back to indexDaphne Koller: Biomedicine and Machine Learning | Lex Fridman Podcast #93
link |
The following is a conversation with Daphne Koller,
link |
a professor of computer science at Stanford University,
link |
a cofounder of Coursera with Andrew Eng,
link |
and founder and CEO of Incitro,
link |
a company at the intersection of machine learning
link |
We're now in the exciting early days
link |
of using the data driven methods of machine learning
link |
to help discover and develop new drugs
link |
and treatment that scale.
link |
Daphne and Incitro are leading the way on this
link |
with breakthroughs that may ripple through all fields
link |
of medicine, including one's most critical for helping
link |
with the current coronavirus pandemic.
link |
This conversation was recorded before the COVID 19 outbreak.
link |
For everyone feeling the medical, psychological,
link |
and financial burden of this crisis,
link |
I'm sending love your way.
link |
Stay strong, we're in this together, we'll beat this thing.
link |
This is the Artificial Intelligence Podcast.
link |
If you enjoy it, subscribe on YouTube,
link |
review it with five stars on Apple Podcasts,
link |
support it on Patreon, or simply connect with me on Twitter
link |
at Lex Friedman, spelled F R I D M A N.
link |
As usual, I'll do a few minutes of ads now
link |
and never any ads in the middle
link |
that can break the flow of this conversation.
link |
I hope that works for you
link |
and doesn't hurt the listening experience.
link |
This show is presented by Cash App,
link |
the number one finance app in the App Store.
link |
When you get it, use code Lex Podcast.
link |
Cash App lets you send money to friends,
link |
buy Bitcoin, and invest in the stock market
link |
with as little as $1.
link |
Since Cash App allows you to send
link |
and receive money digitally, peer to peer,
link |
and security in all digital transactions is very important.
link |
Let me mention that PCI data security standard
link |
that Cash App is compliant with.
link |
I'm a big fan of standards for safety and security.
link |
PCI DSS is a good example of that,
link |
where a bunch of competitors got together and agreed
link |
that there needs to be a global standard
link |
around the security of transactions.
link |
Now we just need to do the same
link |
for autonomous vehicles and AI systems in general.
link |
So again, if you get Cash App from the App Store,
link |
Google Play, and use the code Lex Podcast,
link |
you get $10, and Cash App will also donate $10 to first,
link |
an organization that is helping to advance robotics
link |
and STEM education for young people around the world.
link |
And now here's my conversation with Daphne Koller.
link |
So you cofonded Coursera.
link |
I made a huge impact in the global education of AI,
link |
and after five years, in August 2016, wrote a blog post
link |
saying that you're stepping away and wrote, quote,
link |
it is time for me to turn to another critical challenge,
link |
the development of machine learning
link |
and its applications to improving human health.
link |
So let me ask two far out philosophical questions.
link |
One, do you think we'll one day find cures
link |
for all major diseases known today?
link |
And two, do you think we'll one day figure out
link |
a way to extend the human lifespan,
link |
perhaps to the point of immortality?
link |
So one day is a very long time,
link |
and I don't like to make predictions of the type
link |
we will never be able to do X,
link |
because I think that's a, you know, that's,
link |
that's a smacks of hubris.
link |
It seems that never in the, in the entire eternity
link |
of human existence will we be able to solve a problem.
link |
That being said, curing disease is very hard
link |
because oftentimes by the time you discover the disease,
link |
a lot of damage has already been done.
link |
And so to assume that we would be able to cure disease
link |
at that stage assumes that we would come up with ways
link |
of basically regenerating entire parts of the human body
link |
in the way that actually returns it to its original state.
link |
And that's a very challenging problem.
link |
We have cured very few diseases.
link |
We've been able to provide treatment
link |
for an increasingly large number,
link |
but the number of things that you could actually define
link |
to be cures is actually not that large.
link |
So I think that's, there's a lot of work
link |
that would need to happen before one could legitimately say
link |
that we have cured even a reasonable number
link |
far less all diseases.
link |
On the scale of zero to 100,
link |
where are we in understanding the fundamental mechanisms
link |
of all major diseases?
link |
What's your sense?
link |
So from the computer science perspective
link |
that you've entered the world of health,
link |
how far along are we?
link |
I think it depends on which disease.
link |
I mean, there are ones where I would say
link |
we're maybe not quite at a hundred
link |
because biology is really complicated
link |
and there's always new things that we uncover
link |
that people didn't even realize existed.
link |
So, but I would say there's diseases
link |
where we might be in the 70s or 80s
link |
and then there's diseases in which I would say
link |
probably the majority where we're really close to zero.
link |
Would Alzheimer's and schizophrenia
link |
and type two diabetes fall closer to zero or to the 80?
link |
I think Alzheimer's is probably closer to zero than to 80.
link |
There are hypotheses, but I don't think
link |
those hypotheses have as of yet been sufficiently validated
link |
that we believe them to be true
link |
and there is an increasing number of people
link |
who believe that the traditional hypotheses
link |
might not really explain what's going on.
link |
I would also say that Alzheimer's and schizophrenia
link |
in even type two diabetes are not really one disease.
link |
They're almost certainly a heterogeneous collection
link |
of mechanisms that manifest in clinically similar ways.
link |
So in the same way that we now understand
link |
that breast cancer is really not one disease,
link |
it is multitude of cellular mechanisms,
link |
all of which ultimately translate
link |
to uncontrolled proliferation, but it's not one disease.
link |
The same is almost undoubtedly true
link |
for those other diseases as well.
link |
And that understanding that needs to precede
link |
any understanding of the specific mechanisms
link |
of any of those other diseases.
link |
Now, in schizophrenia, I would say
link |
we're almost certainly closer to zero than to anything else.
link |
Type two diabetes is a bit of a mix.
link |
There are clear mechanisms that are implicated
link |
that I think have been validated
link |
that have to do with insulin resistance and such,
link |
but there's almost certainly there as well,
link |
many mechanisms that we have not yet understood.
link |
So you've also thought and worked a little bit
link |
on the longevity side.
link |
Do you see the disease and longevity
link |
as overlapping completely, partially,
link |
or not at all as efforts?
link |
Those mechanisms are certainly overlapping.
link |
There's a well known phenomenon that says that
link |
for most diseases, other than childhood diseases,
link |
the risk for contracting that disease
link |
increases exponentially year on year,
link |
every year from the time you're about 40.
link |
So obviously there is a connection
link |
between those two things.
link |
That's not to say that they're identical.
link |
There's clearly aging that happens
link |
that is not really associated with any specific disease.
link |
And there's also diseases and mechanisms of disease
link |
that are not specifically related to aging.
link |
So I think overlap is where we're at.
link |
Okay, it is a little unfortunate that we get older.
link |
And it seems that there's some correlation
link |
with the occurrence of diseases
link |
or the fact that we get older and both are quite sad.
link |
Well, I mean, there's processes that happen as cells age
link |
that I think are contributing to disease.
