back to indexRegina Barzilay: Deep Learning for Cancer Diagnosis and Treatment | Lex Fridman Podcast #40
link |
The following is a conversation with Regina Barzilay.
link |
She's a professor at MIT and a world class researcher
link |
in natural language processing
link |
and applications of deep learning to chemistry and oncology
link |
or the use of deep learning for early diagnosis,
link |
prevention and treatment of cancer.
link |
She has also been recognized for teaching
link |
of several successful AI related courses at MIT,
link |
including the popular Introduction
link |
to Machine Learning course.
link |
This is the Artificial Intelligence podcast.
link |
If you enjoy it, subscribe on YouTube,
link |
give it five stars on iTunes, support it on Patreon
link |
or simply connect with me on Twitter
link |
at Lex Friedman spelled F R I D M A N.
link |
And now here's my conversation with Regina Barzilay.
link |
In an interview you've mentioned
link |
that if there's one course you would take,
link |
it would be a literature course with a friend of yours
link |
that a friend of yours teaches.
link |
Just out of curiosity, because I couldn't find anything
link |
on it, are there books or ideas that had profound impact
link |
on your life journey, books and ideas perhaps
link |
outside of computer science and the technical fields?
link |
I think because I'm spending a lot of my time at MIT
link |
and previously in other institutions where I was a student,
link |
I have limited ability to interact with people.
link |
So a lot of what I know about the world
link |
actually comes from books.
link |
And there were quite a number of books
link |
that had profound impact on me and how I view the world.
link |
Let me just give you one example of such a book.
link |
I've maybe a year ago read a book
link |
called The Emperor of All Melodies.
link |
It's a book about, it's kind of a history of science book
link |
on how the treatments and drugs for cancer were developed.
link |
And that book, despite the fact that I am in the business
link |
of science, really opened my eyes on how imprecise
link |
and imperfect the discovery process is
link |
and how imperfect our current solutions
link |
and what makes science succeed and be implemented.
link |
And sometimes it's actually not the strengths of the idea,
link |
but devotion of the person who wants to see it implemented.
link |
So this is one of the books that, you know,
link |
at least for the last year, quite changed the way
link |
I'm thinking about scientific process
link |
just from the historical perspective
link |
and what do I need to do to make my ideas really implemented.
link |
Let me give you an example of a book
link |
which is not kind of, which is a fiction book.
link |
It's a book called Americana.
link |
And this is a book about a young female student
link |
who comes from Africa to study in the United States.
link |
And it describes her past, you know, within her studies
link |
and her life transformation that, you know,
link |
in a new country and kind of adaptation to a new culture.
link |
And when I read this book, I saw myself
link |
in many different points of it,
link |
but it also kind of gave me the lens on different events.
link |
And some of it that I never actually paid attention.
link |
One of the funny stories in this book
link |
is how she arrives to her new college
link |
and she starts speaking in English
link |
and she had this beautiful British accent
link |
because that's how she was educated in her country.
link |
This is not my case.
link |
And then she notices that the person who talks to her,
link |
you know, talks to her in a very funny way,
link |
in a very slow way.
link |
And she's thinking that this woman is disabled
link |
and she's also trying to kind of to accommodate her.
link |
And then after a while, when she finishes her discussion
link |
with this officer from her college,
link |
she sees how she interacts with the other students,
link |
with American students.
link |
And she discovers that actually she talked to her this way
link |
because she saw that she doesn't understand English.
link |
And I thought, wow, this is a funny experience.
link |
And literally within few weeks,
link |
I went to LA to a conference
link |
and I asked somebody in the airport,
link |
you know, how to find like a cab or something.
link |
And then I noticed that this person is talking
link |
in a very strange way.
link |
And my first thought was that this person
link |
have some, you know, pronunciation issues or something.
link |
And I'm trying to talk very slowly to him
link |
and I was with another professor, Ernst Frankel.
link |
And he's like laughing because it's funny
link |
that I don't get that the guy is talking in this way
link |
because he thinks that I cannot speak.
link |
So it was really kind of mirroring experience.
link |
And it led me think a lot about my own experiences
link |
moving, you know, from different countries.
link |
So I think that books play a big role
link |
in my understanding of the world.
link |
On the science question, you mentioned that
link |
it made you discover that personalities of human beings
link |
are more important than perhaps ideas.
link |
Is that what I heard?
link |
It's not necessarily that they are more important
link |
than ideas, but I think that ideas on their own
link |
are not sufficient.
link |
And many times, at least at the local horizon,
link |
it's the personalities and their devotion to their ideas
link |
is really that locally changes the landscape.
link |
Now, if you're looking at AI, like let's say 30 years ago,
link |
you know, dark ages of AI or whatever,
link |
what is symbolic times, you can use any word.
link |
You know, there were some people,
link |
now we're looking at a lot of that work
link |
and we're kind of thinking this was not really
link |
maybe a relevant work, but you can see that some people
link |
managed to take it and to make it so shiny
link |
and dominate the academic world
link |
and make it to be the standard.
link |
If you look at the area of natural language processing,
link |
it is well known fact that the reason that statistics
link |
in NLP took such a long time to become mainstream
link |
because there were quite a number of personalities
link |
which didn't believe in this idea
link |
and didn't stop research progress in this area.
link |
So I do not think that, you know,
link |
kind of asymptotically maybe personalities matters,
link |
but I think locally it does make quite a bit of impact
link |
and it's generally, you know,
link |
speeds up the rate of adoption of the new ideas.
link |
Yeah, and the other interesting question
link |
is in the early days of particular discipline,
link |
I think you mentioned in that book
link |
is ultimately a book of cancer.
link |
It's called The Emperor of All Melodies.
link |
Yeah, and those melodies included the trying to,
link |
the medicine, was it centered around?
link |
So it was actually centered on, you know,
link |
how people thought of curing cancer.
link |
Like for me, it was really a discovery how people,
link |
what was the science of chemistry behind drug development
link |
that it actually grew up out of dying,
link |
like coloring industry that people
link |
who developed chemistry in 19th century in Germany
link |
and Britain to do, you know, the really new dyes.
link |
They looked at the molecules and identified
link |
that they do certain things to cells.
link |
And from there, the process started.
link |
And, you know, like historically saying,
link |
yeah, this is fascinating
link |
that they managed to make the connection
link |
and look under the microscope and do all this discovery.
link |
But as you continue reading about it
link |
and you read about how chemotherapy drugs
link |
which were developed in Boston,
link |
and some of them were developed.
link |
And Farber, Dr. Farber from Dana Farber,
link |
you know, how the experiments were done
link |
that, you know, there was some miscalculation,
link |
let's put it this way.
link |
And they tried it on the patients and they just,
link |
and those were children with leukemia and they died.
link |
And then they tried another modification.
link |
You look at the process, how imperfect is this process?
link |
And, you know, like, if we're again looking back
link |
like 60 years ago, 70 years ago,
link |
you can kind of understand it.
link |
But some of the stories in this book
link |
which were really shocking to me
link |
were really happening, you know, maybe decades ago.
link |
And we still don't have a vehicle
link |
to do it much more fast and effective and, you know,
link |
scientific the way I'm thinking computer science scientific.
link |
So from the perspective of computer science,
link |
you've gotten a chance to work the application to cancer
link |
and to medicine in general.
link |
From a perspective of an engineer and a computer scientist,
link |
how far along are we from understanding the human body,
link |
biology of being able to manipulate it
link |
in a way we can cure some of the maladies,
link |
some of the diseases?
link |
So this is very interesting question.
link |
And if you're thinking as a computer scientist
link |
about this problem, I think one of the reasons
link |
that we succeeded in the areas
link |
we as a computer scientist succeeded
link |
is because we don't have,
link |
we are not trying to understand in some ways.
link |
Like if you're thinking about like eCommerce, Amazon,
link |
Amazon doesn't really understand you.
