Regina Barzilay: Deep Learning for Cancer Diagnosis and Treatment

00:00:00.000 | The following is a conversation with Regina Barsley.

00:00:03.240 | She's a professor at MIT and a world-class researcher

00:00:06.720 | in natural language processing

00:00:08.360 | and applications of deep learning to chemistry and oncology

00:00:12.480 | or the use of deep learning for early diagnosis,

00:00:15.360 | prevention and treatment of cancer.

00:00:18.320 | She has also been recognized for teaching

00:00:21.040 | of several successful AI related courses at MIT,

00:00:24.720 | including the popular introduction

00:00:26.840 | to machine learning course.

00:00:28.920 | This is the Artificial Intelligence Podcast.

00:00:32.160 | If you enjoy it, subscribe on YouTube,

00:00:34.560 | give it five stars on iTunes,

00:00:36.400 | support it on Patreon or simply connect with me on Twitter

00:00:39.840 | at Lex Friedman spelled F-R-I-D-M-A-N.

00:00:43.760 | And now here's my conversation with Regina Barsley.

00:00:47.760 | In an interview, you've mentioned

00:00:50.320 | that if there's one course you would take,

00:00:51.960 | it would be a literature course with a friend of yours

00:00:54.600 | that a friend of yours teaches.

00:00:56.360 | Just out of curiosity, 'cause I couldn't find anything on it,

00:01:00.240 | are there books or ideas that had profound impact

00:01:04.400 | on your life journey, books and ideas,

00:01:06.760 | perhaps outside of computer science and the technical fields?

00:01:10.760 | - I think because I'm spending a lot of my time at MIT

00:01:14.680 | and previously in other institutions where I was a student,

00:01:18.280 | I have a limited ability to interact with people.

00:01:21.040 | So a lot of what I know about the world

00:01:22.640 | actually comes from books.

00:01:25.240 | And there were quite a number of books

00:01:27.240 | that had profound impact on me and how I view the world.

00:01:31.360 | Let me just give you one example of such a book.

00:01:35.800 | I've maybe a year ago read a book

00:01:39.680 | called "The Emperor of All Melodies."

00:01:42.480 | It's a book about, it's kind of a history of science book

00:01:45.760 | on how the treatments and drugs for cancer were developed.

00:01:50.760 | And that book, despite the fact

00:01:53.600 | that I am in the business of science,

00:01:55.480 | really opened my eyes on how imprecise

00:01:59.800 | and imperfect the discovery process is

00:02:03.080 | and how imperfect our current solutions

00:02:05.840 | and what makes science succeed and be implemented.

00:02:11.080 | And sometimes it's actually not the strength of the idea,

00:02:14.120 | but devotion of the person who wants to see it implemented.

00:02:17.440 | So this is one of the books that,

00:02:19.800 | at least for the last year, quite changed the way

00:02:22.280 | I'm thinking about scientific process

00:02:24.920 | just from the historical perspective

00:02:26.680 | and what do I need to do

00:02:28.920 | to make my ideas really implemented.

00:02:33.440 | Let me give you an example of a book,

00:02:36.040 | which is not kind of, which is a fiction book,

00:02:39.560 | is a book called "Americana."

00:02:43.200 | And this is a book about a young female student

00:02:48.760 | who comes from Africa to study in the United States.

00:02:53.240 | And it describes her path within her studies

00:02:57.760 | and her life transformation that,

00:03:00.600 | in a new country and kind of adaptation to a new culture.

00:03:06.560 | And when I read this book,

00:03:09.520 | I saw myself in many different points of it,

00:03:14.920 | but it also kind of gave me the lens on different events.

00:03:19.920 | And some events that I never actually paid attention,

00:03:22.080 | one of the funny stories in this book

00:03:24.720 | is how she arrives to her new college

00:03:29.720 | and she starts speaking in English

00:03:32.920 | and she had this beautiful British accent

00:03:35.720 | because that's how she was educated in her country.

00:03:39.840 | This is not my case.

00:03:40.960 | And then she notices that the person who talks to her,

00:03:44.640 | talks to her in a very funny way, in a very slow way.