link |
Some of those have to do with DNA damage
link |
that accumulates as cells divide
link |
where the repair mechanisms don't fully correct for those.
link |
There are accumulations of proteins
link |
that are misfolded and potentially aggregate
link |
and those two contribute to disease
link |
and contribute to inflammation.
link |
There is a multitude of mechanisms
link |
that have been uncovered that are sort of wear and tear
link |
at the cellular level that contribute
link |
to disease processes and I'm sure there's many
link |
that we don't yet understand.
link |
On a small tangent, perhaps philosophical,
link |
the fact that things get older
link |
and the fact that things die
link |
is a very powerful feature for the growth of new things
link |
that, you know, it's a kind of learning mechanism.
link |
So it's both tragic and beautiful.
link |
So, do you, so in, you know, in trying to fight disease
link |
and trying to fight aging,
link |
do you think about sort of the useful fact of our mortality
link |
or would you, like if you were, could be immortal?
link |
Would you choose to be immortal?
link |
Again, I think immortal is a very long time
link |
and I don't know that that would necessarily be something
link |
that I would want to aspire to,
link |
but I think all of us aspire to an increased health span,
link |
I would say, which is an increased amount of time
link |
where you're healthy and active
link |
and feel as you did when you were 20,
link |
we're nowhere close to that.
link |
People deteriorate physically and mentally over time
link |
and that is a very sad phenomenon.
link |
So I think a wonderful aspiration would be
link |
if we could all live to, you know, the biblical 120
link |
maybe in perfect health would be high quality of life.
link |
I think that would be an amazing goal for us
link |
to achieve as a society,
link |
now is the right age 120 or 100 or 150.
link |
I think that's up for debate,
link |
but I think an increased health span is a really worthy goal.
link |
And anyway, in a grand time of the age of the universe,
link |
it's all pretty short.
link |
So from the perspective that you've done,
link |
obviously a lot of incredible work on machine learning.
link |
So what role do you think data and machine learning
link |
play in this goal of trying to understand diseases
link |
and trying to eradicate diseases?
link |
Up until now, I don't think it's played
link |
very much of a significant role
link |
because largely the data sets that one really needed
link |
to enable a powerful machine learning methods,
link |
those data sets haven't really existed.
link |
There's been dribs and drabs
link |
and some interesting machine learning
link |
that has been applied, I would say machine learning
link |
slash data science,
link |
but the last few years are starting to change that.
link |
So we now see an increase in some large data sets,
link |
but equally importantly, an increase in technologies
link |
that are able to produce data at scale.
link |
It's not typically the case that people have deliberately,
link |
proactively use those tools for the purpose
link |
of generating data for machine learning.
link |
They, to the extent that those techniques
link |
have been used for data production,
link |
they've been used for data production
link |
to drive scientific discovery.
link |
And the machine learning came as a sort of
link |
byproduct second stage of, oh, you know,
link |
now we have a data set less to machine learning on that
link |
rather than a more simplistic data analysis method.
link |
But what we are doing it in Cetro
link |
is actually flipping that around and saying,
link |
here's this incredible repertoire of methods
link |
that bioengineers, cell biologists have come up with.
link |
Let's see if we can put them together in brand new ways
link |
with the goal of creating data sets
link |
that machine learning can really be applied on productively
link |
to create powerful predictive models
link |
that can help us address fundamental problems
link |
So really focus, make data the primary focus
link |
and the primary goal and find,
link |
use the mechanisms of biology and chemistry
link |
to create the kinds of data set
link |
that could allow machine learning to benefit the most.
link |
I wouldn't put it in those terms
link |
because that says that data is the end goal.
link |
Data is the means.
link |
So for us, the end goal is helping address challenges
link |
And the method that we've elected to do that
link |
is to apply machine learning to build predictive models.
link |
And machine learning, in my opinion,
link |
can only be really successfully applied,
link |
especially the more powerful models,
link |
if you give it data that is of sufficient scale
link |
and sufficient quality.
link |
So how do you create those data sets
link |
so as to drive the ability to generate predictive models
link |
which subsequently help improve human health?
link |
So before we dive into the details of that,
link |
let me take us to back and ask when and where
link |
was your interest in human health born?
link |
Are there moments, events, perhaps,
link |
if I may ask, tragedies in your own life
link |
that catalyze this passion,
link |
or was it the broader desire to help humankind?
link |
So I would say it's a bit of both.
link |
So on, I mean, my interest in human health
link |
actually dates back to the early 2000s
link |
when a lot of my peers
link |
in machine learning and I were using data sets
link |
that frankly were not very inspiring.
link |
Some of us old timers still remember
link |
the quote, unquote, 20 news groups data set
link |
where this was literally a bunch of texts
link |
from 20 news groups,
link |
a concept that doesn't really even exist anymore.
link |
And the question was, can you classify
link |
which news group a particular bag of words came from?
link |
And it wasn't very interesting.
link |
The data sets at the time on the biology side
link |
were much more interesting both from a technical
link |
and also from an aspirational perspective.
link |
They were still pretty small,
link |
but they were better than 20 news groups.
link |
And so I started out, I think, just by wanting to do something
link |
that was more, I don't know,
link |
societally useful and technically interesting.
link |
And then over time became more and more interested
link |
in the biology and the human health aspects for themselves
link |
and began to work even sometimes on papers
link |
that were just in biology
link |
without having a significant machine learning component.
link |
I think my interest in drug discovery
link |
is partly due to an incident I had
link |
with when my father sadly passed away about 12 years ago.
link |
He had an autoimmune disease that settled in his lungs.
link |
And the doctor's basics said, well,
link |
there's only one thing that we could do,
link |
which is give him prednisone.
link |
At some point, I remember a doctor even came and said,
link |
hey, let's do a lung biopsy to figure out
link |
which autoimmune disease he has.
link |
And I said, would that be helpful?
link |
Would that change treatment?
link |
I said, no, there's only prednisone.
link |
That's the only thing we can give him.
link |
And I had friends who were rheumatologists who said,
link |
the FDA would never approve prednisone today
link |
because the ratio of side effects to the brain
link |
to the ratio of side effects to benefit
link |
is probably not large enough.
link |
Today, we're in a state where there's probably four or five,
link |
maybe even more, well, depends for which autoimmune disease,
link |
but there are multiple drugs
link |
that can help people with autoimmune disease
link |
that many of which didn't exist at 12 years ago.
link |
And I think we're at a golden time
link |
in some ways in drug discovery
link |
where there's the ability to create drugs
link |
that are much more safe and much more effective
link |
than we've ever been able to before.
link |
And what's lacking is enough understanding
link |
of biology and mechanism to know where to aim that engine.
link |
And I think that's where machine learning can help.
link |
So in 2018, you started and now lead a company in Citro,
link |
which is like you mentioned, perhaps the focus
link |
is drug discovery and the utilization
link |
of machine learning for drug discovery.
link |
So you mentioned that, quote,
link |
we're really interested in creating what you might call
link |
a disease in a dish model, disease in a dish models,
link |
places where diseases are complex,
link |
where we really haven't had a good model system
link |
or typical animal models that have been used for years,
link |
including testing on mice, just aren't very effective.
