measurement.
Within a few seconds,
the system tells you whether you have retinopathy or not.
[Downey] In the past,
Santhi's condition could've taken months to diagnose,
if diagnosed at all.
[Downey] By the time an eye doctor would've been able to see her,
Santhi's diabetes might have caused her to go blind.
Now, with the help of new technology,
it's immediate,
and she can take the hour-long bus ride
to Dr. Kim's clinic in Madurai
for same-day treatment.
[Downey] Now thousands of patients
who may have waited weeks or months to be seen
can get the help they need before it's too late.
Thank you, sir.
[Downey] Retinopathy
is when high blood sugar damages the retina.
Blood leaks, and the laser treatment
basically "welds" the blood vessels
to stop the leakage.
Routine eye exams can spot the problem early.
In rural or remote areas, like here,
AI can step in and be that early detection system.
[Pedro Domingos] I think one of the most important applications of AI.
is in places where doctors are scarce.
In a way, what AI does is make intelligence cheap,
and now imagine what you can do when you make intelligence cheap.
People can go to doctors they couldn't before.
It may not be the impact that catches the most headlines,
but in many ways it'll be the most important impact.
[family chattering happily]
[Mega] AI now is this next generation of tools
that we can apply to clinically meaningful problems,
so AI really starts to democratize healthcare.
[Mega] The work with diabetic retinopathy
is opening our eyes to so much potential.
Even within these images,
we're starting to see some interesting signals
that might tell us about someone's risk factors for heart disease.
And from there, you start to think about
all of the images that we collect in medicine.
Can you use AI or an algorithm
to help patients and doctors
get ahead of a given diagnosis?
Take cancer as an example of how AI can help save lives.
We could take a sample of somebody's blood
and look for the minuscule amounts of cancer DNA
or tumor DNA in that blood. This is a great application for machine learning.
[Downey] And why stop there?
Could AI accomplish
what human researchers have not yet been able to?
Figuring out how cells work well enough
that you can understand why a tumor grows and how to stop it
without hurting the surrounding cells.
[Downey] And if cancer could be cured,
maybe mental health disorders,
like depression, or anxiety.
There are facial and vocal biomarkers
of these mental health disorders.
People check their phones 15 times an hour.
So that's an opportunity
to almost do, like, a well-being checkpoint.
You can flag that to the individual,
to a loved one,
or in some cases even to a doctor.
[Bran Ferren] If you look at the overall field of medicine,
how do you do a great job of diagnosing illness?
Having artificial intelligence,
the world's greatest diagnostician, helps.
[Downey] At Google,
Julie and the Euphonia team have been working for months
trying to find a way for former NFL star Tim Shaw
to get his voice back.
[Dimitri Kanevsky speaking]
Yes! So Zach's team, the DeepMind team,
has built a model that can imitate your voice.
For Tim, we were lucky,
because, you know, Tim has a career of NFL player,
so he did multiple radio interviews and TV interviews,
so he sent us this footage.
Hey, this is Tim Shaw, special teams animal.
Christmas is coming, so we need to find out
what the Titans players are doing.
If you gotta hesitate, that's probably a "no."
[Cattiau] Tim will be able to type what he wants,
and the prototype will say it in Tim's voice.
I've always loved attention.
Don't know if you know that about me.
[laughs] She's gonna shave it for you.
[Downey] Interpreting speech is one thing,
but re-creating the way a real person sounds
is an order of magnitude harder.
Playing Tecmo Bowl, eating Christmas cookies, and turkey.
[Downey] Voice imitation is also known as voice synthesis,
which is basically speech recognition in reverse.
First, machine learning converts text back into waveforms.
These waveforms are then used to create sound.
This is how Alexa and Google Home are able to talk to us.
Now the teams from DeepMind and Google AI
are working to create a model
to imitate the unique sound of Tim's voice.
Looks like it's computing.
But it worked this morning?
We have to set expectations quite low.
[Cattiau] I don't know how our model is going to perform.
I hope that Tim will understand
and actually see the technology for what it is,
which is a work in progress and a research project.
[Downey] After six months of waiting, Tim Shaw is about to find out.
The team working on his speech recognition model
is coming to his house for a practice run.
[doorbell rings]
[dog barks]
[Sharon] Good girl, come say hello.
-Hi! -Oh, hi!
Welcome.
-Hi! -Come in.
Thanks for having us.
[Sharon] He's met some of you before, right?
How are you doing, Tim?
-Hi, Tim. -Good to see you.
-Hello. -Hello.
Hi.
Hi, I'm Julie. We saw each other on the camera.
It's warmer here than it is in Boston.
[Sharon] As it should be.
[all laughing]
Okay.
Lead the way, Tim.
[Cattiau] I'm excited to share with Tim and his parents
what we've been working on.
I'm a little bit nervous. I don't know if the app
is going to behave the way we hope it will behave,
but I'm also very excited, to learn new things
and to hear Tim's feedback.
So I brought two versions with me.
