Lex Fridman PodcastDaphne Koller: Biomedicine and Machine Learning | Lex Fridman Podcast #93
EVERY SPOKEN WORD
115 min read · 23,069 words- 0:00 – 2:22
Introduction
- LFLex Fridman
The following is a conversation with Daphne Koller, a professor of computer science at Stanford University, a co-founder of Coursera with Andrew Ng, and founder and CEO of insitro, a company at the intersection of machine learning and biomedicine. We're now in the exciting early days of using the data-driven methods of machine learning to help discover and develop new drugs and treatments at scale. Daphne and insitro are leading the way on this with breakthroughs that may ripple through all fields of medicine, including ones most critical for helping with the current coronavirus pandemic. This conversation was recorded before the COVID-19 outbreak. For everyone feeling the medical, psychological, and financial burden of this crisis, I'm sending love your way. Stay strong. We're in this together. We'll beat this thing. This is the Artificial Intelligence podcast. If you enjoy it, subscribe on YouTube, review it with five stars on Apple podcasts, support it on Patreon, or simply connect with me on Twitter, @lexfridman, spelled F-R-I-D-M-A-N. As usual, I'll do a few minutes of ads now, and never any ads in the middle that can break the flow of this conversation. I hope that works for you and doesn't hurt the listening experience. This show is presented by Cash App, the number one finance app in the App Store. When you get it, use code LEXPODCAST. Cash App lets you send money to friends, buy Bitcoin, and invest in the stock market with as little as $1. Since Cash App allows you to send and receive money digitally peer-to-peer, and security in all digital transactions is very important, let me mention that PCI data security standard that Cash App is compliant with. I'm a big fan of standards for safety and security. PCI DSS is a good example of that, where a bunch of competitors got together and agreed that there needs to be a global standard around the security of transactions. Now we just need to do the same for autonomous vehicles and AI systems in general. So again, if you get Cash App from the App Store or Google Play, and use the code LEXPODCAST, you get $10 and Cash App will also donate $10 to FIRST, an organization that is helping to advance robotics and STEM education for young people around the world. And now, here's my conversation with Daphne Koller.
- 2:22 – 6:31
Will we one day cure all disease?
- LFLex Fridman
So you co-founded Coursera and made a huge impact in the global education of AI, and after five years, in August 2016, wrote a blog post saying that you're stepping away and wrote, quote, "It is time for me to turn to another critical challenge, the development of machine learning and its applications to improving human health." So let me ask two far out philosophical questions. One, do you think we will one day find cures for all major diseases known today? And two, do you think we will one day figure out a way to extend the human lifespan, perhaps to the point of immortality?
- DKDaphne Koller
So one day is a very long time-
- LFLex Fridman
(laughs)
- DKDaphne Koller
... and I don't like to make predictions of the type we will never be able to do X because I think that's a, uh, you know, that's- that smacks of hubris. It seems that never in the inf- in the entire eternity of human existence will we be able to solve a problem. That being said, curing disease is very hard, because oftentimes by the time you discover the disease, a lot of damage has already been done. And so to assume that we would be able to cure disease at that stage assumes that we would come up with ways of basically regenerating entire parts of the human body in a way that actually returns it to its original state, and that's a very challenging problem. We've cured very few diseases. We've been able to provide treatment for an increasingly large number, but the number of things that you could actually define to be cures is actually not that large. Um, so I think that's- it- there's a lot of work that would need to happen before one could legitimately say that we have cured even a reasonable number, far less all diseases.
- LFLex Fridman
On the scale of 0 to 100, where are we in understanding the fundamental mechanisms of all- of major diseases? What's your sense? So from the computer science perspective that you've entered the world of health, how far along are we?
- DKDaphne Koller
I think it depends on which disease. I mean, there are ones where I would say we're maybe not quite at 100 because biology is really complicated and there's always new things that we uncover that people didn't even realize existed. Um, so but I would say there's diseases where we might be in the 70s or 80s, um, and then there's diseases in which I would say, with g- probably the majority, where we're really close to 0.
- LFLex Fridman
Would, uh, Alzheimers and schizophrenia and type 2 diabetes fall closer to 0 or to the 80?
- DKDaphne Koller
I think Alzheimers is probably closer to 0 than to 80. There are hypotheses, but I don't think those hypotheses have, as of yet, been sufficiently validated that we believe them to be true, and there is an increasing number of people who believe that the traditional hypotheses might not really explain what's going on. I would also say that Alzheimers and schizophrenia and even type 2 diabetes are not really one disease. They're almost certainly a heterogeneous collection of mechanisms that manifest in clinically similar ways. So in the same way that we now understand that breast cancer is really not one disease, it is multitude of, uh, cellular mechanisms, all of which ultimately translate to, uh, uncontrolled proliferation, uh, but it's not one disease. The same is almost undoubtedly true for those other diseases as well, and it's that understanding that needs to precede any understanding of the specific mechanisms of any of those other diseases. Now, in schizophrenia, I would say we're almost certainly closer to 0 than to anything else. Um, type 2 diabetes is a bit, um, of a mix. There are clear mechanisms that are implicated that I think have been validated that have to do with insulin resistance and such, but there's, um, almost certainly there as well many mechanisms that we have not yet understood.
- 6:31 – 10:16
Longevity
- DKDaphne Koller
- LFLex Fridman
You've also thought and worked a little bit on the longevity side. Do you see the disease and longevity as-... overlapping completely, partially, or not at all, a- as efforts?