link |
So can you try to describe what is an animal model
link |
and what is a disease in a dish model?
link |
So an animal model for disease is where you create,
link |
effectively, it's what it sounds like.
link |
It's oftentimes a mouse
link |
where we have introduced some external perturbation
link |
that creates the disease.
link |
And then we cure that disease.
link |
And the hope is that by doing that,
link |
we will cure a similar disease in the human.
link |
The problem is that oftentimes the way in which
link |
we generate the disease in the animal
link |
has nothing to do with how that disease actually comes
link |
It's what you might think of as a copy of the phenotype,
link |
a copy of the clinical outcome,
link |
but the mechanisms are quite different.
link |
And so curing the disease in the animal,
link |
which in most cases doesn't happen naturally.
link |
Mice don't get Alzheimer's, they don't get diabetes,
link |
they don't get atherosclerosis,
link |
they don't get autism or schizophrenia.
link |
Those cures don't translate over
link |
to what happens in the human.
link |
And that's where most drugs fails
link |
just because the findings that we had in the mouse
link |
don't translate to a human.
link |
The disease in the dish models is a fairly new approach.
link |
It's been enabled by technologies
link |
that have not existed for more than five to 10 years.
link |
So for instance, the ability for us to take a cell
link |
from any one of us, you or me,
link |
revert that say skin cell to what's called stem cell status,
link |
which is what's called a pluripotent cell
link |
that can then be differentiated into different types of cells.
link |
So from that pluripotent cell,
link |
one can create a lex neuron or a lex cardiomyocytes
link |
or a lex hepatocyte that has your genetics,
link |
but that right cell type.
link |
And so if there's a genetic burden of disease
link |
that would manifest in that particular cell type,
link |
you might be able to see it by looking at those cells
link |
and saying, oh, that's what potentially sick cells
link |
look like versus healthy cells
link |
and understand how and then explore what kind of interventions
link |
might revert the unhealthy looking cell to a healthy cell.
link |
Now, of course, curing cells is not the same as curing people.
link |
And so there's still potentially a translatability gap,
link |
but at least for diseases that are driven,
link |
say by human genetics,
link |
and where the human genetics
link |
is what drives the cellular phenotype,
link |
there is some reason to hope that if we revert those cells
link |
in which the disease begins
link |
and where the disease is driven by genetics
link |
and we can revert that cell back to a healthy state,
link |
maybe that will help also revert
link |
the more global clinical phenotypes.
link |
So that's really what we're hoping to do.
link |
That step, that backward step,
link |
I was reading about it, the Yamanaka factor.
link |
The reverse step back to stem cells.
link |
It seems like magic.
link |
Honestly, before that happened,
link |
I think very few people would have predicted that
link |
Can you maybe elaborate, is it actually possible?
link |
Like where, like how state,
link |
so this result was maybe like,
link |
I don't know how many years ago,
link |
maybe 10 years ago was first demonstrated,
link |
something like that.
link |
Is this, how hard is this?
link |
Like how noisy is this backward step?
link |
It seems quite incredible and cool.
link |
It is incredible and cool.
link |
It was much more, I think finicky and bespoke
link |
at the early stages when the discovery was first made,
link |
but at this point it's become almost industrialized.
link |
There are what's called contract research organizations,
link |
vendors that will take a sample from a human
link |
and revert it back to stem cell status
link |
and it works a very good fraction of the time.
link |
Now there are people who will ask,
link |
I think good questions, is this really truly a stem cell
link |
or does it remember certain aspects of changes
link |
that were made in the human beyond the genetics?
link |
It's passed as a skin cell, yeah.
link |
It's passed as a skin cell or it's passed
link |
in terms of exposures to different
link |
environmental factors and so on.
link |
So I think the consensus right now
link |
is that these are not always perfect
link |
and there is little bits and pieces of memory sometimes,
link |
but by and large these are actually pretty good.
link |
So one of the key things,
link |
well maybe you can correct me,
link |
but one of the useful things for machine learning
link |
is size, scale of data.
link |
How easy it is to do these kinds of reversals
link |
to stem cells and then does these in a dish models at scale?
link |
Is this a huge challenge or not?
link |
So the reversal is not as of this point,
link |
something that can be done at the scale of tens of thousands
link |
or hundreds of thousands.
link |
I think total number of stem cells or IPS cells
link |
that are what's called induced pluripotent stem cells
link |
in the world I think is somewhere between five
link |
and 10,000 last I looked.
link |
Now again, that might not count things that exist
link |
and this or that academic center
link |
and they may add up to a bit more,
link |
but that's about the range.
link |
So it's not something that you could at this point
link |
generate IPS cells from a million people,
link |
but maybe you don't need to
link |
because maybe that background is enough
link |
because it can also be now perturbed in different ways
link |
and some people have done really interesting experiments
link |
in for instance, taking cells from a healthy human
link |
and then introducing a mutation into it
link |
using one of the other miracle technologies
link |
that's emerged in the last decade,
link |
which is CRISPR gene editing and introduced a mutation
link |
that is known to be pathogenic.
link |
And so you can now look at the healthy cells
link |
and unhealthy cells, the one with the mutation
link |
and do a one on one comparison
link |
where everything else is held constant.
link |
And so you could really start to understand specifically
link |
what the mutation does at the cellular level.
link |
So the IPS cells are a great starting point
link |
and obviously more diversity is better
link |
because you also wanna capture ethnic background
link |
and how that affects things,
link |
but maybe you don't need one from every single patient
link |
with every single type of disease
link |
because we have other tools at our disposal.
link |
Well, how much difference is there between people
link |
and mentioned ethnic background in terms of IPS cells?
link |
Like it seems like these magical cells
link |
that can create anything between different populations,
link |
Is there a lot of variability between stem cells?
link |
Well, first of all, there's the variability
link |
that's driven simply by the fact
link |
that genetically we're different.
link |
So a stem cell that's derived from my genotype
link |
is gonna be different from a stem cell
link |
that's derived from your genotype.
link |
There's also some differences
link |
that have more to do with, for whatever reason,
link |
some people's stem cells differentiate better
link |
than other people's stem cells.
link |
We don't entirely understand why.
link |
So there's certainly some differences there as well.
link |
But the fundamental difference
link |
and the one that we really care about
link |
and is a positive is that the fact
link |
that the genetics are different
link |
and therefore recapitulate my disease burden
link |
versus your disease burden.
link |
What's the disease burden?
link |
Well, a disease burden is just, if you think,
link |
I mean, it's not a well defined mathematical term,
link |
although there are mathematical formulations of it.
link |
If you think about the fact
link |
that some of us are more likely
link |
to get a certain disease than others
link |
because we have more variations in our genome
link |
that are causative of the disease,
link |
maybe fewer that are protective of the disease.
link |
People have quantified that
link |
using what are called polygenic risk scores,
link |
which look at all of the variations
link |
in an individual person's genome
link |
and add them all up in terms of how much risk
link |
they confer for a particular disease.