link |
And that's why it recommends you certain books
link |
or certain products, correct?
link |
And, you know, traditionally when people
link |
were thinking about marketing, you know,
link |
they divided the population to different kind of subgroups,
link |
identify the features of this subgroup
link |
and come up with a strategy
link |
which is specific to that subgroup.
link |
If you're looking about recommendation system,
link |
they're not claiming that they're understanding somebody,
link |
they're just managing to,
link |
from the patterns of your behavior
link |
to recommend you a product.
link |
Now, if you look at the traditional biology,
link |
and obviously I wouldn't say that I
link |
at any way, you know, educated in this field,
link |
but you know what I see, there's really a lot of emphasis
link |
on mechanistic understanding.
link |
And it was very surprising to me
link |
coming from computer science,
link |
how much emphasis is on this understanding.
link |
And given the complexity of the system,
link |
maybe the deterministic full understanding
link |
of this process is, you know, beyond our capacity.
link |
And the same ways in computer science
link |
when we're doing recognition, when you do recommendation
link |
and many other areas,
link |
it's just probabilistic matching process.
link |
And in some way, maybe in certain cases,
link |
we shouldn't even attempt to understand
link |
or we can attempt to understand, but in parallel,
link |
we can actually do this kind of matchings
link |
that would help us to find key role
link |
to do early diagnostics and so on.
link |
And I know that in these communities,
link |
it's really important to understand,
link |
but I'm sometimes wondering, you know,
link |
what exactly does it mean to understand here?
link |
Well, there's stuff that works and,
link |
but that can be, like you said,
link |
separate from this deep human desire
link |
to uncover the mysteries of the universe,
link |
of science, of the way the body works,
link |
the way the mind works.
link |
It's the dream of symbolic AI,
link |
of being able to reduce human knowledge into logic
link |
and be able to play with that logic
link |
in a way that's very explainable
link |
and understandable for us humans.
link |
I mean, that's a beautiful dream.
link |
So I understand it, but it seems that
link |
what seems to work today and we'll talk about it more
link |
is as much as possible, reduce stuff into data,
link |
reduce whatever problem you're interested in to data
link |
and try to apply statistical methods,
link |
apply machine learning to that.
link |
On a personal note,
link |
you were diagnosed with breast cancer in 2014.
link |
What did facing your mortality make you think about?
link |
How did it change you?
link |
You know, this is a great question
link |
and I think that I was interviewed many times
link |
and nobody actually asked me this question.
link |
I think I was 43 at a time.
link |
And the first time I realized in my life that I may die
link |
and I never thought about it before.
link |
And there was a long time since you're diagnosed
link |
until you actually know what you have
link |
and how severe is your disease.
link |
For me, it was like maybe two and a half months.
link |
And I didn't know where I am during this time
link |
because I was getting different tests
link |
and one would say it's bad and I would say, no, it is not.
link |
So until I knew where I am,
link |
I really was thinking about
link |
all these different possible outcomes.
link |
Were you imagining the worst
link |
or were you trying to be optimistic or?
link |
It would be really,
link |
I don't remember what was my thinking.
link |
It was really a mixture with many components at the time
link |
speaking in our terms.
link |
And one thing that I remember,
link |
and every test comes and then you're saying,
link |
oh, it could be this or it may not be this.
link |
And you're hopeful and then you're desperate.
link |
So it's like, there is a whole slew of emotions
link |
that goes through you.
link |
But what I remember is that when I came back to MIT,
link |
I was kind of going the whole time through the treatment
link |
to MIT, but my brain was not really there.
link |
But when I came back, really finished my treatment
link |
and I was here teaching and everything,
link |
I look back at what my group was doing,
link |
what other groups was doing.
link |
And I saw these trivialities.
link |
It's like people are building their careers
link |
on improving some parts around two or 3% or whatever.
link |
I was, it's like, seriously,
link |
I did a work on how to decipher ugaritic,
link |
like a language that nobody speak and whatever,
link |
like what is significance?
link |
When all of a sudden, I walked out of MIT,
link |
which is when people really do care
link |
what happened to your ICLR paper,
link |
what is your next publication to ACL,
link |
to the world where people, you see a lot of suffering
link |
that I'm kind of totally shielded on it on daily basis.
link |
And it's like the first time I've seen like real life
link |
and real suffering.
link |
And I was thinking, why are we trying to improve the parser
link |
or deal with trivialities when we have capacity
link |
to really make a change?
link |
And it was really challenging to me because on one hand,
link |
I have my graduate students really want to do their papers
link |
and their work, and they want to continue to do
link |
what they were doing, which was great.
link |
And then it was me who really kind of reevaluated
link |
what is the importance.
link |
And also at that point, because I had to take some break,
link |
I look back into like my years in science
link |
and I was thinking, like 10 years ago,
link |
this was the biggest thing, I don't know, topic models.
link |
We have like millions of papers on topic models
link |
and variation of topics models.
link |
Now it's totally like irrelevant.
link |
And you start looking at this, what do you perceive
link |
as important at different point of time
link |
and how it fades over time.
link |
And since we have a limited time,
link |
all of us have limited time on us,
link |
it's really important to prioritize things
link |
that really matter to you, maybe matter to you
link |
at that particular point.
link |
But it's important to take some time
link |
and understand what matters to you,
link |
which may not necessarily be the same
link |
as what matters to the rest of your scientific community
link |
and pursue that vision.
link |
So that moment, did it make you cognizant?
link |
You mentioned suffering of just the general amount
link |
of suffering in the world.
link |
Is that what you're referring to?
link |
So as opposed to topic models
link |
and specific detailed problems in NLP,
link |
did you start to think about other people
link |
who have been diagnosed with cancer?
link |
Is that the way you started to see the world perhaps?
link |
And it actually creates, because like, for instance,
link |
there is parts of the treatment
link |
where you need to go to the hospital every day
link |
and you see the community of people that you see
link |
and many of them are much worse than I was at a time.
link |
And you all of a sudden see it all.
link |
And people who are happier someday
link |
just because they feel better.
link |
And for people who are in our normal realm,
link |
you take it totally for granted that you feel well,
link |
that if you decide to go running, you can go running
link |
and you're pretty much free
link |
to do whatever you want with your body.
link |
Like I saw like a community,
link |
my community became those people.
link |
And I remember one of my friends, Dina Katabi,
link |
took me to Prudential to buy me a gift for my birthday.
link |
And it was like the first time in months
link |
that I went to kind of to see other people.
link |
And I was like, wow, first of all, these people,
link |
they are happy and they're laughing
link |
and they're very different from these other my people.
link |
And second of thing, I think it's totally crazy.
link |
They're like laughing and wasting their money
link |
on some stupid gifts.
link |
They already may have cancer and they don't understand it.
link |
So you can really see how the mind changes
link |
that you can see that,
link |
before that you can ask,
link |
didn't you know that you're gonna die?
link |
Of course I knew, but it was a kind of a theoretical notion.
link |
It wasn't something which was concrete.
link |
And at that point, when you really see it
link |
and see how little means sometimes the system has
link |
to have them, you really feel that we need to take a lot
link |
of our brilliance that we have here at MIT
link |
and translate it into something useful.
link |
Yeah, and you still couldn't have a lot of definitions,
link |
but of course, alleviating, suffering, alleviating,
link |
trying to cure cancer is a beautiful mission.
link |
So I of course know theoretically the notion of cancer,
link |
but just reading more and more about it's 1.7 million
link |
new cancer cases in the United States every year,
link |
600,000 cancer related deaths every year.
link |
So this has a huge impact, United States globally.
link |
When broadly, before we talk about how machine learning,
link |
when do you think we as a civilization will cure cancer?
link |
How hard of a problem is it from everything you've learned
link |
I cannot really assess it.
link |
What I do believe will happen with the advancement
link |
in machine learning is that a lot of types of cancer
link |
we will be able to predict way early
link |
and more effectively utilize existing treatments.