- DKDaphne Koller
Those mechanisms are certainly overlapping. There's, um, a well-known phenomenon that says that, uh, for most diseases other than childhood diseases, the, um, risk, um, for getting, for contracting that disease increases exponentially year on year every year from the time you're about 40. So obviously, there is a connection between those two things. Um, I, that's not to say that they're identical. There's clearly aging that happens that is not really associated with any specific disease. Um, and there's also diseases and mechanisms of disease that are not specifically related to aging. So, I think overlap is where we're at.
- LFLex Fridman
(laughs) Okay. It, it is a little unfortunate that we get older and it seems that there's some correlation with the, the fa- the, the occurrence of diseases with the fact that we get older.
- DKDaphne Koller
Mm-hmm.
- LFLex Fridman
And both are quite sad.
- DKDaphne Koller
Well, I mean, there's processes that happen as cells age that I think are contributing to disease. Some of those have to do with, uh, DNA damage that accumulates as cells divide, where the repair mechanisms don't fully, uh, correct for those. Um, there are accumulations of, uh, proteins that are misfolded and potentially aggregate and those too contribute to disease and contribute to inflammation. There is a nu- there is a multitude of mechanisms that have been uncovered that are sort of wear and tear at the cellular level that contribute to disease processes s- that, and I'm sure there's many that we don't yet understand.
- LFLex Fridman
On a small tangent, and perhaps philosophical (laughs) . Does, um, the, the fact that things get older and the fact that things die is a very powerful feature for the growth of new things that, you know, it's a learn- it's a kind of learning mechanism so it's both tragic and beautiful. So (laughs) do you, do you th- do you th- do you, uh, so in, you know, in, in trying to fight disease and trying to fight aging, do you think about sort of the useful fact of our mortality? Or would you, like w- if you were, could be immortal, would you choose to be immortal?
- DKDaphne Koller
Again, I think immortal is a very long time.
- LFLex Fridman
(laughs)
- DKDaphne Koller
And (laughs) I don't know that that would necessarily be something that I would want to aspire to. But I think all of us aspire to, um, an increased health span, I would say, which is an increased amount of time where you're healthy and active and feel as you did when you were 20, and we're nowhere close to that. Uh, people deteriorate physically and mentally over time and that is a very sad phenomenon. So, I think a wonderful aspiration would be if we could all live to, you know, the biblical 120 maybe-
- LFLex Fridman
(laughs)
- DKDaphne Koller
... in perfect health would be-
- LFLex Fridman
And high quality of life.
- DKDaphne Koller
... high quality of life. I think that would be an amazing goal for us to achieve as a society. Now, is the right age 120 or 100 or 150? I think that's up for debate. But I think an increased health span is a really worthy goal.
- LFLex Fridman
And anyway, in a grand time of the age of the universe, it's all pretty short.
- 10:16 – 13:05
Role of machine learning in treating diseases
- LFLex Fridman
S- so from the perspective, uh, you've done obviously a lot of incredible work in machine learning. So, what role do you think data and machine learning play in this, in this goal of trying to understand diseases and trying to eradicate diseases?
- DKDaphne Koller
Up until now, I don't think it's played very much of a significant role because largely the datasets that one really needed t- to enable a powerful machine learning methods, h- those datasets haven't really existed. There's been dribs and drabs and some interesting machine learning that has been applied, I would say machine learning/data science. But the last few years are starting to change that, so we now see an increase in, uh, some large datasets, but equally importantly, an increase in technologies that are able to produce data at scale. It's not typically the case that people have deliberately, proactively used those tools for the purpose of generating data for machine learning, they, uh, to the extent that those techniques have been used for data production, they've been used for data production to drive scientific discovery and the machine learning came as a sort of byproduct second stage of, "Oh, you know, now we have a dataset, let's do machine learning on that rather than a more simplistic data analysis method." But what we are doing at Insitro is actually flipping that around and saying, "Here's this incredible repertoire of methods that bioengineers, cell biologists have come up with. Let's see if we can put them together in brand new ways with the goal of creating datasets that machine learning can really be applied on productively to create powerful predictive models that can help us address fundamental problems in human health."
- LFLex Fridman
So really focus, uh, so get, uh, make data the, the primary focus and the primary goal. And find, use the mechanisms of biology and chemistry to, uh, to, uh, to c- create the kinds of dataset that could, uh, allow machine learning to benefit the most.
- DKDaphne Koller
I wouldn't put it in those terms because that says that data is the end goal. Data is the means.
- LFLex Fridman
The means. Okay.
- DKDaphne Koller
So for us, the end goal is helping address challenges in human health and the method that we've elected to do that is...... uh, to apply machine learning to build predictive models. And machine learning, in my opinion, can only be really successfully applied, especially the more powerful models, if you give it data that is of sufficient scale and sufficient quality. So how do you create those data sets so as to drive the, uh, ability to generate predictive models which subsequently help improve
- 13:05 – 16:25
A personal journey to medicine
- DKDaphne Koller
human health?
- LFLex Fridman
So before we dive into details of that, le- let me take a step back and ask, um, when and where (laughs) was your interest in human health born? Are there moments, events, perhaps, if I may ask, tragedies in your own life that catalyzed this passion or was it the broader desire to help humankind?