00:03:48.280 | And she's thinking that this woman is disabled

00:03:50.560 | and she's also trying to kind of to accommodate her.

00:03:54.440 | And then after a while, when she finishes her discussion

00:03:56.640 | with this officer from her college,

00:03:58.560 | she sees how she interacts with other students,

00:04:02.040 | with American students, and she discovers that actually

00:04:06.920 | she talked to her this way

00:04:09.640 | because she saw that she doesn't understand English.

00:04:12.160 | And I thought, wow, this is a funny experience.

00:04:15.040 | And literally within a few weeks,

00:04:17.960 | I went to LA to a conference

00:04:21.880 | and I asked somebody in the airport

00:04:24.320 | how to find a cab or something.

00:04:26.560 | And then I noticed that this person is talking

00:04:29.360 | in a very strange way.

00:04:30.280 | And my first thought was that this person

00:04:32.080 | have some pronunciation issues or something.

00:04:35.480 | And I'm trying to talk very slowly to him

00:04:37.160 | and I was with another professor, Ernst Frankel,

00:04:39.640 | and he's like laughing because it's funny

00:04:43.160 | that I don't get that the guy is talking in this way

00:04:45.720 | because he thinks that I cannot speak.

00:04:47.000 | So it was really kind of mirroring experience

00:04:50.080 | and it led me think a lot about my own experiences

00:04:54.320 | moving from different countries.

00:04:57.000 | So I think that books play a big role

00:05:00.240 | in my understanding of the world.

00:05:02.720 | - On the science question,

00:05:05.600 | you mentioned that it made you discover

00:05:07.520 | that personalities of human beings

00:05:09.840 | are more important than perhaps ideas.

00:05:12.480 | Is that what I heard?

00:05:13.720 | - It's not necessarily that they are more important

00:05:15.760 | than ideas, but I think that ideas on their own

00:05:19.200 | are not sufficient.

00:05:20.520 | And many times, at least at the local horizon,

00:05:24.720 | it's the personalities and their devotion to their ideas

00:05:29.200 | is really that locally changes the landscape.

00:05:33.040 | If you're looking at AI, like let's say 30 years ago,

00:05:37.680 | dark ages of AI or whatever word is symbolic times,

00:05:40.640 | you can use any word.

00:05:41.840 | There were some people,

00:05:44.640 | now we're looking at a lot of that work

00:05:46.600 | and we're kind of thinking this was not really

00:05:48.720 | maybe a relevant work,

00:05:50.640 | but you can see that some people managed to take it

00:05:53.040 | and to make it so shiny and dominate the academic world

00:05:58.040 | and make it to be the standard.

00:06:02.320 | If you look at the area of natural language processing,

00:06:05.160 | it is well-known fact that the reason that statistics

00:06:09.120 | in NLP took such a long time to become mainstream

00:06:13.960 | because there were quite a number of personalities

00:06:16.800 | which didn't believe in this idea

00:06:18.400 | and didn't stop research progress in this area.

00:06:22.040 | So I do not think that kind of asymptotically

00:06:27.040 | maybe personalities matters,

00:06:28.920 | but I think locally it does make quite a bit of impact

00:06:33.920 | and it's generally speeds up the rate of adoption

00:06:38.920 | of the new ideas.

00:06:41.360 | - Yeah, and the other interesting question

00:06:43.480 | is in the early days of particular discipline,

00:06:46.480 | I think you mentioned in that book

00:06:49.920 | was ultimately a book of cancer.

00:06:52.320 | - It's called "The Emperor of All Melodies."

00:06:55.080 | - Yeah, and those melodies included

00:06:57.840 | the trying to, the medicine, was it centered on--

00:07:00.520 | - So it was actually centered on

00:07:04.920 | how people thought of curing cancer.

00:07:07.200 | Like for me, it was really a discovery how people,

00:07:10.680 | what was the science of chemistry behind drug development,

00:07:14.120 | that it actually grew up out of dyeing,

00:07:17.240 | like coloring industry, that people who develop chemistry

00:07:21.920 | in 19th century in Germany and Britain

00:07:25.080 | to do the really new dyes,

00:07:28.120 | they looked at the molecules and identified

00:07:30.160 | that they do certain things to cells.