link |
I think, I hope at least that with all the advancements
link |
in AI and drug discovery, we would be able
link |
to much faster find relevant molecules.
link |
What I'm not sure about is how long it will take
link |
the medical establishment and regulatory bodies
link |
to kind of catch up and to implement it.
link |
And I think this is a very big piece of puzzle
link |
that is currently not addressed.
link |
That's the really interesting question.
link |
So first a small detail that I think the answer is yes,
link |
but is cancer one of the diseases that when detected earlier
link |
that's a significantly improves the outcomes?
link |
So like, cause we will talk about there's the cure
link |
and then there is detection.
link |
And I think where machine learning can really help
link |
is earlier detection.
link |
So does detection help?
link |
Detection is crucial.
link |
For instance, the vast majority of pancreatic cancer patients
link |
are detected at the stage that they are incurable.
link |
That's why they have such a terrible survival rate.
link |
It's like just few percent over five years.
link |
It's pretty much today the sentence.
link |
But if you can discover this disease early,
link |
there are mechanisms to treat it.
link |
And in fact, I know a number of people who were diagnosed
link |
and saved just because they had food poisoning.
link |
They had terrible food poisoning.
link |
They went to ER, they got scan.
link |
There were early signs on the scan
link |
and that would save their lives.
link |
But this wasn't really an accidental case.
link |
So as we become better, we would be able to help
link |
to many more people that are likely to develop diseases.
link |
And I just want to say that as I got more into this field,
link |
I realized that cancer is of course terrible disease,
link |
but there are really the whole slew of terrible diseases
link |
out there like neurodegenerative diseases and others.
link |
So we, of course, a lot of us are fixated on cancer
link |
because it's so prevalent in our society.
link |
And you see these people where there are a lot of patients
link |
with neurodegenerative diseases
link |
and the kind of aging diseases
link |
that we still don't have a good solution for.
link |
And I felt as a computer scientist,
link |
we kind of decided that it's other people's job
link |
to treat these diseases because it's like traditionally
link |
people in biology or in chemistry or MDs
link |
are the ones who are thinking about it.
link |
And after kind of start paying attention,
link |
I think that it's really a wrong assumption
link |
and we all need to join the battle.
link |
So how it seems like in cancer specifically
link |
that there's a lot of ways that machine learning can help.
link |
So what's the role of machine learning
link |
in the diagnosis of cancer?
link |
So for many cancers today, we really don't know
link |
what is your likelihood to get cancer.
link |
And for the vast majority of patients,
link |
especially on the younger patients,
link |
it really comes as a surprise.
link |
Like for instance, for breast cancer,
link |
80% of the patients are first in their families,
link |
And I never saw that I had any increased risk
link |
because nobody had it in my family.
link |
And for some reason in my head,
link |
it was kind of inherited disease.
link |
But even if I would pay attention,
link |
the very simplistic statistical models
link |
that are currently used in clinical practice,
link |
they really don't give you an answer, so you don't know.
link |
And the same true for pancreatic cancer,
link |
the same true for non smoking lung cancer and many others.
link |
So what machine learning can do here
link |
is utilize all this data to tell us early
link |
who is likely to be susceptible
link |
and using all the information that is already there,
link |
be it imaging, be it your other tests,
link |
and eventually liquid biopsies and others,
link |
where the signal itself is not sufficiently strong
link |
for human eye to do good discrimination
link |
because the signal may be weak,
link |
but by combining many sources,
link |
machine which is trained on large volumes of data
link |
can really detect it early.
link |
And that's what we've seen with breast cancer
link |
and people are reporting it in other diseases as well.
link |
That really boils down to data, right?
link |
And in the different kinds of sources of data.
link |
And you mentioned regulatory challenges.
link |
So what are the challenges
link |
in gathering large data sets in this space?
link |
Again, another great question.
link |
So it took me after I decided that I want to work on it
link |
two years to get access to data.
link |
Any data, like any significant data set?
link |
Any significant amount, like right now in this country,
link |
there is no publicly available data set
link |
of modern mammograms that you can just go
link |
on your computer, sign a document and get it.
link |
It just doesn't exist.
link |
I mean, obviously every hospital has its own collection
link |
There are data that came out of clinical trials.
link |
What we're talking about here is a computer scientist
link |
who just wants to run his or her model
link |
and see how it works.
link |
This data, like ImageNet, doesn't exist.
link |
And there is a set which is called like Florida data set
link |
which is a film mammogram from 90s
link |
which is totally not representative
link |
of the current developments.
link |
Whatever you're learning on them doesn't scale up.
link |
This is the only resource that is available.
link |
And today there are many agencies
link |
that govern access to data.
link |
Like the hospital holds your data
link |
and the hospital decides whether they would give it
link |
to the researcher to work with this data or not.
link |
Individual hospital?
link |
I mean, the hospital may, you know,
link |
assuming that you're doing research collaboration,
link |
you can submit, you know,
link |
there is a proper approval process guided by RB
link |
and if you go through all the processes,
link |
you can eventually get access to the data.
link |
But if you yourself know our OEI community,
link |
there are not that many people who actually ever got access
link |
to data because it's very challenging process.
link |
And sorry, just in a quick comment,
link |
MGH or any kind of hospital,
link |
are they scanning the data?
link |
Are they digitally storing it?
link |
Oh, it is already digitally stored.
link |
You don't need to do any extra processing steps.
link |
It's already there in the right format is that right now
link |
there are a lot of issues that govern access to the data
link |
because the hospital is legally responsible for the data.
link |
And, you know, they have a lot to lose
link |
if they give the data to the wrong person,
link |
but they may not have a lot to gain if they give it
link |
as a hospital, as a legal entity has given it to you.
link |
And the way, you know, what I would imagine
link |
happening in the future is the same thing that happens
link |
when you're getting your driving license,
link |
you can decide whether you want to donate your organs.
link |
You can imagine that whenever a person goes to the hospital,
link |
they, it should be easy for them to donate their data
link |
for research and it can be different kind of,
link |
do they only give you a test results or only mammogram
link |
or only imaging data or the whole medical record?
link |
Because at the end,
link |
we all will benefit from all this insights.
link |
And it's not like you say, I want to keep my data private,
link |
but I would really love to get it from other people
link |
because other people are thinking the same way.
link |
So if there is a mechanism to do this donation
link |
and the patient has an ability to say
link |
how they want to use their data for research,
link |
it would be really a game changer.
link |
People, when they think about this problem,
link |
there's a, it depends on the population,
link |
depends on the demographics,
link |
but there's some privacy concerns generally,
link |
not just medical data, just any kind of data.
link |
It's what you said, my data, it should belong kind of to me.
link |
I'm worried how it's going to be misused.
link |
How do we alleviate those concerns?
link |
Because that seems like a problem that needs to be,
link |
that problem of trust, of transparency needs to be solved
link |
before we build large data sets that help detect cancer,
link |
help save those very people in the future.
link |
So I think there are two things that could be done.
link |
There is a technical solutions
link |
and there are societal solutions.
link |
So on the technical end,
link |
we today have ability to improve disambiguation.
link |
Like, for instance, for imaging,
link |
it's, you know, for imaging, you can do it pretty well.
link |
What's disambiguation?
link |
And disambiguation, sorry, disambiguation,
link |
removing the identification,
link |
removing the names of the people.
link |
There are other data, like if it is a raw tax,
link |
you cannot really achieve 99.9%,
link |
but there are all these techniques
link |
that actually some of them are developed at MIT,
link |
how you can do learning on the encoded data
link |
where you locally encode the image,
link |
you train a network which only works on the encoded images
link |
and then you send the outcome back to the hospital
link |
and you can open it up.
link |
So those are the technical solutions.
link |
There are a lot of people who are working in this space
link |
where the learning happens in the encoded form.
link |
We are still early,
link |
but this is an interesting research area
link |
where I think we'll make more progress.