- DKDaphne Koller
So I would say it's a bit of both. So on, uh, I mean, my interest in human health actually dates back to the early 2000s when, uh, a l- when a lot of my, um, peers in machine learning and I were using data sets that frankly were not very inspiring. Some of us old-timers still remember the "20 news groups" data set where this was le- literally a bunch of text from 20 news groups, a concept that doesn't really even exist anymore. And the question was, can you classify, uh, which, um, which news group a particular bag of words came from? And it wasn't very interesting. The data sets at the time on the biology side were much more interesting both from a technical and also from an aspirational perspective. They were still pretty small but they were better than 20 news groups and, um, so I started out, I think, just by, just by wanting to do something that was more, I don't know, societally useful and technically interesting. And then over time became more and more interested in the biology and the, and the human health aspects for themselves and began to work even sometimes on papers that were just in biology without having a significant machine learning component. Um, I think my interest in drug discovery is partly due to, um, an incident I had with, um, when my father sadly passed away about 12 years ago. Um, he had an autoimmune disease that settled in his lungs, um, and, um, the doctors basically said, "Well, there's only one thing that we could do, which is give him prednisone." At some point, I remember a doctor even came and said, "Hey, let's do a lung biopsy to figure out which autoimmune disease he has." And I said, "W- would that be helpful? Would that change treatment?" He said, "No. There's only prednisone. That's the only thing we can give him." And I had friends who were rheumatologists who said, "The FDA would never approve prednisone today because the, the ratio of, of side effects to benefit is probably not large enough." Today, we're in a state where there's probably four or five, maybe even more, um, uh, well it depends for which autoimmune disease, but there are multiple drugs that can help people with autoimmune diseases, many of which didn't exist at, 12 years ago. And I think we're at a golden time in some ways in drug discovery where there's the ability to create drugs, um, that are much more safe and much more effective than we've ever been able to do before. And what's lacking is enough understanding of biology and mechanism to know where to aim that, um, where to aim that engine, and I think that's where machine learning can help.
- 16:25 – 33:25
Insitro and disease-in-a-dish models
- LFLex Fridman
So in 2018 you started and now lead a company, insitro, uh, which is a, like you mentioned, perhaps the focus is drug discovery and the utilization of machine learning for drug discovery. So you mentioned that, "We're really interested in creating what you might call a disease-in-a-dish model, disease-in-a-dish models, places where diseases are complex, where we really haven't had a good model system, where typical animal models that have been used for years, including testing on mice, just aren't very effective." So can you, can you try to describe what is an animal model and what, what is a disease-in-a-dish model?
- DKDaphne Koller
Sure. So an animal model for disease is where you create effectively... It's what it sounds like. It's a, it's a, oftentimes a mouse, uh, where we have introduced some external perturbation that creates the disease and then we cure that, um, disease, uh, and the hope is that by doing that, we will cure a similar disease in the human. The problem is of- is that oftentimes the way in which, um, we generate the disease in the animal has nothing to do with how that disease actually comes about in a human. Uh, it's what you might think of as a copy of the, of the phenotype, a copy of the clinical outcome, uh, but the mechanisms are quite different. And so curing the disease in the animal, which in most cases doesn't happen naturally, mice don't get Alzheimer's, they don't get diabetes, they don't get atherosclerosis, they don't get autism or schizophrenia, um, those, uh, cures don't translate over to what happens in the human. And that's where most drugs fails just because the findings that we had in the mouse don't translate to a human. Um, the disease-in-a-dish models is a fairly new approach. It's been enabled by technologies that have not existed for more than five to 10 years. So for instance, um, the ability for us to take a cell from any one of us, you or me, um, revert that, say, skin cell to what's called stem cell status, which is a-... what, which is what's called a pluripotent cell that can then be differentiated into different types of cells. So from that pluripotent cell, one can create a less neuron or a less cardiomyocyte or a less hepatocyte that has your genetics but that right, uh, cell type. And so if there's a genetic burden of disease that would manifest in that particular cell type, you might be able to see it by looking at those cells and saying, "Oh, that's what potentially sick cells look like versus healthy cells." And understand how, uh, and then explore what kind of interventions might revert the unhealthy-looking cell to a healthy cell. Now, of course, curing cells is not the same as curing people. Um, and so there's still potentially a translatability gap. But at least for diseases that, um, are driven, say, by human genetics and where the human genetics is what drives the cellular phenotype, there is s- some reason to hope that if we revert those cells in which the disease begins and where the disease is driven by genetics and we can revert that cell back to a healthy state, maybe that will help, um, also revert the gl- more global clinical phenotypes. That's really what we're hoping to do.
- LFLex Fridman
That step, that backward step, I was reading about it, the Yamanaka factor-
- DKDaphne Koller
Yes.
- LFLex Fridman
... the, so like that, the reverse step back to stem cells-
- DKDaphne Koller
Yes.
- LFLex Fridman
... is in- seems like magic. Is it-
- DKDaphne Koller
It is.
- LFLex Fridman
(laughs)
- DKDaphne Koller
Unbel- honestly, before that happened, I think very few people would have predicted that to be possible. It's amazing.
- LFLex Fridman
Uh, can you maybe elaborate? Is it actually possible? (laughs) Like, uh, where, like, uh, how sta- so this result was, uh, maybe, like, I do- I don't know how many years ago, maybe 10 years ago-
- DKDaphne Koller
Mm-hmm.
- LFLex Fridman
... it was first demonstrated, something like that. Is this, how hard is this? Like, how noisy is this backward step? It seems quite incredible and cool.
- DKDaphne Koller
It is, uh, it is incredible and cool. It was, uh, much more, I think, finicky and bespoke at the early stages when the discovery was first made. But at this point, it's become almost industrialized. There are c- what's called contract research organizations, vendors that will take a sample from a human and revert it back to stem cell status and it works a very good fraction of the time. Now, there are people who will ask, um, I think, g- good questions, "Is this really truly a stem cell or does it remember certain aspects of, uh, what, of changes-"
- LFLex Fridman
(laughs)
- DKDaphne Koller
"... that were made in the human, um, beyond the genetics?"
- LFLex Fridman
It's fast as a skin cell, yeah.