00:07:32.120 | And from there, the process started

00:07:34.480 | and like historically, yeah, this is fascinating

00:07:36.880 | that they managed to make the connection

00:07:38.680 | and look under the microscope and do all this discovery.

00:07:42.280 | But as you continue reading about it

00:07:44.440 | and you read about how chemotherapy drugs

00:07:48.760 | were actually developed in Boston

00:07:50.520 | and some of them were developed

00:07:52.480 | and Farber, Dr. Farber from Dana-Farber,

00:07:57.480 | you know, how the experiments were done,

00:08:00.480 | that there was some miscalculation,

00:08:03.320 | let's put it this way, and they tried it on the patients

00:08:05.960 | and those were children with leukemia and they died.

00:08:10.000 | And then they tried another modification.

00:08:11.640 | You look at the process, how imperfect is this process?

00:08:15.000 | And you know, like if we're again looking back

00:08:17.480 | like 60 years ago, 70 years ago,

00:08:19.160 | you can kind of understand it.

00:08:20.760 | But some of the stories in this book,

00:08:23.000 | which were really shocking to me,

00:08:24.600 | were really happening, you know, maybe decades ago.

00:08:28.000 | And we still don't have a vehicle

00:08:30.640 | to do it much more fast and effective

00:08:33.520 | and you know, scientific,

00:08:35.640 | the way I'm thinking computer science, scientific.

00:08:38.200 | - So from the perspective of computer science,

00:08:40.400 | you've gotten a chance to work the application

00:08:43.160 | to cancer and to medicine in general.

00:08:44.880 | From a perspective of an engineer and a computer scientist,

00:08:48.440 | how far along are we from understanding the human body,

00:08:51.800 | biology, of being able to manipulate it

00:08:55.160 | in a way we can cure some of the maladies,

00:08:57.920 | some of the diseases?

00:08:59.720 | - So this is very interesting question.

00:09:02.220 | And if you're thinking as a computer scientist

00:09:06.040 | about this problem, I think one of the reasons

00:09:09.840 | that we succeeded in the areas

00:09:11.880 | we as a computer scientist succeeded

00:09:14.000 | is because we don't have,

00:09:16.280 | we are not trying to understand in some ways.

00:09:19.000 | Like if you're thinking about like e-commerce, Amazon,

00:09:22.280 | Amazon doesn't really understand you

00:09:24.240 | and that's why it recommends you certain books

00:09:27.720 | or certain products, correct?

00:09:29.600 | And in, you know, traditionally,

00:09:34.240 | when people were thinking about marketing,

00:09:36.160 | you know, they divided the population

00:09:37.800 | to different kind of subgroups,

00:09:39.800 | identify the features of the subgroup

00:09:41.760 | and come up with a strategy

00:09:43.120 | which is specific to that subgroup.

00:09:45.560 | If you're looking about recommendation system,

00:09:47.320 | they're not claiming that they're understanding somebody,

00:09:50.600 | they're just managing to,

00:09:52.680 | from the patterns of your behavior,

00:09:54.760 | to recommend you a product.

00:09:57.560 | Now, if you look at the traditional biology,

00:09:59.600 | and obviously I wouldn't say that I,

00:10:03.200 | at any way, you know, educated in this field,

00:10:06.160 | but you know, what I see,

00:10:07.040 | there's really a lot of emphasis

00:10:09.280 | on mechanistic understanding.

00:10:10.640 | And it was very surprising to me

00:10:12.560 | coming from computer science

00:10:13.800 | how much emphasis is on this understanding.

00:10:17.600 | And given the complexity of the system,

00:10:20.720 | maybe the deterministic full understanding

00:10:23.200 | of this process is, you know, beyond our capacity.

00:10:27.360 | And the same way as in computer science,

00:10:29.440 | when we're doing recognition,

00:10:30.520 | when you do recommendation in many other areas,

00:10:32.760 | it's just probabilistic matching process.