link |
And then they've put people on a spectrum
link |
of their disease risk.
link |
And for certain diseases where we've been sufficiently
link |
powered to really understand the connection
link |
between the many, many small variations
link |
that give rise to an increased disease risk,
link |
there is some pretty significant differences
link |
in terms of the risk between the people,
link |
say at the highest decile of this polygenic risk score
link |
and the people at the lowest decile.
link |
Sometimes those other differences are, you know,
link |
factor of 10 or 12 higher.
link |
So there's definitely a lot that our genetics
link |
contributes to disease risk,
link |
even if it's not by any stretch, the full explanation.
link |
And from a machinery perspective, there's signal there.
link |
There is definitely signal in the genetics.
link |
And there's even more signal, we believe,
link |
in looking at the cells that are derived
link |
from those different genetics, because in principle,
link |
you could say all the signal is there at the genetics level.
link |
So we don't need to look at the cells,
link |
but our understanding of the biology is so limited
link |
at this point, then seeing what actually happens
link |
at the cellular level is a heck of a lot closer
link |
to the human clinical outcome than looking
link |
at the genetics directly, and so we can learn
link |
a lot more from it than we could by looking at genetics alone.
link |
So just to get a sense, I don't know if it's easy to do,
link |
but what kind of data is useful
link |
in this disease in a dish model?
link |
Like what are, what's the source of raw data information?
link |
And also, from my outsider's perspective,
link |
sort of biology and cells are squishy things.
link |
They're literally squishy things.
link |
How do you connect the computer to that?
link |
Which sensory mechanisms, I guess.
link |
So that's another one of those revolutions
link |
that have happened in the last 10 years,
link |
in that our ability to measure cells very quantitatively
link |
has also dramatically increased.
link |
So back when I started doing biology in the late 90s,
link |
early 2000s, that was the initial era
link |
where we started to measure biology
link |
in really quantitative ways using things like microarrays,
link |
where you would measure, in a single experiment,
link |
the activity level, what's called expression level
link |
of every gene in the genome in that sample.
link |
And that ability is what actually allowed us
link |
to even understand that there are molecular subtypes
link |
of diseases like cancer, where up until that point,
link |
it's like, oh, you have breast cancer.
link |
But then when we looked at the molecular data,
link |
it was clear that there's different subtypes
link |
of breast cancer that at the level of gene activity
link |
look completely different to each other.
link |
So that was the beginning of this process.
link |
Now we have the ability to measure individual cells
link |
in terms of their gene activity
link |
using what's called single cell RNA sequencing,
link |
which basically sequences the RNA,
link |
which is that activity level of different genes
link |
for every gene in a genome.
link |
And you could do that at single cell levels.
link |
That's an incredibly powerful way of measuring cells.
link |
I mean, you literally count the number of transcripts.
link |
So it really turns that squishy thing
link |
into something that's digital.
link |
Another tremendous data source
link |
that's emerged in the last few years is microscopy
link |
and specifically even super resolution microscopy,
link |
where you could use digital reconstruction
link |
to look at subcellular structure,
link |
sometimes even things that are below
link |
the diffraction limit of light
link |
by doing sophisticated reconstruction.
link |
And again, that gives you tremendous amount of information
link |
at the subcellular level.
link |
There's now more and more ways
link |
that amazing scientists out there are developing
link |
for getting new types of information from even single cells.
link |
And so that is a way of turning those squishy things
link |
into digital data.
link |
Into beautiful data sets.
link |
But so that data set then with machine learning tools
link |
allows you to maybe understand the developmental,
link |
like the mechanism of a particular disease.
link |
And if it's possible to sort of at a high level describe,
link |
how does that help
link |
lead to a drug discovery
link |
that can help prevent, reverse that mechanism?
link |
So I think there's different ways
link |
in which this data could potentially be used.
link |
Some people use it for scientific discovery and say,
link |
oh, look, we see this phenotype at the cellular level.
link |
So let's try and work our way backwards
link |
and think which genes might be involved in pathways
link |
that give rise to that.
link |
So that's a very sort of analytical method
link |
to sort of work our way backwards
link |
using our understanding of no biology.
link |
Some people use it in a somewhat more sort of forward.
link |
If that was backward, this would be forward,
link |
which is to say, okay, if I can perturb this gene,
link |
does it show a phenotype
link |
that is similar to what I see in disease patients?
link |
And so maybe that gene is actually causal of the disease.
link |
So that's a different way.
link |
And then there's what we do,
link |
which is basically to take that very large collection of data
link |
and use machine learning to uncover the patterns
link |
that emerge from it.
link |
So for instance, what are those subtypes
link |
that might be similar at the human clinical outcome
link |
but quite distinct when you look at the molecular data?
link |
And then if we can identify such a subtype,
link |
are there interventions that if I apply it to cells
link |
that come from this subtype of the disease
link |
and you apply that intervention, it could be a drug
link |
or it could be a CRISPR gene intervention,
link |
does it revert the disease state to something
link |
that looks more like normal, happy, healthy cells?
link |
And so hopefully if you see that,
link |
that gives you a certain hope that that intervention
link |
will also have a meaningful clinical benefit to people.
link |
And there's obviously a bunch of things
link |
that you would want to do after that to validate that,
link |
but it's a very different and much less hypothesis driven way
link |
of uncovering new potential interventions
link |
and might give rise to things that are not the same things
link |
that everyone else is already looking at.
link |
I don't know, I'm just like to psychoanalyze
link |
my own feeling about our discussion currently.
link |
It's so exciting to talk about, fundamentally,
link |
something that's been turned into a machine learning problem
link |
and that has so much real world impact.
link |
That's how I feel too.
link |
That's kind of exciting because I'm so,
link |
most of my days spent with data sets
link |
that I guess closer to the news groups.
link |
So it just feels good to talk about.
link |
In fact, I almost don't want to talk to you about machine learning.
link |
I want to talk about the fundamentals of the data set,
link |
which is an exciting place to be.
link |
It's what gets me up in the morning.
link |
It's also what attracts a lot of the people who work it in Cetro
link |
to in Cetro because I think all of the,
link |
certainly all of our machine learning people are outstanding
link |
and could go get a job selling ads online
link |
or doing eCommerce or even self driving cars.
link |
But I think they would want,
link |
they come to us because they want to work on something
link |
that has more of an aspirational nature
link |
and can really benefit humanity.
link |
With these approaches,
link |
what do you hope, what kind of diseases can be helped?
link |
We mentioned Alzheimer's, Schizophrenia, Type 2 Diabetes.
link |
Can you just describe the various kinds of diseases
link |
that this approach can help?
link |
Well, we don't know and I try and be very cautious
link |
about making promises about some things.
link |
Oh, we will cure X.
link |
People make that promise and I think it's,
link |
I tried to first deliver and then promise
link |
as opposed to the other way around.
link |
There are characteristics of a disease
link |
that make it more likely that this type of approach
link |
can potentially be helpful.