- DKDaphne Koller
It's fast as a skin cell or it's fast in terms of exposures to different environmental factors and so on. So I think the consensus right now is that these are not always perfect and there is little bits and pieces of memory sometimes but by and large, these are actually pretty good.
- LFLex Fridman
So, uh, one of the key things, well, maybe, maybe you can correct me but one of the useful things for machine learning is size, scale of data.
- DKDaphne Koller
Mm-hmm.
- LFLex Fridman
How easy it is to do these kinds of reversals to stem cells and then disease in a dish models its scale. Is this, is that a huge challenge or, or not?
- DKDaphne Koller
So the reverse, the reversal is not as of this point something that can be done at the scale of tens of thousands or hundreds of thousands. Um, I think total number of stem cells or iPS cells that are what's called induced pluripotent stem cells in the world, I think is somewhere between 5,000 and 10,000 last I looked. Now again, that might not count things that exist in this or that academic center and they may add up to a bit more but that's about the range. So it's not something that you could at this point generate iPS cells from a million people but maybe you don't need to because maybe that background is enough because it can also be now perturbed in different ways and some people have done really interesting experiments in, um, for instance taking cells from a healthy human and then, uh, introducing a mutation into it using some of the, using one of the other miracle technologies that's emerged in the last decade which is CRISPR gene editing and introduce the mutation that is known to be pathogenic and so you can now look at the healthy cells and the unhealthy cells, the one with the mutation and do a one-on-one comparison where everything else is held constant and so you could really start to understand specifically what the mutation does at the cellular level. So the iPS cells are a great starting point and obviously more diversity is better 'cause you also want to capture ethnic background and how that affects things but maybe you don't need one from every single, uh, patient with every single type of disease because we have other tools at our disposal.
- LFLex Fridman
Well, how much difference is there between people, I mentioned ethnic background, in terms of iPS cells? So we're all, like, it seems like these magical cells that can do any- uh, create anything, uh, between different populations, different people, is there a lot of variability between stem cells?
- DKDaphne Koller
Well, first of all, there is the variability that's driven simply by the fact that, um, genetically, we're different. So a stem cell that's derived from my genotype is gonna be different from a stealth cell, stem cell that's derived from your genotype. Um, there's also some differences that have more to do with, for whatever reason, um, some people's stem cells differentiate better than other people's stem cells. We don't entirely understand why so there is certainly some differences there as well, but the fundamental difference and the one that we really care about and is a positive is that the, um, is the fact that the genetics are- are different and therefore recapitulate my disease burden versus your disease burden.
- LFLex Fridman
What's a disease burden?
- DKDaphne Koller
Well, a disease burden is just, if you thi-, I mean, it's not a well-defined mathematical term, although there are mathematical formulations of it. It, if you think about the fact that some of us are more likely to get-
- LFLex Fridman
That's right.
- DKDaphne Koller
... a certain disease than others because we have more variations in our genome that are causative of the disease, maybe fewer that are protective of the disease. Um, people have, uh, quantified that using what are called polygenic risk scores, which look at all of the variations in a- an individual person's genome and add them all up in terms of how much risk they confer for a particular disease and then they've put people on a spectrum of their disease risk. And for certain diseases where we've been sufficiently powered to really understand the connection between the many, many small variations that give rise to an increased disea- disease risk, there is some pretty significant differences in terms of the risk between the people, say, at the highest decile of this polygenic risk score and the people of the lowest decile. Sometimes those, um, differences are, you know, a factor of 10 or 12 higher, so there's definitely, uh, a l- a lot that our genetics contributes to disease risk even if it's not by any stretch the full explanation.
- LFLex Fridman
And from a machinery perspective, there's signal there?
- DKDaphne Koller
There is definitely signal in the genetics and there's even more signal we believe in looking at the cells that are derived from those different genetics because, uh, in principle you could say all the signals, they're at the, at the genetics level so we don't need to look at the cells-
- 33:25 – 36:43
What diseases can be helped with disease-in-a-dish approaches?
- LFLex Fridman
with these m- w- with these approaches, what do you hope, what kind of diseases can be helped? We mentioned Alzheimer's, sch- schizophrenia, type 2 diabetes. Can you just describe various kinds of diseases that this approach can, can help?
- DKDaphne Koller
Well, we don't know and I try and be very cautious about making promises about some things. Say, "Oh, we will cure X." People make that promise and I think it's, I try to first deliver and then promise as opposed to the other way around. There are characteristics of a disease that make it more likely that this type of approach can potentially be helpful. So for instance, diseases that have a very strong genetic basis, um, are ones that are more likely to manifest in a stem cell derived model. Um, we would want the cellular models to be relatively reproducible and robust so that you could actually get a, um, enough of those cells in the, in a way that isn't very highly variable and noisy. Um, you would want the disease to be relatively contained in one or a small number of cell types that you could actually create in an in vitro, in a dish setting. Whereas if it's something that's really broad and systemic and involves multiple cells that are in very distal parts of your body, putting that all in a dish is really challenging. So we want to focus on the ones that are most likely to be successful today with the hope, I think, that, uh, really smart bioengineers out there are developing better and better systems all the time so the diseases that might not be tractable today might be tractable in three years. So for instance, five years ago, the stem cell derived models didn't really exist. People were doing most of the work in cancer cells and cancer cells are very, very, uh, poor models of most human biology because they're, A, they were cancer to begin with, and B, as you passage them and they, uh, proliferate in a dish, they become, because of the genomic instability, even less similar to human biology. Um, now we have these stem cell derived models. Uh, we have the capability to reasonably robustly, not quite at the right scale yet but close, to derive what's called organoids which are these teeny little sort of multicellular-
- LFLex Fridman
Yeah.