link |
There is a lot of work in natural language processing
link |
community how to do the identification better.
link |
But even today, there are already a lot of data
link |
which can be deidentified perfectly,
link |
like your test data, for instance, correct,
link |
where you can just, you know the name of the patient,
link |
you just want to extract the part with the numbers.
link |
The big problem here is again,
link |
hospitals don't see much incentive
link |
to give this data away on one hand
link |
and then there is general concern.
link |
Now, when I'm talking about societal benefits
link |
and about the education,
link |
the public needs to understand that I think
link |
that there are situation and I still remember myself
link |
when I really needed an answer, I had to make a choice.
link |
There was no information to make a choice,
link |
you're just guessing.
link |
And at that moment you feel that your life is at the stake,
link |
but you just don't have information to make the choice.
link |
And many times when I give talks,
link |
I get emails from women who say,
link |
you know, I'm in this situation,
link |
can you please run statistic and see what are the outcomes?
link |
We get almost every week a mammogram that comes by mail
link |
to my office at MIT, I'm serious.
link |
That people ask to run because they need to make
link |
life changing decisions.
link |
And of course, I'm not planning to open a clinic here,
link |
but we do run and give them the results for their doctors.
link |
But the point that I'm trying to make,
link |
that we all at some point or our loved ones
link |
will be in the situation where you need information
link |
to make the best choice.
link |
And if this information is not available,
link |
you would feel vulnerable and unprotected.
link |
And then the question is, you know, what do I care more?
link |
Because at the end, everything is a trade off, correct?
link |
Just out of curiosity, it seems like one possible solution,
link |
I'd like to see what you think of it,
link |
based on what you just said,
link |
based on wanting to know answers
link |
for when you're yourself in that situation.
link |
Is it possible for patients to own their data
link |
as opposed to hospitals owning their data?
link |
Of course, theoretically, I guess patients own their data,
link |
but can you walk out there with a USB stick
link |
containing everything or upload it to the cloud?
link |
Where a company, you know, I remember Microsoft
link |
had a service, like I try, I was really excited about
link |
and Google Health was there.
link |
I tried to give, I was excited about it.
link |
Basically companies helping you upload your data
link |
to the cloud so that you can move from hospital to hospital
link |
from doctor to doctor.
link |
Do you see a promise of that kind of possibility?
link |
I absolutely think this is, you know,
link |
the right way to exchange the data.
link |
I don't know now who's the biggest player in this field,
link |
but I can clearly see that even for totally selfish
link |
health reasons, when you are going to a new facility
link |
and many of us are sent to some specialized treatment,
link |
they don't easily have access to your data.
link |
And today, you know, we might want to send this mammogram,
link |
need to go to the hospital, find some small office
link |
which gives them the CD and they ship as a CD.
link |
So you can imagine we're looking at kind of decades old
link |
mechanism of data exchange.
link |
So I definitely think this is an area where hopefully
link |
all the right regulatory and technical forces will align
link |
and we will see it actually implemented.
link |
It's sad because unfortunately, and I need to research
link |
why that happened, but I'm pretty sure Google Health
link |
and Microsoft Health Vault or whatever it's called
link |
both closed down, which means that there was
link |
either regulatory pressure or there's not a business case
link |
or there's challenges from hospitals,
link |
which is very disappointing.
link |
So when you say you don't know what the biggest players are,
link |
the two biggest that I was aware of closed their doors.
link |
So I'm hoping, I'd love to see why
link |
and I'd love to see who else can come up.
link |
It seems like one of those Elon Musk style problems
link |
that are obvious needs to be solved
link |
and somebody needs to step up and actually do
link |
this large scale data collection.
link |
So I know there is an initiative in Massachusetts,
link |
I think, which you led by the governor
link |
to try to create this kind of health exchange system
link |
where at least to help people who kind of when you show up
link |
in emergency room and there is no information
link |
about what are your allergies and other things.
link |
So I don't know how far it will go.
link |
But another thing that you said
link |
and I find it very interesting is actually
link |
who are the successful players in this space
link |
and the whole implementation, how does it go?
link |
To me, it is from the anthropological perspective,
link |
it's more fascinating that AI that today goes in healthcare,
link |
we've seen so many attempts and so very little successes.
link |
And it's interesting to understand that I've by no means
link |
have knowledge to assess it,
link |
why we are in the position where we are.
link |
Yeah, it's interesting because data is really fuel
link |
for a lot of successful applications.
link |
And when that data acquires regulatory approval,
link |
like the FDA or any kind of approval,
link |
it seems that the computer scientists
link |
are not quite there yet in being able
link |
to play the regulatory game,
link |
understanding the fundamentals of it.
link |
I think that in many cases when even people do have data,
link |
we still don't know what exactly do you need to demonstrate
link |
to change the standard of care.
link |
Like let me give you an example
link |
related to my breast cancer research.
link |
So in traditional breast cancer risk assessment,
link |
there is something called density,
link |
which determines the likelihood of a woman to get cancer.
link |
And this pretty much says,
link |
how much white do you see on the mammogram?
link |
The whiter it is, the more likely the tissue is dense.
link |
And the idea behind density, it's not a bad idea.
link |
In 1967, a radiologist called Wolf decided to look back
link |
at women who were diagnosed
link |
and see what is special in their images.
link |
Can we look back and say that they're likely to develop?
link |
So he come up with some patterns.
link |
And it was the best that his human eye can identify.
link |
Then it was kind of formalized
link |
and coded into four categories.
link |
And that's what we are using today.
link |
And today this density assessment
link |
is actually a federal law from 2019,
link |
approved by President Trump
link |
and for the previous FDA commissioner,
link |
where women are supposed to be advised by their providers
link |
if they have high density,
link |
putting them into higher risk category.
link |
And in some states,
link |
you can actually get supplementary screening
link |
paid by your insurance because you're in this category.
link |
Now you can say, how much science do we have behind it?
link |
Whatever, biological science or epidemiological evidence.
link |
So it turns out that between 40 and 50% of women
link |
have dense breasts.
link |
So about 40% of patients are coming out of their screening
link |
and somebody tells them, you are in high risk.
link |
Now, what exactly does it mean
link |
if you as half of the population in high risk?
link |
It's from saying, maybe I'm not,
link |
or what do I really need to do with it?
link |
Because the system doesn't provide me
link |
a lot of the solutions
link |
because there are so many people like me,
link |
we cannot really provide very expensive solutions for them.
link |
And the reason this whole density became this big deal,
link |
it's actually advocated by the patients
link |
who felt very unprotected
link |
because many women went and did the mammograms
link |
which were normal.
link |
And then it turns out that they already had cancer,
link |
quite developed cancer.
link |
So they didn't have a way to know who is really at risk
link |
and what is the likelihood that when the doctor tells you,
link |
you're okay, you are not okay.
link |
So at the time, and it was 15 years ago,
link |
this maybe was the best piece of science that we had.
link |
And it took quite 15, 16 years to make it federal law.
link |
But now this is a standard.
link |
Now with a deep learning model,
link |
we can so much more accurately predict
link |
who is gonna develop breast cancer
link |
just because you're trained on a logical thing.
link |
And instead of describing how much white
link |
and what kind of white machine
link |
can systematically identify the patterns,
link |
which was the original idea behind the thought
link |
of the cardiologist,
link |
machines can do it much more systematically
link |
and predict the risk when you're training the machine
link |
to look at the image and to say the risk in one to five years.
link |
Now you can ask me how long it will take
link |
to substitute this density,
link |
which is broadly used across the country
link |
and really is not helping to bring this new models.
link |
And I would say it's not a matter of the algorithm.
link |
Algorithms use already orders of magnitude better
link |
than what is currently in practice.
link |
I think it's really the question,
link |
who do you need to convince?
link |
How many hospitals do you need to run the experiment?
link |
What, you know, all this mechanism of adoption
link |
and how do you explain to patients
link |
and to women across the country
link |
that this is really a better measure?