link |
So for instance, diseases that have a very strong genetic basis
link |
are ones that are more likely to manifest
link |
in a stem cell derived model.
link |
We would want the cellular models to be relatively reproducible
link |
and robust so that you could actually get enough of those cells
link |
in a way that isn't very highly variable and noisy.
link |
You would want the disease to be relatively contained
link |
in one or a small number of cell types
link |
that you could actually create in vitro in a dish setting,
link |
whereas if it's something that's really broad and systemic
link |
and involves multiple cells that are in very distal parts
link |
of your body, putting that all in a dish is really challenging.
link |
So we want to focus on the ones that are most likely
link |
to be successful today with the hope, I think,
link |
that really smart bioengineers out there
link |
are developing better and better systems all the time
link |
so that diseases that might not be tractable today
link |
might be tractable in three years.
link |
So for instance, five years ago,
link |
these stem cell derived models didn't really exist.
link |
People were doing most of the work in cancer cells,
link |
and cancer cells are very, very poor models
link |
of most human biology because, A, they were cancer to begin with
link |
and B, as you passage them and they proliferate in a dish,
link |
they become, because of the genomic instability,
link |
even less similar to human biology.
link |
Now we have these stem cell derived models.
link |
We have the capability to reasonably robustly,
link |
not quite at the right scale yet,
link |
but close to derive what's called organoids,
link |
which are these teeny little sort of multicellular organ,
link |
which can wrap sort of models of an organ system.
link |
So there's cerebral organoids and liver organoids
link |
and kidney organoids.
link |
Yeah, brain organoids.
link |
That organoids is possibly the coolest thing I've ever seen.
link |
Is that not like the coolest thing?
link |
And then I think on the horizon,
link |
we're starting to see things like connecting these organoids
link |
to each other so that you could actually start,
link |
and there's some really cool papers that start to do that,
link |
where you can actually start to say,
link |
okay, can we do multi organ system stuff?
link |
There's many challenges to that.
link |
It's not easy by any stretch,
link |
but I'm sure people will figure it out.
link |
And in three years or five years,
link |
there will be disease models that we could make
link |
for things that we can't make today.
link |
And this conversation would seem almost outdated
link |
with the kind of scale that could be achieved
link |
in like three years.
link |
That would be so cool.
link |
So you've cofounded Coursera with Andrew Eng,
link |
and we're part of the whole MOOC revolution.
link |
So to jump topics a little bit,
link |
can you maybe tell the origin story of the history,
link |
the origin story of MOOCs, of Coursera,
link |
and in general, you're teaching to huge audiences
link |
on a very sort of impactful topic of AI in general.
link |
So I think the origin story of MOOCs emanates
link |
from a number of efforts that occurred
link |
at Stanford University around the late 2000s,
link |
where different individuals within Stanford,
link |
myself included, were getting really excited
link |
about the opportunities of using online technologies
link |
as a way of achieving both improved quality of teaching
link |
and also improved scale.
link |
And so Andrew, for instance, led the Stanford Engineering
link |
Everywhere, which was sort of an attempt to take
link |
10 Stanford courses and put them online,
link |
just as video lectures.
link |
I led an effort within Stanford to take some of the courses
link |
and really create a very different teaching model
link |
that broke those up into smaller units
link |
and had some of those embedded interactions and so on,
link |
which got a lot of support from university leaders
link |
because they felt like it was potentially a way
link |
of improving the quality of instruction at Stanford
link |
by moving to what's now called the flipped classroom model.
link |
And so those efforts eventually started to interplay
link |
with each other and created a tremendous sense
link |
of excitement and energy within the Stanford community
link |
about the potential of online teaching
link |
and led in the fall of 2011 to the launch
link |
of the first Stanford MOOCs.
link |
By the way, MOOCs, it's probably impossible
link |
that people don't know, but I guess massive...
link |
Open online courses.
link |
Open online courses.
link |
We did not come up with the acronym.
link |
I'm not particularly fond of the acronym,
link |
but it is what it is.
link |
Big Bang is not a great term for the start of the universe,
link |
but it is what it is.
link |
So anyway, those courses launched in the fall of 2011
link |
and there were, within a matter of weeks,
link |
a real publicity campaign, just a New York Times article
link |
About 100,000 students or more in each of those courses.
link |
And I remember this conversation that Andrew and I had,
link |
which is like, wow, there's this real need here.
link |
And I think we both felt like, sure,
link |
we were accomplished academics and we could go back
link |
and go back to our labs, write more papers.
link |
But if we did that, then this wouldn't happen.
link |
And it seemed too important not to happen.
link |
And so we spent a fair bit of time debating,
link |
do we want to do this as a Stanford effort,
link |
kind of building on what we'd started?
link |
Do we want to do this as a for profit company?
link |
Do we want to do this as a nonprofit?
link |
And we decided ultimately to do it as we did with Coursera.
link |
And so, you know, we started really operating as a company
link |
at the beginning of 2012.
link |
In the rest of history.
link |
In the rest of history.
link |
But how did you, was that really surprising to you?
link |
How did you at that time, and at this time,
link |
make sense of this need for sort of global education?
link |
You mentioned that you felt that, wow,
link |
the popularity indicates that there's a hunger
link |
for sort of globalization of learning.
link |
I think there is a hunger for learning that,
link |
you know, globalization is part of it,
link |
but I think it's just a hunger for learning.
link |
The world has changed in the last 50 years.
link |
It used to be that you finished college,
link |
you got a job, by and large, the skills that you learned
link |
in college were pretty much what got you
link |
through the rest of your job history.
link |
And yeah, you learned some stuff,
link |
but it wasn't a dramatic change.
link |
Today, we're in a world where the skills that you need
link |
for a lot of jobs, they didn't even exist
link |
when you went to college, and the jobs,
link |
and many of the jobs that exist when you went to college
link |
don't even exist today, or dying.
link |
So part of that is due to AI, but not only.
link |
And we need to find a way of keeping people,
link |
giving people access to the skills that they need today.
link |
And I think that's really what's driving a lot of this hunger.
link |
So I think if we even take a step back,
link |
for you all this started in trying to think of new ways
link |
to teach, or new ways to sort of organize the material
link |
and present the material in a way that would help
link |
the education process pedagogy.
link |
So what have you learned about effective education
link |
from this process of playing, of experimenting
link |
with different ideas?
link |
So we learned a number of things,
link |
some of which I think could translate back
link |
and have translated back effectively
link |
to how people teach on campus,
link |
and some of which I think are more specific
link |
to people who learn online,
link |
more sort of people who learn as part of their daily life.
link |
So we learned, for instance, very quickly that short is better.
link |
So people who are especially in the workforce
link |
can't do a 15 week semester long course.
link |
They just can't fit that into their lives.
link |
Can you describe the shortness of what?
link |
So the little lecture short.
link |
The lecture short.
link |
The course is short.
link |
We started out, you know, the first online education efforts
link |
were actually MIT's open courseware initiatives.