- DKDaphne Koller
... uh, organ, rep- rep, sort of models of an organ system. So there are cerebral organoids and liver organoids and kidney organoids and, um-
- LFLex Fridman
Yeah, brain organoids-
- DKDaphne Koller
... gut organoids.
- LFLex Fridman
... is possibly the coolest thing I've ever seen. (laughs)
- DKDaphne Koller
Is that not like the coolest thing?
- LFLex Fridman
Yeah.
- DKDaphne Koller
Um, and then I think on the horizon we're starting to see things like connecting these organoids to each other so that you could actually start... And there's some really cool papers that start to do that where you can actually start to say, "Okay, can we do multi-organ system stuff?" Um, there's many challenges to that. It's not easy by any stretch, but it might, I'm sure people will figure it out and in three years or five years there will be disease models that we could make for things that we can't make today.
- LFLex Fridman
Yeah, and this conversation would seem almost outdated with the kind of scale that could be achieved in like three years.
- DKDaphne Koller
I hope so.
- LFLex Fridman
That's the hope.
- DKDaphne Koller
That would be so cool.
- 36:43 – 49:04
Coursera and education
- DKDaphne Koller
- LFLex Fridman
So you've, uh, co-founded Coursera with Andrew Ng.... and we're part of the whole MOOC revolution. (laughs)
- DKDaphne Koller
Mm-hmm.
- LFLex Fridman
So to jump topics a little bit, can you maybe tell the origin story of the history, the origin story of MOOCs, of Coursera, and in general, the- your teaching to huge audiences on a very sort of impactful topic of AI-
- DKDaphne Koller
Yeah.
- LFLex Fridman
... in general?
- DKDaphne Koller
So I think the origin story of MOOCs emanates from a number of efforts that occurred at Stanford University around, um, you know, th- the late 2000s where different individuals within Stanford, myself included, were getting really excited about the opportunities of using online technologies as a way of achieving both improved quality of teaching and also improved scale. And so, um, Andrew, for instance, led the, um, uh, the Stanford Engineering Everywhere, which was sort of an attempt to take 10 Stanford courses and put them online, um, just as, you know, video lectures. I led an effort within Stanford to take some of the courses and really create a very different teaching model that broke those up into smaller units and had some of those embedded interactions and- and so on, which got a lot of support from, um, university leaders because they felt like it was potentially a way of improving the quality of instruction at Stanford by moving to what's now called the flipped classroom model. Um, and so those efforts eventually sort of started to interplay with each other and created a tremendous sense of excitement and energy within the Stanford community about the potential of online teaching and led in the fall of 2011 to the launch of the first Stanford MOOCs, the, um-
- LFLex Fridman
By the way, MOOCs, I- it's probably impossible that people don't know, but it's, I guess, massive...
- DKDaphne Koller
Open Online Courses.
- LFLex Fridman
... Open Online Courses. So the, uh-
- DKDaphne Koller
We did not come up with the acronym.
- LFLex Fridman
Yeah.
- DKDaphne Koller
I'm not particularly fond of the acronym but it is what it is.
- LFLex Fridman
It is what it is. Big Bang is not a great term for the start of the universe but it is what it is.
- DKDaphne Koller
Probably so.
- LFLex Fridman
(laughs)
- DKDaphne Koller
Um, so anyway, we ... so those courses launched in- in the fall of 2011 and there were, within a matter of weeks, with no real publicity campaign, just, um, a New York Times article that went viral, um, about 100,000 students or more in each of those courses. And I remember this conversation that Andrew and I had which was like, "Wow, this is just, there's this real need here." And I think we both felt like, sure, we were, um, accomplished academics and we could go back and, you know, go back to our labs, write more papers, but if we did that, then this wouldn't happen, and it seemed too important not to happen. And so we spent a fair bit of time debating, do we wanna do this as a Stanford effort, kind of building on what we'd started? Do we wanna do this as a for-profit company? Do we wanna do this as a nonprofit? And we decided ultimately to do it as we did with Coursera, um, and so, you know, we, uh, started really operating as a company at the beginning of 2012. Um, we-
- LFLex Fridman
And the rest is history. (laughs)
- DKDaphne Koller
And the rest is history.
- LFLex Fridman
But how did you ... was that really surprising to you? How- how do you at that, how did you at that time and at this time make sense of this need for sort of global education? You mentioned that you felt that, wow, the- the popularity indicates that, um, there's a hunger for-
- DKDaphne Koller
Mm-hmm.
- LFLex Fridman
... sort of globalization of l- learning.
- DKDaphne Koller
I think there is a hunger for learning that, you know, globalization is part of it, but I think it's just a hunger for learning. The world has changed in the last 50 years. It used to be that you finished college, you got a job, by and large the skills that you learned in college were pretty much what got you through the rest of your job history and- and yeah, you learned some stuff, but it wasn't a dramatic change. Today, we're in a world where the skills that you need for a lot of jobs, they didn't even exist when you went to college, and the jobs, and many of the jobs that existed when you went to college don't even exist today, um, or are dying. So part of that is due to AI, but not only, and we need to find a way of keeping people, giving people access to the skills that they need today, and I think that's really what's driving a lot of this hunger.
- LFLex Fridman
So I think if we even take a step back, all of, uh ... for you, all of this started in trying to think of new ways to teach or to, yeah, new ways to sort of, uh, organize the material and present the material in a way that would help the education process-
- DKDaphne Koller
Right.
- LFLex Fridman
... pedagogy. Yeah. So what have you learned about effective education from this process of playing, of experimenting with different, uh, ideas?