link |
And again, I don't think it's an AI question.
link |
We can work more and make the algorithm even better,
link |
but I don't think that this is the current, you know,
link |
the barrier, the barrier is really this other piece
link |
that for some reason is not really explored.
link |
It's like anthropological piece.
link |
And coming back to your question about books,
link |
there is a book that I'm reading.
link |
It's called American Sickness by Elizabeth Rosenthal.
link |
And I got this book from my clinical collaborator,
link |
Dr. Connie Lehman.
link |
And I said, I know everything that I need to know
link |
about American health system,
link |
but you know, every page doesn't fail to surprise me.
link |
And I think there is a lot of interesting
link |
and really deep lessons for people like us
link |
from computer science who are coming into this field
link |
to really understand how complex is the system of incentives
link |
in the system to understand how you really need to play
link |
to drive adoption.
link |
You just said it's complex,
link |
but if we're trying to simplify it,
link |
who do you think most likely would be successful
link |
if we push on this group of people?
link |
Is it the doctors?
link |
Is it the hospitals?
link |
Is it the governments or policymakers?
link |
Is it the individual patients, consumers?
link |
Who needs to be inspired to most likely lead to adoption?
link |
Or is there no simple answer?
link |
There's no simple answer,
link |
but I think there is a lot of good people in medical system
link |
who do want to make a change.
link |
And I think a lot of power will come from us as consumers
link |
because we all are consumers or future consumers
link |
of healthcare services.
link |
And I think we can do so much more
link |
in explaining the potential and not in the hype terms
link |
and not saying that we now killed all Alzheimer
link |
and I'm really sick of reading this kind of articles
link |
which make these claims,
link |
but really to show with some examples
link |
what this implementation does and how it changes the care.
link |
Because I can't imagine,
link |
it doesn't matter what kind of politician it is,
link |
we all are susceptible to these diseases.
link |
There is no one who is free.
link |
And eventually, we all are humans
link |
and we're looking for a way to alleviate the suffering.
link |
And this is one possible way
link |
where we currently are under utilizing,
link |
which I think can help.
link |
So it sounds like the biggest problems are outside of AI
link |
in terms of the biggest impact at this point.
link |
But are there any open problems
link |
in the application of ML to oncology in general?
link |
So improving the detection or any other creative methods,
link |
whether it's on the detection segmentations
link |
or the vision perception side
link |
or some other clever of inference?
link |
Yeah, what in general in your view are the open problems
link |
Yeah, I just want to mention that beside detection,
link |
not the area where I am kind of quite active
link |
and I think it's really an increasingly important area
link |
in healthcare is drug design.
link |
Because it's fine if you detect something early,
link |
but you still need to get drugs
link |
and new drugs for these conditions.
link |
And today, all of the drug design,
link |
ML is non existent there.
link |
We don't have any drug that was developed by the ML model
link |
or even not developed,
link |
but at least even knew that ML model
link |
plays some significant role.
link |
I think this area with all the new ability
link |
to generate molecules with desired properties
link |
to do in silica screening is really a big open area.
link |
To be totally honest with you,
link |
when we are doing diagnostics and imaging,
link |
primarily taking the ideas that were developed
link |
for other areas and you applying them with some adaptation,
link |
the area of drug design is really technically interesting
link |
and exciting area.
link |
You need to work a lot with graphs
link |
and capture various 3D properties.
link |
There are lots and lots of opportunities
link |
to be technically creative.
link |
And I think there are a lot of open questions in this area.
link |
We're already getting a lot of successes
link |
even with kind of the first generation of these models,
link |
but there is much more new creative things that you can do.
link |
And what's very nice to see is that actually
link |
the more powerful, the more interesting models
link |
actually do do better.
link |
So there is a place to innovate in machine learning
link |
And some of these techniques are really unique to,
link |
let's say, to graph generation and other things.
link |
What, just to interrupt really quick, I'm sorry,
link |
graph generation or graphs, drug discovery in general,
link |
how do you discover a drug?
link |
Is this chemistry?
link |
Is this trying to predict different chemical reactions?
link |
Or is it some kind of...
link |
What do graphs even represent in this space?
link |
And what's a drug?
link |
Okay, so let's say you're thinking
link |
there are many different types of drugs,
link |
but let's say you're gonna talk about small molecules
link |
because I think today the majority of drugs
link |
are small molecules.
link |
So small molecule is a graph.
link |
The molecule is just where the node in the graph
link |
is an atom and then you have the bonds.
link |
So it's really a graph representation.
link |
If you look at it in 2D, correct,
link |
you can do it 3D, but let's say,
link |
let's keep it simple and stick in 2D.
link |
So pretty much my understanding today,
link |
how it is done at scale in the companies,
link |
without machine learning,
link |
you have high throughput screening.
link |
So you know that you are interested
link |
to get certain biological activity of the compound.
link |
So you scan a lot of compounds,
link |
like maybe hundreds of thousands,
link |
some really big number of compounds.
link |
You identify some compounds which have the right activity
link |
and then at this point, the chemists come
link |
and they're trying to now to optimize
link |
this original heat to different properties
link |
that you want it to be maybe soluble,
link |
you want it to decrease toxicity,
link |
you want it to decrease the side effects.
link |
Are those, sorry again to interrupt,
link |
can that be done in simulation
link |
or just by looking at the molecules
link |
or do you need to actually run reactions
link |
in real labs with lab coats and stuff?
link |
So when you do high throughput screening,
link |
you really do screening.
link |
It's really the lab screening.
link |
You screen the molecules, correct?
link |
I don't know what screening is.
link |
The screening is just check them for certain property.
link |
Like in the physical space, in the physical world,
link |
like actually there's a machine probably
link |
that's actually running the reaction.
link |
Actually running the reactions, yeah.
link |
So there is a process where you can run
link |
and that's why it's called high throughput
link |
that it become cheaper and faster
link |
to do it on very big number of molecules.
link |
You run the screening,
link |
you identify potential good starts
link |
and then when the chemists come in
link |
who have done it many times
link |
and then they can try to look at it and say,
link |
how can you change the molecule
link |
to get the desired profile
link |
in terms of all other properties?
link |
So maybe how do I make it more bioactive and so on?
link |
And there the creativity of the chemists
link |
really is the one that determines the success
link |
of this design because again,
link |
they have a lot of domain knowledge
link |
of what works, how do you decrease the CCD and so on
link |
and that's what they do.
link |
So all the drugs that are currently
link |
in the FDA approved drugs
link |
or even drugs that are in clinical trials,
link |
they are designed using these domain experts
link |
which goes through this combinatorial space
link |
of molecules or graphs or whatever
link |
and find the right one or adjust it to be the right ones.
link |
It sounds like the breast density heuristic
link |
from 67 to the same echoes.
link |
It's not necessarily that.
link |
It's really driven by deep understanding.
link |
It's not like they just observe it.
link |
I mean, they do deeply understand chemistry
link |
and they do understand how different groups
link |
and how does it changes the properties.
link |
So there is a lot of science that gets into it
link |
and a lot of kind of simulation,
link |
how do you want it to behave?
link |
It's very, very complex.
link |
So they're quite effective at this design, obviously.
link |
Now effective, yeah, we have drugs.
link |
Like depending on how do you measure effective,
link |
if you measure it in terms of cost, it's prohibitive.
link |
If you measure it in terms of times,
link |
we have lots of diseases for which we don't have any drugs
link |
and we don't even know how to approach
link |
and don't need to mention few drugs
link |
or neurodegenerative disease drugs that fail.
link |
So there are lots of trials that fail in later stages,
link |
which is really catastrophic from the financial perspective.
link |
So is it the effective, the most effective mechanism?
link |
Absolutely no, but this is the only one that currently works.
link |
And I was closely interacting
link |
with people in pharmaceutical industry.
link |
I was really fascinated on how sharp
link |
and what a deep understanding of the domain do they have.