I was supposed to pick, but I decided to just bring both
just in case one is better than the other,
and, just so you know,
this one here was trained
only using recordings of your voice,
and this one here was trained using recordings of your voice,
and also from other participants from ALS TDI
who went through the same exercise of... [laughing]
So, okay.
I was hoping we could give them a try.
Are we ready?
Who are you talking about?
[app chimes]
It got it.
[John] It got it.
[gasps]
[Tim] Is he coming?
[app chimes]
Yes.
Are you working today?
[app chimes]
[chuckling]
It's wonderful.
[Cattiau] Cool.
Thank you for trying this.
-Wow! -It's fabulous.
[John] What I love, it made mistakes,
-and then it corrected itself. -Yeah.
I was watching it like, "That's not it,"
and then it went... [mimics app] Then it does it right.
These were phrases,
part of the 70% that we actually used
to train the model,
but we also set aside 30% of the phrases,
so this might not do as well,
but I was hoping that we could try some of these too.
[John] So what we've already done
is him using phrases that were used to train the app.
That's right.
Now we're trying to see if it can recognize phrases
-that weren't part of that. -[Cattiau] Yes, that's right.
So let's give it a try?
Do you want me to?
Do you have the time to play?
[app chimes]
What happens afterwards?
[app chimes]
Huh. So, on the last one,
this one got it, and this one didn't.
-We'll pause it. So... -I love the first one, where it says,
-"Can you help me take a shower?" -[laughing]
-[Cattiau] That's not at all what he said. -[John] I know,
you've gotta be really careful what you ask for.
[all laughing]
[John] So if, when it's interpreting his voice,
and it makes some errors,
is there a way we can correct it?
Yeah. We want to add the option
for you guys to fix the recordings,
but as of today, because this is the very first time
we actually tried this,
we don't have it yet.
[Cattiau] This is still a work in progress.
We have a speech recognition model
that works for Tim Shaw,
which is, you know, one person,
and we're really hoping that, you know,
this technology can work for many people.
There's something else I want you to try,
if that's okay?
We're working with another team at Google called DeepMind.
They're specialized in voice imitation and synthesis.
[Downey] In 2019,
Tim wrote a letter to his younger self.
They are words written by a 34-year-old man with ALS
who has trouble communicating
sent back in time
to a 22-year-old on the cusp of NFL greatness.
[Cattiau] So let me give this a try.
I just like using this letter because it's just so beautiful,
so let me see if this is gonna work.
[Tim's younger voice] So, I've decided to write you this letter
'cause I have so much to tell you.
I want to explain to you
why it's so difficult for me to speak,
the diagnosis, all of it,
and what my life is like now,
'cause one day, you will be in my shoes,
living with the same struggles.
It's his voice, that I'd forgotten.
We do.
[app chimes]
[app chimes]
We're so happy to be working with you.
It's really an honor.
[John] The thought that one day,
that can be linked with this,
and when you speak as you are now,
it will sound like that, is...
It's okay. We'll wait.
[Cattiau] There is a lot of unknown
and still a lot of research to be conducted.
We're really trying to have a proof of concept first,
and then expand to not only people who have ALS,
but people who had a stroke,
or a traumatic brain injury, multiple sclerosis,
any types of neurologic conditions.
Maybe other languages, too, you know?
I would really like this to work for French, for example.
[Mega] Wouldn't it be a wonderful opportunity
to bring technology to problems that we're solving
in life science and healthcare,
and in fact, it's a missed opportunity
if we don't try to bring the best technologies
to help people.
This is really just the beginning.
[Downey] Just the beginning indeed.
Imagine the possibilities.
I think in the imaginable future for AI and healthcare
is that there is no healthcare anymore,
because nobody needs it.
You could have an AI that is directly talking to your immune system,
and is actually preemptively creating the antibodies
for the epidemics that are coming your way,
and will not be stopped.
This will not happen tomorrow, but it's the long-term goal
that we can point towards.
[Downey] Tim had never heard his own words
read out loud before today.
Neither had his parents.
[Tim] Every single day is a struggle for me.
I can barely move my arms.
[John] Have fun.
I can't walk on my own,
so I recently started using a wheelchair.
I have trouble chewing and swallowing.
I'd kill for a good pork chop.
Yes, my body is failing, but my mind is not giving up.
Find what's most important in your life,
and live for that.
Don't let three letters, NFL,
define you...
[crowd cheering]
...the same way I refuse to let three letters define me.
[John] One of the things Tim has taught us,
and I think it's a lesson for everyone...
Medically speaking, Tim's life has an end to it.
In fact, five years ago we were told he only had two to five years left.
We're already past that.
He has learned very quickly
that today is the day that we have,
and we can ruin today
by thinking about yesterday
and how wonderful it used to be,
and, "Oh, woe is me," and "I wish it was like that."
We can also ruin today
by looking into the future,
and in Tim's case,
how horrible this is going to be.
"This is going to happen,"
"I won't be able to do this anymore."
So if we go either of those directions,
it spoils us from being present today.
That's a lesson for all of us.
Whether we have an ALS diagnosis or not,
try to see the good and the blessing of every day.
You're here with us today.
It's going to be a good day.