- DKDaphne Koller
So we learned a number of things, some of which I think could translate back and have translated back effectively to how people teach on campus and some of which I think are more specific to people who learn online, um, more sort of people who learn as part of their daily life. So we learned, for instance, very quickly that short is better. So people who are especially in the workforce can't do a 15-week semester-long course. They just can't fit that into their lives.
- LFLex Fridman
Well, short- can you, uh, can you describe the shortness of what? The- the- the entirety?
- DKDaphne Koller
Both. Both.
- LFLex Fridman
So ev- every aspect, so the little lecture is short-
- 49:04 – 50:52
Advice to people interested in AI
- DKDaphne Koller
- LFLex Fridman
And briefly, I know it might be a- a difficult question to ask but there's a lot of people fascinated by artificial intelligence, by machine learning, by deep learning. Is there a recommendation for the next year or for a lifelong journey? As somebody interested in this, how do they, um, how do they begin? How do they enter that learning journey?
- DKDaphne Koller
I think the important thing is first to just get started and, um, there is plenty of online content that one can get for both the core foundations of mathematics and statistics and programming and then from there to machine learning. I would encourage people not to skip too quickly past the foundations because I find that there is a lot of people who learn machine learning, whether it's online or on campus, without getting those foundations and they basically just turn the crank on existing models in ways that, A, don't allow for a lot of innovation and, uh, and- and adjustment to the problem at hand but also, B, are sometimes just wrong and they don't even realize that their application is- is wrong because there's artifacts that they haven't fully understood. So, I think the foundations, machine learning is an important step and then, um, and then actually start solving problems. Um, try and find someone to solve them with because especially at the beginning it's useful to have someone to bounce ideas off and fix mistakes that you make and, um, and you can fix mistakes that they make but, uh, but then just find practical problems, whether it's in your workplace or if you don't have that, Kaggle competitions or such are a really great place to find interesting problems and just practice. (laughs)
- LFLex Fridman
Practice.
- 50:52 – 55:10
Beautiful idea in deep learning
- LFLex Fridman
Perhaps a bit of a romanticized question but what idea in deep learning do you find, have you found in your journey the most beautiful or surprising or interesting? Uh, perhaps not just deep learning but AI in general, statistics...
- DKDaphne Koller
I'm gonna answer with two things.
- LFLex Fridman
Sure.
- DKDaphne Koller
Um, one would be the foundational concept of end-to-end training which is that you start from the raw data and you train something that is not like a single piece but rather the, um, t- towards the actual goal that you're looking to-
- LFLex Fridman
So, from the raw data to the outcome.
- DKDaphne Koller
Yeah.
- LFLex Fridman
Like, and noth- no, no details in between.
- DKDaphne Koller
Well, not no details but the fact that you, I mean, you could certainly introduce building blocks that were trained towards other tasks, I'm actually coming to that in my second half of the answer.
- LFLex Fridman
(laughs)
- DKDaphne Koller
But, um, that doesn't have to be like a single monolithic blob in the middle. Actually, I think that's not ideal. But rather, the fact that at the end of the day, you can actually train something that goes all the way from the beginning to the end. And the other one that I find really compelling is the notion of learning a representation that in its turn, even if it was trained to another task, can potentially be used as a much more rapid starting point to solving a different task, and that's, I think, reminiscent of what makes people successful learners. It's something that is relatively new in the machine learning space. I think it's underutilized even relative to today's capabilities but more and more of how do we learn, um, sort of reusable representation?
- LFLex Fridman
So, end-to-end and transfer learning...
- DKDaphne Koller
Yeah.
- LFLex Fridman
Is it surprising to you that neural networks are able to, in many cases, do these things? Is this, is it, uh, maybe taking back to when you f- when you first would dive deep into neural networks or in general even today, is it surprising that neural networks work at all and work wonderfully to do this kind of raw, uh, end-to-end learning and even transfer learning?
- DKDaphne Koller
I think I was surprised by how well, when you have large enough amounts of data, it's possible to find a meaningful representation in what is an exceedingly high dimensional space. And so I find that to be really exciting and people are still working out the math for that. There's more papers on that every year and I think it's, would be really cool if we figured that out. But, um, that to me was a surprise because in the early days when I was starting my way in machine learning and the datasets were rather small, I think we- we believed, I believed that you needed to have a much more constrained and knowledge-rich search space to really make, t- to really get to a meaningful answer and I think it was true at the time. What I think is, um, is still a question is will a completely knowledge-free approach where there's no prior knowledge going into the construction of the model, is that gonna be the solution or not? It's not actually the solution today in the sense that the architecture of a, you know, convolutional neural network that's used for images is actually quite different to the type of network that's used for language and yet different from the one that's used for speech or biology or any other application. There's still some insight that goes into the structure of the network to get the- the right performance. Will we be able to come up with a universal learning machine? I don't know.
- LFLex Fridman
Yeah. (laughs) ... I wonder if there's always has to be some insight injected somewhere-
- DKDaphne Koller
Yeah. Uh-huh.
- LFLex Fridman
... or whether it can converge.
- 55:10 – 58:29
Uncertainty in AI
- LFLex Fridman
So you've done a lot of interesting work with probabilistic graphical models i- in general, Bayesian deep learning and, and so on. So can you maybe speak high level, how can learning systems deal with uncertainty?
- DKDaphne Koller
One of the limitations, I think, of a lot of machine learning models is that they come up with an answer and you don't know how much you can believe that answer and, um, oftentimes the, uh, e- th- the answer is actually quite poorly calibrated relative to its uncertainties. Even if you look at where the, um, you know, the, the, the confidence that comes out of the, say, the neural network at the end and you ask, "How much more likely is an answer of 0.8 versus 0.9?" it's not really in any way calibrated to the, um, to the actual reliability of that network and how true it is. And the further away you move from the training data, the more, uh, not only the more wrong the network is, often it's more wrong and more confident in its wrong answer.