link |
It's not observation driven.
link |
There is really a lot of science behind what they do.
link |
But if you ask me, can machine learning change it,
link |
I firmly believe yes,
link |
because even the most experienced chemists
link |
cannot hold in their memory and understanding
link |
everything that you can learn
link |
from millions of molecules and reactions.
link |
And the space of graphs is a totally new space.
link |
I mean, it's a really interesting space
link |
for machine learning to explore, graph generation.
link |
Yeah, so there are a lot of things that you can do here.
link |
So we do a lot of work.
link |
So the first tool that we started with
link |
was the tool that can predict properties of the molecules.
link |
So you can just give the molecule and the property.
link |
It can be by activity property,
link |
or it can be some other property.
link |
And you train the molecules
link |
and you can now take a new molecule
link |
and predict this property.
link |
Now, when people started working in this area,
link |
it is something very simple.
link |
They do kind of existing fingerprints,
link |
which is kind of handcrafted features of the molecule.
link |
When you break the graph to substructures
link |
and then you run it in a feed forward neural network.
link |
And what was interesting to see that clearly,
link |
this was not the most effective way to proceed.
link |
And you need to have much more complex models
link |
that can induce a representation,
link |
which can translate this graph into the embeddings
link |
and do these predictions.
link |
So this is one direction.
link |
Then another direction, which is kind of related
link |
is not only to stop by looking at the embedding itself,
link |
but actually modify it to produce better molecules.
link |
So you can think about it as machine translation
link |
that you can start with a molecule
link |
and then there is an improved version of molecule.
link |
And you can again, with encoder translate it
link |
into the hidden space and then learn how to modify it
link |
to improve the in some ways version of the molecules.
link |
So that's, it's kind of really exciting.
link |
We already have seen that the property prediction
link |
works pretty well.
link |
And now we are generating molecules
link |
and there is actually labs
link |
which are manufacturing this molecule.
link |
So we'll see where it will get us.
link |
Okay, that's really exciting.
link |
There's a lot of promise.
link |
Speaking of machine translation and embeddings,
link |
I think you have done a lot of really great research
link |
in NLP, natural language processing.
link |
Can you tell me your journey through NLP?
link |
What ideas, problems, approaches were you working on?
link |
Were you fascinated with, did you explore
link |
before this magic of deep learning reemerged and after?
link |
So when I started my work in NLP, it was in 97.
link |
This was very interesting time.
link |
It was exactly the time that I came to ACL.
link |
And at the time I could barely understand English,
link |
but it was exactly like the transition point
link |
because half of the papers were really rule based approaches
link |
where people took more kind of heavy linguistic approaches
link |
for small domains and try to build up from there.
link |
And then there were the first generation of papers
link |
which were corpus based papers.
link |
And they were very simple in our terms
link |
when you collect some statistics
link |
and do prediction based on them.
link |
And I found it really fascinating that one community
link |
can think so very differently about the problem.
link |
And I remember my first paper that I wrote,
link |
it didn't have a single formula.
link |
It didn't have evaluation.
link |
It just had examples of outputs.
link |
And this was a standard of the field at the time.
link |
In some ways, I mean, people maybe just started emphasizing
link |
the empirical evaluation, but for many applications
link |
like summarization, you just show some examples of outputs.
link |
And then increasingly you can see that how
link |
the statistical approaches dominated the field
link |
and we've seen increased performance
link |
across many basic tasks.
link |
The sad part of the story maybe that if you look again
link |
through this journey, we see that the role of linguistics
link |
in some ways greatly diminishes.
link |
And I think that you really need to look
link |
through the whole proceeding to find one or two papers
link |
which make some interesting linguistic references.
link |
This was definitely one of the.
link |
Things like syntactic trees, just even basically
link |
against our conversation about human understanding
link |
of language, which I guess what linguistics would be
link |
structured, hierarchical representing language
link |
in a way that's human explainable, understandable
link |
I don't know if it is, what is explainable
link |
and understandable.
link |
In the end, we perform functions and it's okay
link |
to have machine which performs a function.
link |
Like when you're thinking about your calculator, correct?
link |
Your calculator can do calculation very different
link |
from you would do the calculation,
link |
but it's very effective in it.
link |
And this is fine if we can achieve certain tasks
link |
with high accuracy, doesn't necessarily mean
link |
that it has to understand it the same way as we understand.
link |
In some ways, it's even naive to request
link |
because you have so many other sources of information
link |
that are absent when you are training your system.
link |
And I would tell you one application
link |
that is really fascinating.
link |
In 97, when it came to ACL, there were some papers
link |
on machine translation.
link |
They were like primitive.
link |
Like people were trying really, really simple.
link |
And the feeling, my feeling was that, you know,
link |
to make real machine translation system,
link |
it's like to fly at the moon and build a house there
link |
and the garden and live happily ever after.
link |
I mean, it's like impossible.
link |
I never could imagine that within, you know, 10 years,
link |
we would already see the system working.
link |
And now, you know, nobody is even surprised
link |
to utilize the system on daily basis.
link |
So this was like a huge, huge progress,
link |
saying that people for very long time
link |
tried to solve using other mechanisms.
link |
And they were unable to solve it.
link |
That's why coming back to your question about biology,
link |
that, you know, in linguistics, people try to go this way
link |
and try to write the syntactic trees
link |
and try to abstract it and to find the right representation.
link |
And, you know, they couldn't get very far
link |
with this understanding while these models using,
link |
you know, other sources actually capable
link |
to make a lot of progress.
link |
Now, I'm not naive to think
link |
that we are in this paradise space in NLP.
link |
And sure as you know,
link |
that when we slightly change the domain
link |
and when we decrease the amount of training,
link |
it can do like really bizarre and funny thing.
link |
But I think it's just a matter
link |
of improving generalization capacity,
link |
which is just a technical question.
link |
Wow, so that's the question.
link |
How much of language understanding can be solved
link |
with deep neural networks?
link |
In your intuition, I mean, it's unknown, I suppose.
link |
But as we start to creep towards romantic notions
link |
of the spirit of the Turing test
link |
and conversation and dialogue
link |
and something that maybe to me or to us,
link |
so the humans feels like it needs real understanding.
link |
How much can that be achieved
link |
with these neural networks or statistical methods?
link |
So I guess I am very much driven by the outcomes.
link |
Can we achieve the performance
link |
which would be satisfactory for us for different tasks?
link |
Now, if you again look at machine translation system,
link |
which are trained on large amounts of data,
link |
they really can do a remarkable job
link |
relatively to where they've been a few years ago.
link |
And if you project into the future,
link |
if it will be the same speed of improvement, you know,
link |
Now, does it bother me
link |
that it's not doing the same translation as we are doing?
link |
Now, if you go to cognitive science,
link |
we still don't really understand what we are doing.
link |
I mean, there are a lot of theories
link |
and there's obviously a lot of progress and studying,
link |
but our understanding what exactly goes on in our brains
link |
when we process language is still not crystal clear
link |
and precise that we can translate it into machines.
link |
What does bother me is that, you know,
link |
again, that machines can be extremely brittle
link |
when you go out of your comfort zone
link |
of when there is a distributional shift
link |
between training and testing.
link |
And it have been years and years,
link |
every year when I teach an LP class,
link |
now show them some examples of translation
link |
from some newspaper in Hebrew or whatever, it was perfect.
link |
And then I have a recipe that Tomi Yakel's system
link |
sent me a while ago and it was written in Finnish
link |
And it's just a terrible translation.
link |
You cannot understand anything what it does.
link |
It's not like some syntactic mistakes, it's just terrible.
link |
And year after year, I tried and will translate
link |
and year after year, it does this terrible work
link |
because I guess, you know, the recipes
link |
are not a big part of their training repertoire.
link |
So, but in terms of outcomes, that's a really clean,
link |
good way to look at it.