link |
And that was, you know, recording of classroom lectures.
link |
Hour and a half or something like that, yeah.
link |
That didn't really work very well.
link |
I mean, some people benefit.
link |
I mean, of course they did.
link |
But it's not really very palatable experience for someone
link |
who has a job and, you know, three kids
link |
and they need to run errands and such.
link |
They can't fit 15 weeks into their life.
link |
And the hour and a half is really hard.
link |
So we learned very quickly.
link |
I mean, we started out with short video modules
link |
and over time we made them shorter
link |
because we realized that 15 minutes was still too long
link |
if you want to fit in when you're waiting in line
link |
for your kids doctor's appointment.
link |
It's better if it's five to seven.
link |
We learned that 15 week courses don't work
link |
and you really want to break this up into shorter units
link |
so that there is a natural completion point.
link |
It gives people a sense of they're really close
link |
to finishing something meaningful.
link |
They can always come back and take part two and part three.
link |
We also learned that compressing the content works really well
link |
because if some people that pace works well
link |
and for others they can always rewind and watch again.
link |
And so people have the ability to then learn at their own pace.
link |
And so that flexibility, the brevity and the flexibility
link |
are both things that we found to be very important.
link |
We learned that engagement during the content is important
link |
and the quicker you give people feedback
link |
the more likely they are to be engaged.
link |
Hence the introduction of these,
link |
which we actually was an intuition that I had going in
link |
and was then validated using data
link |
that introducing some of these sort of little micro quizzes
link |
into the lectures really helps.
link |
Self graded as automatically graded assessments
link |
really helped too because it gives people feedback.
link |
See, there you are.
link |
So all of these are valuable.
link |
And then we learned a bunch of other things too.
link |
We did some really interesting experiments,
link |
for instance on the gender bias
link |
how having a female role model as an instructor
link |
can change the balance of men to women
link |
in terms of especially in STEM courses.
link |
And you could do that online by doing A.B. testing
link |
in ways that would be really difficult to go on campus.
link |
Oh, that's exciting.
link |
But so the shortness, the compression,
link |
I mean, that's actually,
link |
so that probably is true for all good editing
link |
is always just compressing the content,
link |
making it shorter.
link |
So that puts a lot of burden on the instructor
link |
and the creator of the educational content.
link |
Probably most lectures at MIT or Stanford
link |
could be five times shorter
link |
if the preparation was put enough.
link |
So maybe people might disagree with that,
link |
but the Christmas declarity that a lot of the Moot like Coursera delivers
link |
is how much effort does that take?
link |
So first of all, let me say that it's not clear
link |
that that crispness would work as effectively
link |
in a face to face setting
link |
because people need time to absorb the material.
link |
And so you need to at least pause
link |
and give people a chance to reflect and maybe practice.
link |
And that's what MOOCs do,
link |
is that they give you these chunks of content
link |
and then ask you to practice with it.
link |
And that's where I think some of the newer pedagogy
link |
that people are adopting in face to face teaching
link |
that have to do with interactive learning and such
link |
can be really helpful.
link |
But both those approaches,
link |
whether you're doing that type of methodology
link |
in online teaching or in that flipped classroom
link |
interactive teaching.
link |
What's a side to pause? What's flipped classroom?
link |
Flipped classroom is a way in which
link |
online content is used to supplement
link |
face to face teaching where people watch the videos
link |
perhaps and do some of the exercises before coming to class
link |
and then when they come to class,
link |
it's actually to do much deeper problem solving
link |
oftentimes in a group.
link |
But any one of those different pedagogies
link |
that are beyond just standing there and droning on
link |
in front of the classroom for an hour and 15 minutes
link |
require a heck of a lot more preparation.
link |
And so it's one of the challenges, I think,
link |
that people have that we had when trying to convince
link |
instructors to teach on Coursera.
link |
And it's part of the challenges that pedagogy experts
link |
on campus have in trying to get faculty to teach
link |
different things that it's actually harder to teach
link |
that way than it is to stand there and drone.
link |
Do you think MOOCs will replace in person education
link |
or become the majority of in person
link |
of education of the way people learn in the future?
link |
Again, the future could be very far away,
link |
but where's the trend going, do you think?
link |
So I think it's a nuanced and complicated answer.
link |
I don't think MOOCs will replace face to face teaching.
link |
I think learning is in many cases a social experience
link |
and even at Coursera we had people who naturally
link |
formed study groups even when they didn't have to
link |
to just come and talk to each other.
link |
And we found that that actually benefited their learning
link |
in very important ways.
link |
So there was more success among learners who had
link |
those study groups than among ones who didn't.
link |
So I don't think it's just going to, oh,
link |
we're all going to just suddenly learn online
link |
with a computer and no one else in the same way
link |
that recorded music has not replaced live concerts.
link |
But I do think that especially when you are thinking
link |
about continuing education, the stuff that people get
link |
when they're traditional whatever high school
link |
college education is done and they yet have to
link |
maintain their level of expertise and skills
link |
in a rapidly changing world, I think people will consume
link |
more and more educational content in this online format
link |
because going back to school for formal education
link |
is not an option for most people.
link |
Briefly, it might be a difficult question to ask,
link |
but people fascinated by artificial intelligence,
link |
by machine learning, by deep learning,
link |
is there a recommendation for the next year
link |
or for a lifelong journey of somebody interested in this,
link |
how do they begin, how do they enter that learning journey?
link |
I think the important thing is first to just get started
link |
and there's plenty of online content that one can get
link |
for both the core foundations of mathematics
link |
and statistics and programming and then from there
link |
to machine learning.
link |
I would encourage people not to skip too quickly
link |
past the foundations because I find that there's a lot
link |
of people who learn machine learning whether it's online
link |
or on campus without getting those foundations
link |
and they basically just turn the crank on existing models
link |
in ways that they don't allow for a lot of innovation
link |
and adjustment to the problem at hand
link |
but also be or sometimes just wrong
link |
and they don't even realize that their application is wrong
link |
because there's artifacts that they haven't fully understood.
link |
I think the foundations, machine learning is an important step
link |
and then actually start solving problems.
link |
Try and find someone to solve them with
link |
because especially at the beginning it's useful
link |
to have someone to bounce ideas off
link |
and fix mistakes that you make
link |
and you can fix mistakes that they make
link |
but then just find practical problems
link |
whether it's in your workplace or if you don't have that
link |
Kaggle competitions or such are a really great place
link |
to find interesting problems and just practice.
link |
Perhaps a bit of a romanticized question
link |
but what idea in deep learning do you find,
link |
have you found in your journey the most beautiful
link |
or surprising or interesting?
link |
Perhaps not just deep learning but AI in general, statistics.
link |
I'm going to answer with two things.
link |
One would be the foundational concept of end to end training
link |
which is that you start from the raw data
link |
and you train something that is not a single piece
link |
but rather towards the actual goal that you're looking to...
link |
From the raw data to the outcome and no details in between.