- LFLex Fridman
(laughs)
- DKDaphne Koller
And that is a serious issue in, um, a lot of application areas. So when you think, for instance, about medical diagnosis as being maybe a- an epitome of how problematic this can be, if you were training your network on a certain set of patients, on a certain patient population and now you have a patient that is an outlier and there's no human that looks at this and that patient is put into a neural network and your network not only gives a completely incorrect diagnosis but is supremely confident-
- LFLex Fridman
Yeah.
- DKDaphne Koller
... in its wrong answer, you could kill people.
- LFLex Fridman
Yes.
- DKDaphne Koller
So I think creating more of a, um, an understanding of how do you produce networks that are calibrated in their uncertainty and can also say, "You know what? I give up. I don't know what to say about this particular data instance because I've never seen something that's sufficiently like it before," I think that's going to be really important in mission-critical applications, especially ones where human life is at stake and that includes e- you know, medical applications.
- LFLex Fridman
Yeah.
- DKDaphne Koller
But it also includes, you know, automated driving because you'd want the network to be able to say, "You know what? I have no idea what this blob is that I'm seeing in the middle of the road, so I'm just gonna stop because I don't wanna potentially run over, uh, a pedestrian that I don't recognize."
- LFLex Fridman
Is there good mechanisms, ideas of how to allow learning systems to, uh, provide that uncertainty with their pre- along with their predictions?
- DKDaphne Koller
Certainly people have, uh, come up with mechanisms that involve Bayesian deep learning, deep learning that involves Gaussian processes. Um, I mean, there is a slew of different approaches that people have come up with. There's methods that use ensembles of networks with, uh, trained with different subsets of data or different random starting points. Those are actually sometimes surprisingly good at creating a sort of set of how confident or not you are in your answer.
- LFLex Fridman
Yeah.
- DKDaphne Koller
It's very much an area of open research.
- 58:29 – 1:06:52
AGI and AI safety
- DKDaphne Koller
- LFLex Fridman
Let's cautiously (laughs) venture back into the land of, uh, philosophy and, uh, speaking of AI systems providing uncertainty, somebody like, uh, Stuart Russell believes that as we create more and more intelligent systems, it's really important for them to be full of self-doubt, uh, (laughs) because, uh, you know, if they're given more and more power, we wanna m-... The way to maintain human control over AI systems or human supervision, which is true, like you just mentioned with autonomous vehicles, it's really important to get human supervision when the car is not sure because if it's really confident, it can, in cases when it can get in trouble, it's gonna be really problematic. So let me ask about sort of the questions of AGI and human level intelligence. I mean, we've talked about curing diseases. Uh, we just said a fundamental thing that can have an impact today. But AI people also dream of both understanding and creating intelligence. Is that something you think about? Is that something you dream about? Is that something, uh, you think is within our reach to be thinking about as computer scientists?
- DKDaphne Koller
Boy, let me tease apart different parts of that question.
- LFLex Fridman
Yeah, that's the worst question. (laughs)
- DKDaphne Koller
Uh, yeah. It, it's a multi-part question.
- LFLex Fridman
Yeah.
- DKDaphne Koller
So let me start with the feasibility of AGI, then I'll talk about the timelines a little bit and then talk about, well, what controls does one need when protecting... Wh- when thinking about protections in the AI space. So, you know, I think AGI obviously is a longstanding dream that even our, uh, early pioneers in this space had, you know, the Turing test and so on are, uh, the earliest discussions of that. We're obviously closer than we were w- 70 or so years ago, but I think it's still very far away. I think machine learning algorithms today are really exquisitely good pattern recognizers in very specific problem domains where they have seen enough training data to make good predictions. You take a, um, machine learning algorithm and you move it to a slightly different version of even that same problem, far less one that's different, and it will just completely choke.... so, I think we're nowhere close to the versatility and flexibility of even a human toddler in terms of their ability to context switch and solve different problems using a single knowledge base, single brain. So am I desperately worried about-
- LFLex Fridman
(laughs)
- DKDaphne Koller
... uh, the machines taking over the universe and, you know, starting to kill people because they want to have more power? I don't think so.
- LFLex Fridman
Well, so to pause on that, so you, you kind of, uh, intuited that super intelligence is a very difficult thing to achieve. That we're-
- DKDaphne Koller
Even intelligent-
- LFLex Fridman
Intelligent. Intelligent.
- DKDaphne Koller
Super intelligence, we're not even close to intelligence.
- LFLex Fridman
E- even just the greater abilities of generalization of our c- current systems. But (laughs) we haven't answered all the parts and we'll take another-
- DKDaphne Koller
I'm getting to the second part of what-
- LFLex Fridman
Okay. We'll go, we'll take it ... but, uh, maybe another tangent you could also pick up is, uh, can we get in trouble with much dumber systems?
- DKDaphne Koller
Yes.
- LFLex Fridman
(laughs)
- DKDaphne Koller
And that is exactly where I was going.
- LFLex Fridman
Okay.
- DKDaphne Koller
So, I ... so just to wrap up on the threats of AGI, I think that it seems to me a little early today to figure out protections against a human level or superhuman level intelligence who's ... who ... where we don't even see the skeleton of what that would look like. So, it seems that it's very speculative on how w- how, how to protect against that. But we can definitely, and have gotten into trouble on much dumber systems. And a lot of that has to do with the fact that the systems that we're building are increasingly complex, um, increasingly poorly understood, um, and there's ripple effects that are unpredictable in changing little things that can have dramatic consequences on the outcome. Uh, and by the way, that's not unique to artificial intelligence. I think artificial intelligence exacerbates that, brings it to a new level. But heck, our electric grid is really complicated.