link |
I guess the question I was asking is,
link |
do you think, imagine a future,
link |
do you think the current approaches can pass
link |
the Turing test in the way,
link |
in the best possible formulation of the Turing test?
link |
Which is, would you wanna have a conversation
link |
with a neural network for an hour?
link |
Oh God, no, no, there are not that many people
link |
that I would want to talk for an hour, but.
link |
There are some people in this world, alive or not,
link |
that you would like to talk to for an hour.
link |
Could a neural network achieve that outcome?
link |
So I think it would be really hard to create
link |
a successful training set, which would enable it
link |
to have a conversation, a contextual conversation
link |
Do you think it's a problem of data, perhaps?
link |
I think in some ways it's not a problem of data,
link |
it's a problem both of data and the problem of
link |
the way we're training our systems,
link |
their ability to truly, to generalize,
link |
to be very compositional.
link |
In some ways it's limited in the current capacity,
link |
at least we can translate well,
link |
we can find information well, we can extract information.
link |
So there are many capacities in which it's doing very well.
link |
And you can ask me, would you trust the machine
link |
to translate for you and use it as a source?
link |
I would say absolutely, especially if we're talking about
link |
newspaper data or other data which is in the realm
link |
of its own training set, I would say yes.
link |
But having conversations with the machine,
link |
it's not something that I would choose to do.
link |
But I would tell you something, talking about Turing tests
link |
and about all this kind of ELISA conversations,
link |
I remember visiting Tencent in China
link |
and they have this chat board and they claim
link |
there is really humongous amount of the local population
link |
which for hours talks to the chat board.
link |
To me it was, I cannot believe it,
link |
but apparently it's documented that there are some people
link |
who enjoy this conversation.
link |
And it brought to me another MIT story
link |
about ELISA and Weisenbaum.
link |
I don't know if you're familiar with the story.
link |
So Weisenbaum was a professor at MIT
link |
and when he developed this ELISA,
link |
which was just doing string matching,
link |
very trivial, like restating of what you said
link |
with very few rules, no syntax.
link |
Apparently there were secretaries at MIT
link |
that would sit for hours and converse with this trivial thing
link |
and at the time there was no beautiful interfaces
link |
so you actually need to go through the pain
link |
And Weisenbaum himself was so horrified by this phenomenon
link |
that people can believe enough to the machine
link |
that you just need to give them the hint
link |
that machine understands you and you can complete the rest
link |
that he kind of stopped this research
link |
and went into kind of trying to understand
link |
what this artificial intelligence can do to our brains.
link |
So my point is, you know,
link |
how much, it's not how good is the technology,
link |
it's how ready we are to believe
link |
that it delivers the goods that we are trying to get.
link |
That's a really beautiful way to put it.
link |
I, by the way, I'm not horrified by that possibility,
link |
but inspired by it because,
link |
I mean, human connection,
link |
whether it's through language or through love,
link |
it seems like it's very amenable to machine learning
link |
and the rest is just challenges of psychology.
link |
Like you said, the secretaries who enjoy spending hours.
link |
I would say I would describe most of our lives
link |
as enjoying spending hours with those we love
link |
for very silly reasons.
link |
All we're doing is keyword matching as well.
link |
So I'm not sure how much intelligence
link |
we exhibit to each other with the people we love
link |
that we're close with.
link |
So it's a very interesting point
link |
of what it means to pass the Turing test with language.
link |
I think you're right.
link |
In terms of conversation,
link |
I think machine translation
link |
has very clear performance and improvement, right?
link |
What it means to have a fulfilling conversation
link |
is very person dependent and context dependent
link |
That's, yeah, it's very well put.
link |
But in your view, what's a benchmark in natural language,
link |
a test that's just out of reach right now,
link |
but we might be able to, that's exciting.
link |
Is it in perfecting machine translation
link |
or is there other, is it summarization?
link |
What's out there just out of reach?
link |
I think it goes across specific application.
link |
It's more about the ability to learn from few examples
link |
for real, what we call few short learning and all these cases
link |
because the way we publish these papers today,
link |
we say, if we have like naively, we get 55,
link |
but now we had a few example and we can move to 65.
link |
None of these methods
link |
actually are realistically doing anything useful.
link |
You cannot use them today.
link |
And the ability to be able to generalize and to move
link |
or to be autonomous in finding the data
link |
that you need to learn,
link |
to be able to perfect new tasks or new language,
link |
this is an area where I think we really need
link |
to move forward to and we are not yet there.
link |
Are you at all excited,
link |
curious by the possibility
link |
of creating human level intelligence?
link |
Is this, cause you've been very in your discussion.
link |
So if we look at oncology,
link |
you're trying to use machine learning to help the world
link |
in terms of alleviating suffering.
link |
If you look at natural language processing,
link |
you're focused on the outcomes of improving practical things
link |
like machine translation.
link |
But human level intelligence is a thing
link |
that our civilization has dreamed about creating,
link |
super human level intelligence.
link |
Do you think about this?
link |
Do you think it's at all within our reach?
link |
So as you said yourself, Elie,
link |
talking about how do you perceive
link |
our communications with each other,
link |
that we're matching keywords and certain behaviors
link |
So at the end, whenever one assesses,
link |
let's say relations with another person,
link |
you have separate kind of measurements and outcomes
link |
inside your head that determine
link |
what is the status of the relation.
link |
So one way, this is this classical level,
link |
what is the intelligence?
link |
Is it the fact that now we are gonna do the same way
link |
as human is doing,
link |
when we don't even understand what the human is doing?
link |
Or we now have an ability to deliver these outcomes,
link |
but not in one area, not in NLP,
link |
not just to translate or just to answer questions,
link |
but across many, many areas
link |
that we can achieve the functionalities
link |
that humans can achieve with their ability to learn
link |
and do other things.
link |
I think this is, and this we can actually measure
link |
And that's what makes me excited that we,
link |
in my lifetime, at least so far what we've seen,
link |
it's like tremendous progress
link |
across these different functionalities.
link |
And I think it will be really exciting
link |
to see where we will be.
link |
And again, one way to think about it,
link |
there are machines which are improving their functionality.
link |
Another one is to think about us with our brains,
link |
which are imperfect,
link |
how they can be accelerated by this technology
link |
as it becomes stronger and stronger.
link |
Coming back to another book
link |
that I love, Flowers for Algernon.
link |
Have you read this book?
link |
So there is this point that the patient gets
link |
this miracle cure, which changes his brain.
link |
And all of a sudden they see life in a different way
link |
and can do certain things better,
link |
but certain things much worse.
link |
So you can imagine this kind of computer augmented cognition
link |
where it can bring you that now in the same way
link |
as the cars enable us to get to places
link |
where we've never been before,
link |
can we think differently?
link |
Can we think faster?
link |
And we already see a lot of it happening
link |
in how it impacts us,
link |
but I think we have a long way to go there.
link |
So that's sort of artificial intelligence
link |
and technology affecting our,
link |
augmenting our intelligence as humans.
link |
Yesterday, a company called Neuralink announced,
link |
they did this whole demonstration.
link |
I don't know if you saw it.
link |
It's, they demonstrated brain computer,
link |
brain machine interface,
link |
where there's like a sewing machine for the brain.
link |
Do you, you know, a lot of that is quite out there
link |
in terms of things that some people would say
link |
are impossible, but they're dreamers
link |
and want to engineer systems like that.
link |
Do you see, based on what you just said,
link |
a hope for that more direct interaction with the brain?
link |
I think there are different ways.
link |
One is a direct interaction with the brain.
link |
And again, there are lots of companies
link |
that work in this space
link |
and I think there will be a lot of developments.
link |
But I'm just thinking that many times
link |
we are not aware of our feelings,
link |
of motivation, what drives us.
link |
Like, let me give you a trivial example, our attention.
link |
There are a lot of studies that demonstrate
link |
that it takes a while to a person to understand
link |
that they are not attentive anymore.