link |
Not no details but the fact that you could certainly
link |
introduce building blocks that were trained towards other tasks
link |
and actually coming to that in my second half of the answer
link |
but it doesn't have to be a single monolithic blob in the middle
link |
actually I think that's not ideal but rather the fact that
link |
at the end of the day you can actually train something
link |
that goes all the way from the beginning to the end
link |
and the other one that I find really compelling
link |
is the notion of learning a representation
link |
that in its turn, even if it was trained to another task
link |
can potentially be used as a much more rapid starting point
link |
to solving a different task.
link |
That's, I think, reminiscent of what makes people successful learners.
link |
It's something that is relatively new in the machine learning space.
link |
I think it's underutilized even relative to today's capabilities
link |
but more and more of how do we learn reusable representation.
link |
So end to end and transfer learning,
link |
is it surprising to you that neural networks are able to
link |
in many cases do these things?
link |
Is it maybe taking back to when you first would dive deep
link |
into neural networks or in general even today,
link |
is it surprising that neural networks work at all
link |
and work wonderfully to do this kind of raw
link |
end to end learning and even transfer learning?
link |
I think I was surprised by how well
link |
when you have large enough amounts of data,
link |
it's possible to find a meaningful representation
link |
in what is an exceedingly high dimensional space.
link |
So I find that to be really exciting
link |
and people are still working on the math for that.
link |
There's more papers on that every year
link |
and I think it would be really cool if we figured that out.
link |
But that to me was a surprise because in the early days
link |
when I was starting my way in machine learning
link |
and the data sets were rather small,
link |
I think we believed, I believe,
link |
that you needed to have a much more constrained
link |
and knowledge rich search space
link |
to really get to a meaningful answer.
link |
And I think it was true at the time.
link |
What I think is still a question is
link |
will a completely knowledge free approach
link |
where there's no prior knowledge
link |
going into the construction of the model,
link |
is that going to be the solution or not?
link |
It's not actually the solution today in the sense
link |
that the architecture of a convolutional neural network
link |
that's used for images is actually quite different
link |
to the type of network that's used for language
link |
and yet different from the one that's used for speech
link |
or biology or any other application.
link |
There's still some insight that goes into the structure
link |
of the network to get to the right performance.
link |
Will you be able to come up with a universal learning machine?
link |
I wonder if there's always has to be some insight
link |
injected somewhere or whether it can converge.
link |
So you've done a lot of interesting work
link |
with probably the graphical models
link |
in general, Bayesian deep learning and so on.
link |
Can you maybe speak high level?
link |
How can learning systems deal with uncertainty?
link |
One of the limitations, I think,
link |
of a lot of machine learning models
link |
is that they come up with an answer
link |
and you don't know how much you can believe that answer
link |
and oftentimes the answer is actually
link |
quite poorly calibrated relative to its uncertainties
link |
even if you look at where the confidence
link |
that comes out of say the neural network at the end
link |
and you ask how much more likely is an answer
link |
of 0.8 versus 0.9.
link |
It's not really in any way calibrated
link |
to the actual reliability of that network
link |
and how true it is.
link |
And the further away you move from the training data,
link |
the more, not only the more wrong the network is,
link |
often it's more wrong and more confident
link |
in its wrong answer.
link |
And that is a serious issue in a lot of application areas.
link |
So when you think, for instance, about medical diagnosis
link |
as being maybe an epitome of how problematic this can be,
link |
if you were training your network on a certain set of patients
link |
in a certain patient population
link |
and I have a patient that is an outlier
link |
and there's no human that looks at this
link |
and that patient is put into a neural network
link |
and your network not only gives a completely incorrect diagnosis
link |
but is supremely confident in its wrong answer,
link |
you could kill people.
link |
So I think creating more of an understanding
link |
of how do you produce networks that are calibrated
link |
in their uncertainty and can also say,
link |
you know what, I give up.
link |
I don't know what to say about this particular data instance
link |
because I've never seen something
link |
that's sufficiently like it before.
link |
I think it's going to be really important in mission critical applications
link |
especially ones where human life is at stake
link |
and that includes medical applications
link |
but it also includes automated driving
link |
because you'd want the network to be able to say,
link |
you know what, I have no idea what this blob is
link |
that I'm seeing in the middle of the rest.
link |
I'm just going to stop
link |
because I don't want to potentially run over a pedestrian
link |
that I don't recognize.
link |
Is there good mechanisms, ideas of how to allow learning systems
link |
to provide that uncertainty along with their predictions?
link |
Certainly people have come up with mechanisms
link |
that involve Bayesian deep learning,
link |
deep learning that involves Gaussian processes.
link |
I mean, there's a slew of different approaches
link |
that people have come up with.
link |
There's methods that use ensembles of networks
link |
trained with different subsets of data,
link |
different random starting points.
link |
Those are actually sometimes surprisingly good
link |
at creating a set of how confident or not you are in your answer.
link |
It's very much an area of open research.
link |
Let's cautiously venture back into the land of philosophy
link |
and speaking of AI systems providing uncertainty
link |
to somebody like Stuart Russell believes that
link |
as we create more and more intelligent systems,
link |
it's really important for them to be full of self doubt
link |
because if they're given more and more power,
link |
the way to maintain human control over AI systems
link |
or human supervision, which is true,
link |
like you just mentioned with autonomous vehicles,
link |
it's really important to get human supervision
link |
when the car is not sure because if it's really confident
link |
in cases when it can get in trouble,
link |
it's going to be really problematic.
link |
Let me ask about sort of the questions of AGI
link |
and human level intelligence.
link |
I mean, we've talked about curing diseases,
link |
which is a sort of fundamental thing.
link |
We can have an impact today,
link |
but AI people also dream of both understanding
link |
and creating intelligence.
link |
Is that something you think about?
link |
Is that something you dream about?
link |
Is that something you think is within our reach
link |
to be thinking about as computer scientists?
link |
Boy, let me tease apart different parts of that question.
link |
That is the worst question.
link |
Yeah, it's a multi part question.
link |
So let me start with the feasibility of AGI.
link |
Then I'll talk about the timelines a little bit
link |
and then talk about what controls does one need
link |
when thinking about protections in the AI space.
link |
So I think AGI obviously is a longstanding dream
link |
that even our early pioneers in the space had.
link |
The Turing test and so on are the earliest discussions of that.
link |
We're obviously closer than we were 70 or so years ago,
link |
but I think it's still very far away.
link |
I think machine learning algorithms today
link |
are really exquisitely good pattern recognizers
link |
in very specific problem domains
link |
where they have seen enough training data
link |
to make good predictions.
link |
You take a machine learning algorithm
link |
and you move it to a slightly different version
link |
of even that same problem, far less one that's different
link |
and it will just completely choke.
link |
So I think we're nowhere close to the versatility
link |
and flexibility of even a human toddler
link |
in terms of their ability to context switch
link |
and have different problems using a single knowledge base,
link |
So am I desperately worried about the machines taking over
link |
the universe and starting to kill people
link |
because they want to have more power?