- LFLex Fridman
Yes.
- DKDaphne Koller
The software that runs our financial markets-
- LFLex Fridman
Right.
- DKDaphne Koller
... is really complicated. And we've seen those ripple effects translate to dramatic negative consequences like, for instance, um, financial crashes that have to do with feedback loops that we didn't anticipate. So, I think that's a issue that we need to be thoughtful about in many places. Um, artificial intelligence being one of them. And we should ... and I think it's really important the people are thinking about ways in which we can have better interpretability of systems, better tests for, for instance, measuring the extent to which a machine learning system that was trained in one set of circumstances, how well does it actually work in a very different set of circumstances where you might, say for instance, well, I'm not gonna be able to test my automated vehicle in every possible city, village, um, w- weather condition and so on. But if you trained it on this set of conditions and then tested it on 50 or 100 others that were quite different from the ones that you trained it on, then I ... and it worked, then that gives you confidence that the next 50 that you didn't test it on might also work. So-
- LFLex Fridman
Okay.
- DKDaphne Koller
... effectively testing for generalized ability. So I think there's ways that we should be constantly thinking about to validate the robustness-
- LFLex Fridman
Mm-hmm.
- DKDaphne Koller
... of our systems. I think it's very different from the let's make sure robots don't take over the world. And then the other (laughs) place where I think we have a threat, which is also important for us to think about, is the extent to which technology can be abused. So, like any really powerful technology, machine learning can be very much used badly as well as to good. And that goes back to many other technologies that have come up with when people invented projectile missiles and it turned into guns. And people invented nuclear power and it turned into nuclear bombs. And I think honestly I would say that to me gene editing and CRISPR is at least as dangerous a technology if used badly than machine ... as machine learning. Um, you could create really nasty viruses and such using gene editing that are, um, you know, you, you would be really-
- LFLex Fridman
(laughs)
- 1:06:52 – 1:09:04
Are most people good?
- LFLex Fridman
maybe another way to ask is, do you think most people are good and fundamentally we tend towards a better world, which is underlying the question, will machine learning with gene editing ultimately land us somewhere good? Are you optimistic? (laughs)
- DKDaphne Koller
I think by and large I'm optimistic. I think that, mm, most people mean well. That doesn't mean that most people are, you know, altruistic do-gooders, but I think most people mean well. But I think it's also really important for us as a society to create social norms where doing good and being perceived well by our peers is... A- are positively correlated.
- LFLex Fridman
Hmm.
- DKDaphne Koller
I mean, it's very easy to create dysfunctional societies. There's certainly multiple psychological experiments as well as, sadly, real world e- events where people have devolved to a world where being perceived well by your peers is correlated with really atrocious, um, often genocidal behaviors. Uh, so we really want to make sure that we maintain a set of social norms where people know that to be a successful member of society you want to be doing good. And one of the things that I sometimes worry about is that some societies (laughs) don't seem to necessarily be moving in the forward direction in that regard, where it's not necessarily the case that doing, um... That being a good person is what makes you be perceived well by your peers. And I think that's a really important thing for us as a society to remember. It's very easy degenerate back into a universe where it's okay to do really bad stuff and still have your peers think you're amazing.
- 1:09:04 – 1:11:57
Meaning of life
- LFLex Fridman
It's fun to ask a world-class computer scientist and engineer a ridiculously philosophical question like, "What is the meaning of life?" Let me ask, "What gives your life meaning? What are... What is the source of fulfillment, happiness, joy, purpose?"
- DKDaphne Koller
When we were starting Coursera in the fall of 2011, that was right around the time that, uh, Steve Jobs passed away, and so the media was full of various famous quotes that he'd s- uttered. And one of them that really stuck with me, because it resonated with stuff that I'd been feeling for even years before that, is that our goal in life should be to make a dent in the universe. So, I think that, to me, what gives my life meaning is that I would hope that when I am lying there on my death bed and looking at what I'd done in my life that I can point to ways in which I have left the world a better place than it was when I entered it. This is something I tell my kids all the time, because I also think that the burden of that is much greater for those of us who were born to privilege. And in some ways, I was. I mean, I wasn't born super wealthy or anything like that, but I grew up in an educated family with parents who loved me and took care of me, and I had a chance at a great education. And, and, so I... And I always had enough to eat, so I was in many ways born to privilege more than the vast majority of humanity. And my kids, I think, are even more so born to privilege than I was fortunate enough to be. And I think it's really important that for... Especially for those of us who have that opportunity, that we use our lives to make the world a better place.
- LFLex Fridman
I don't think there's a w- better way to end it, Daphne. It was a honor to talk to you. Thank you so much for talking today.
- DKDaphne Koller
Thank you.
- LFLex Fridman
Thanks for listening to this conversation with Daphne Koller, and thank you to our presenting sponsor, Cash App. Please consider supporting the podcast by downloading Cash App and using code LEXPODCAST. If you enjoyed this podcast, subscribe on YouTube, review it with five stars on Apple Podcasts, support it on Patreon, or simply connect with me on Twitter @LexFridman. And now, let me leave you with some words from Hippocrates, a physician from Ancient Greece who's considered to be the father of medicine. "Wherever the art of medicine is loved, there's also love of humanity." Thank you for listening, and hope to see you next time.
Episode duration: 1:12:03
Install uListen for AI-powered chat & search across the full episode — Get Full Transcript
Transcript of episode xlMTWfkQqbY
Get more out of YouTube videos.
High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.
Add to Chrome