link |
And we know that there are people
link |
who really have strong capacity to hold attention.
link |
There are other end of the spectrum people with ADD
link |
and other issues that they have problem
link |
to regulate their attention.
link |
Imagine to yourself that you have like a cognitive aid
link |
that just alerts you based on your gaze,
link |
that your attention is now not on what you are doing.
link |
And instead of writing a paper,
link |
you're now dreaming of what you're gonna do in the evening.
link |
So even this kind of simple measurement things,
link |
how they can change us.
link |
And I see it even in simple ways with myself.
link |
I have my zone app that I got in MIT gym.
link |
It kind of records, you know, how much did you run
link |
and you have some points
link |
and you can get some status, whatever.
link |
Like, I said, what is this ridiculous thing?
link |
Who would ever care about some status in some app?
link |
So to maintain the status,
link |
you have to do set a number of points every month.
link |
And not only is that I do it every single month
link |
for the last 18 months,
link |
it went to the point that I was injured.
link |
And when I could run again,
link |
in two days, I did like some humongous amount of running
link |
just to complete the points.
link |
It was like really not safe.
link |
It was like, I'm not gonna lose my status
link |
because I want to get there.
link |
So you can already see that this direct measurement
link |
and the feedback is, you know,
link |
we're looking at video games
link |
and see why, you know, the addiction aspect of it,
link |
but you can imagine that the same idea can be expanded
link |
to many other areas of our life.
link |
When we really can get feedback
link |
and imagine in your case in relations,
link |
when we are doing keyword matching,
link |
imagine that the person who is generating the keywords,
link |
that person gets direct feedback
link |
before the whole thing explodes.
link |
Is it maybe at this happy point,
link |
we are going in the wrong direction.
link |
Maybe it will be really a behavior modifying moment.
link |
So yeah, it's a relationship management too.
link |
So yeah, that's a fascinating whole area
link |
of psychology actually as well,
link |
of seeing how our behavior has changed
link |
with basically all human relations now have
link |
other nonhuman entities helping us out.
link |
So you teach a large,
link |
a huge machine learning course here at MIT.
link |
I can ask you a million questions,
link |
but you've seen a lot of students.
link |
What ideas do students struggle with the most
link |
as they first enter this world of machine learning?
link |
Actually, this year was the first time
link |
I started teaching a small machine learning class.
link |
And it came as a result of what I saw
link |
in my big machine learning class that Tomi Yakel and I built
link |
maybe six years ago.
link |
What we've seen that as this area become more
link |
and more popular, more and more people at MIT
link |
want to take this class.
link |
And while we designed it for computer science majors,
link |
there were a lot of people who really are interested
link |
to learn it, but unfortunately,
link |
their background was not enabling them
link |
to do well in the class.
link |
And many of them associated machine learning
link |
with the word struggle and failure,
link |
primarily for non majors.
link |
And that's why we actually started a new class
link |
which we call machine learning from algorithms to modeling,
link |
which emphasizes more the modeling aspects of it
link |
and focuses on, it has majors and non majors.
link |
So we kind of try to extract the relevant parts
link |
and make it more accessible,
link |
because the fact that we're teaching 20 classifiers
link |
in standard machine learning class,
link |
it's really a big question to really need it.
link |
But it was interesting to see this
link |
from first generation of students,
link |
when they came back from their internships
link |
and from their jobs,
link |
what different and exciting things they can do.
link |
I would never think that you can even apply
link |
machine learning to, some of them are like matching,
link |
the relations and other things like variety.
link |
Everything is amenable as the machine learning.
link |
That actually brings up an interesting point
link |
of computer science in general.
link |
It almost seems, maybe I'm crazy,
link |
but it almost seems like everybody needs to learn
link |
how to program these days.
link |
If you're 20 years old, or if you're starting school,
link |
even if you're an English major,
link |
it seems like programming unlocks so much possibility
link |
So when you interacted with those non majors,
link |
is there skills that they were simply lacking at the time
link |
that you wish they had and that they learned
link |
in high school and so on?
link |
Like how should education change
link |
in this computerized world that we live in?
link |
I think because I knew that there is a Python component
link |
in the class, their Python skills were okay
link |
and the class isn't really heavy on programming.
link |
They primarily kind of add parts to the programs.
link |
I think it was more of the mathematical barriers
link |
and the class, again, with the design on the majors
link |
was using the notation, like big O for complexity
link |
and others, people who come from different backgrounds
link |
just don't have it in the lexical,
link |
so necessarily very challenging notion,
link |
but they were just not aware.
link |
So I think that kind of linear algebra and probability,
link |
the basics, the calculus, multivariate calculus,
link |
things that can help.
link |
What advice would you give to students
link |
interested in machine learning,
link |
interested, you've talked about detecting,
link |
curing cancer, drug design,
link |
if they want to get into that field, what should they do?
link |
Get into it and succeed as researchers
link |
and entrepreneurs.
link |
The first good piece of news is that right now
link |
there are lots of resources
link |
that are created at different levels
link |
and you can find online in your school classes
link |
which are more mathematical, more applied and so on.
link |
So you can find a kind of a preacher
link |
which preaches in your own language
link |
where you can enter the field
link |
and you can make many different types of contribution
link |
depending of what is your strengths.
link |
And the second point, I think it's really important
link |
to find some area which you really care about
link |
and it can motivate your learning
link |
and it can be for somebody curing cancer
link |
or doing self driving cars or whatever,
link |
but to find an area where there is data
link |
where you believe there are strong patterns
link |
and we should be doing it and we're still not doing it
link |
or you can do it better
link |
and just start there and see where it can bring you.
link |
So you've been very successful in many directions in life,
link |
but you also mentioned Flowers of Argonon.
link |
And I think I've read or listened to you mention somewhere
link |
that researchers often get lost
link |
in the details of their work.
link |
This is per our original discussion with cancer and so on
link |
and don't look at the bigger picture,
link |
bigger questions of meaning and so on.
link |
So let me ask you the impossible question
link |
of what's the meaning of this thing,
link |
of life, of your life, of research.
link |
Why do you think we descendant of great apes
link |
are here on this spinning ball?
link |
You know, I don't think that I have really a global answer.
link |
You know, maybe that's why I didn't go to humanities
link |
and I didn't take humanities classes in my undergrad.
link |
But the way I'm thinking about it,
link |
each one of us inside of them have their own set of,
link |
you know, things that we believe are important.
link |
And it just happens that we are busy
link |
with achieving various goals, busy listening to others
link |
and to kind of try to conform and to be part of the crowd,
link |
that we don't listen to that part.
link |
And, you know, we all should find some time to understand
link |
what is our own individual missions.
link |
And we may have very different missions
link |
and to make sure that while we are running 10,000 things,
link |
we are not, you know, missing out
link |
and we're putting all the resources to satisfy
link |
And if I look over my time, when I was younger,
link |
most of these missions, you know,
link |
I was primarily driven by the external stimulus,
link |
you know, to achieve this or to be that.
link |
And now a lot of what I do is driven by really thinking
link |
what is important for me to achieve independently
link |
of the external recognition.
link |
And, you know, I don't mind to be viewed in certain ways.
link |
The most important thing for me is to be true to myself,
link |
to what I think is right.
link |
How long did it take?
link |
How hard was it to find the you that you have to be true to?
link |
And even now, sometimes, you know,
link |
the vanity and the triviality can take, you know.
link |
Yeah, it can everywhere, you know,
link |
it's just the vanity at MIT is different,
link |
the vanity in different places,
link |
but we all have our piece of vanity.
link |
But I think actually for me, many times the place
link |
to get back to it is, you know, when I'm alone
link |
and also when I read.
link |
And I think by selecting the right books,
link |
you can get the right questions and learn from what you read.
link |
So, but again, it's not perfect.
link |
Like vanity sometimes dominates.
link |
Well, that's a beautiful way to end.
link |
Thank you so much for talking today.