link |
Well, to pause on that,
link |
you've kind of intuited that superintelligence
link |
is a very difficult thing to achieve.
link |
Even intelligence.
link |
Superintelligence, we're not even close to intelligence.
link |
Even just the greater abilities of generalization
link |
of ours current systems.
link |
But we haven't answered all the parts.
link |
I'm getting to the second part.
link |
But maybe another tangent you can also pick up
link |
is can we get in trouble with much dumber systems?
link |
Yes, and that is exactly where I was going.
link |
So just to wrap up on the threats of AGI,
link |
I think that it seems to me a little early today
link |
to figure out protections against a human level
link |
or superhuman level intelligence
link |
where we don't even see the skeleton
link |
of what that would look like.
link |
So it seems that it's very speculative
link |
on how to protect against that.
link |
But we can definitely and have gotten into trouble
link |
on much dumber systems.
link |
And a lot of that has to do with the fact that
link |
the systems that we're building are increasingly complex,
link |
increasingly poorly understood,
link |
and there's ripple effects that are unpredictable
link |
in changing little things
link |
that can have dramatic consequences on the outcome.
link |
And by the way, that's not unique to artificial intelligence.
link |
I think artificial intelligence exacerbates that,
link |
brings it to a new level.
link |
But heck, our electric grid is really complicated.
link |
The software that runs our financial markets
link |
is really complicated.
link |
And we've seen those ripple effects translate
link |
to dramatic negative consequences,
link |
like for instance, financial crashes
link |
that have to do with feedback loops
link |
that we didn't anticipate.
link |
So I think that's an issue that we need to be thoughtful about
link |
Artificial intelligence being one of them.
link |
And we should, and I think it's really important
link |
that people are thinking about ways in which
link |
we can have better interpretability of systems,
link |
better tests for, for instance, measuring the extent
link |
to which a machine learning system that was trained
link |
in one set of circumstances.
link |
How well does it actually work
link |
in a very different set of circumstances
link |
where you might say, for instance,
link |
well, I'm not going to be able to test my automated vehicle
link |
in every possible city, village, weather condition, and so on.
link |
But if you trained it on this set of conditions
link |
and then tested it on 50 or 100 others
link |
that were quite different from the ones that you trained it on,
link |
and it worked, then that gives you confidence
link |
that the next 50 that you didn't test it on might also work.
link |
Effectively, it's testing for generalizability.
link |
So I think there's ways that we should be constantly thinking about
link |
to validate the robustness of our systems.
link |
I think it's very different from the,
link |
let's make sure robots don't take over the world.
link |
And then the other place where I think we have a threat,
link |
which is also important for us to think about,
link |
is the extent to which technology can be abused.
link |
So like any really powerful technology,
link |
machine learning can be very much used badly as well as to good.
link |
And that goes back to many other technologies
link |
that I've come up with when people invented projectile missiles
link |
and it turned into guns.
link |
And people invented nuclear power and it turned into nuclear bombs.
link |
And I think, honestly, I would say that to me,
link |
gene editing in CRISPR is at least as dangerous
link |
at technology if used badly as machine learning.
link |
You could create really nasty viruses and such using gene editing
link |
that you would be really careful about.
link |
So anyway, that's something that we need to be really thoughtful
link |
about whenever we have any really powerful new technology.
link |
And in the case of machine learning is adversarial machine learning,
link |
so all the kinds of attacks like security almost threats.
link |
And there's a social engineering with machine learning algorithms.
link |
And there's face recognition and Big Brother is watching you.
link |
And there's the killer drones that can potentially go
link |
and targeted execution of people in a different country.
link |
I don't want to argue that bombs are not necessarily that much better,
link |
but if people want to kill someone, they'll find a way to do it.
link |
So in general, if you look at trends in the data,
link |
there's less wars, there's less violence, there's more human rights.
link |
So we've been doing overall quite good as a human species.
link |
Surprisingly sometimes.
link |
Are you optimistic?
link |
Maybe another way to ask is, do you think most people are good
link |
and fundamentally we tend towards a better world,
link |
which is underlying the question, will machine learning,
link |
with gene editing ultimately land us somewhere good?
link |
Are you optimistic?
link |
I think by and large, I'm optimistic.
link |
I think that most people mean well.
link |
That doesn't mean that most people are altruistic do gooders,
link |
but I think most people mean well.
link |
But I think it's also really important for us as a society
link |
to create social norms where doing good
link |
and being perceived well by our peers
link |
are positively correlated.
link |
I mean, it's very easy to create dysfunctional societies.
link |
There's certainly multiple psychological experiments,
link |
as well as sadly real world events where people have devolved
link |
to a world where being perceived well by your peers
link |
is correlated with really atrocious, often genocidal behaviors.
link |
So we really want to make sure that we maintain a set of social norms
link |
where people know that to be a successful number of society,
link |
you want to be doing good.
link |
And one of the things that I sometimes worry about is
link |
that some societies don't seem to necessarily be moving
link |
in the forward direction in that regard,
link |
where it's not necessarily the case that doing,
link |
that being a good person is what makes you be perceived well by your peers.
link |
And I think that's a really important thing for us as a society to remember.
link |
It's very easy to degenerate back into a universe
link |
where it's okay to do really bad stuff
link |
and still have your peers think you're amazing.
link |
It's fun to ask a world class computer scientist and engineer
link |
a ridiculously philosophical question like, what is the meaning of life?
link |
Let me ask, what gives your life meaning?
link |
What is the source of fulfillment, happiness, joy, purpose?
link |
When we were starting Coursera in the fall of 2011,
link |
that was right around the time that Steve Jobs passed away.
link |
And so the media was full of various famous quotes that he uttered.
link |
And one of them that really stuck with me
link |
because it resonated with stuff that I'd been feeling for even years before that
link |
is that our goal in life should be to make a dent in the universe.
link |
So I think that to me, what gives my life meaning
link |
is that I would hope that when I am lying there on my deathbed
link |
and looking at what I've done in my life
link |
that I can point to ways in which I have left the world a better place
link |
than it was when I entered it.
link |
This is something I tell my kids all the time
link |
because I also think that the burden of that is much greater
link |
for those of us who were born to privilege and in some ways I was.
link |
I mean, it wasn't born super wealthy or anything like that,
link |
but I grew up in an educated family with parents who loved me and took care of me
link |
and I had a chance at a great education and so I always had enough to eat.
link |
So I was in many ways born to privilege more than the vast majority of humanity.
link |
And my kids I think are even more so born to privilege
link |
than I was fortunate enough to be.
link |
And I think it's really important that especially for those of us who have that opportunity
link |
that we use our lives to make the world a better place.
link |
I don't think there's a better way to end it.
link |
Daphne is honored to talk to you.
link |
Thank you so much for talking to me.
link |
And now let me leave you with some words from Hippocrates, a physician from ancient Greece
link |
who is considered to be the father of medicine.
link |
Wherever the art of medicine is loved, there's also love of humanity.
link |
Thank you for listening and hope to see you next time.