00:10:35.960 | And in some way, maybe in certain cases,

00:10:40.080 | we shouldn't even attempt to understand,

00:10:42.920 | though we can attempt to understand,

00:10:44.480 | but in parallel, we can actually do this kind of matching

00:10:48.040 | that would help us to find hero

00:10:51.080 | to do early diagnostics and so on.

00:10:54.160 | And I know that in these communities,

00:10:55.880 | it's really important to understand,

00:10:59.080 | but I'm sometimes wondering,

00:11:00.720 | what exactly does it mean to understand here?

00:11:02.960 | - Well, there's stuff that works,

00:11:05.520 | but that can be, like you said,

00:11:07.640 | separate from this deep human desire

00:11:10.360 | to uncover the mysteries of the universe,

00:11:12.720 | of science, of the way the body works,

00:11:16.160 | the way the mind works.

00:11:17.600 | It's the dream of symbolic AI,

00:11:19.560 | of being able to reduce human knowledge into logic

00:11:24.000 | and be able to play with that logic

00:11:26.880 | in a way that's very explainable

00:11:28.680 | and understandable for us humans.

00:11:30.280 | I mean, that's a beautiful dream.

00:11:31.760 | So I understand it, but it seems that

00:11:34.840 | what seems to work today, and we'll talk about it more,

00:11:37.880 | is as much as possible, reduce stuff into data,

00:11:40.780 | reduce whatever problem you're interested in to data

00:11:43.880 | and try to apply statistical methods,

00:11:47.040 | apply machine learning to that.

00:11:49.120 | On a personal note,

00:11:51.120 | you were diagnosed with breast cancer in 2014.

00:11:54.160 | What did facing your mortality make you think about?

00:11:58.400 | How did it change you?

00:12:00.240 | - You know, this is a great question,

00:12:01.840 | and I think that I was interviewed many times,

00:12:03.800 | so nobody actually asked me this question.

00:12:05.720 | I think I was 43 at the time,

00:12:09.680 | and it's the first time I realized in my life

00:12:11.480 | that I may die, and I never thought about it before.

00:12:14.480 | And there was a long time since you diagnosed

00:12:17.280 | until you actually know what you have

00:12:18.600 | and how severe is your disease.

00:12:20.160 | For me, it was like maybe two and a half months.

00:12:23.520 | And I didn't know where I am during this time

00:12:28.320 | because I was getting different tests,

00:12:30.640 | and one would say it's bad, and I would say,

00:12:32.600 | no, it is not.

00:12:33.420 | So until I knew where I am,

00:12:34.840 | I really was thinking about

00:12:36.280 | all these different possible outcomes.

00:12:38.240 | - Were you imagining the worst,

00:12:39.720 | or were you trying to be optimistic?

00:12:41.960 | - It would be really, I don't remember

00:12:45.640 | what was my thinking.

00:12:47.400 | It was really a mixture with many components at the time,

00:12:51.120 | speaking in our terms.

00:12:54.120 | And one thing that I remember,

00:12:59.120 | and every test comes, and then you think,

00:13:01.560 | oh, it could be this, or it may not be this,

00:13:03.360 | and you're hopeful, and then you're desperate.

00:13:04.720 | So it's like there is a whole slew of emotions

00:13:07.720 | that go through you.

00:13:08.720 | But what I remember is that when I came back to MIT,

00:13:14.800 | I was kind of going the whole time

00:13:17.160 | through the treatment to MIT,

00:13:18.280 | but my brain was not really there.

00:13:19.800 | But when I came back, really finished my treatment,

00:13:21.800 | and I was here teaching and everything,

00:13:24.560 | you know, I look back at what my group was doing,

00:13:27.080 | what other groups was doing,

00:13:28.840 | and I saw these trivialities.

00:13:30.820 | It's like people are building their careers

00:13:33.240 | on improving some parts, around 2% or 3% or whatever.

00:13:36.920 | I was, it's like, seriously,

00:13:38.400 | I did a work on how to decipher Ugaritic,

00:13:40.760 | like a language that nobody speak, and whatever.

00:13:42.880 | Like, what is significance?

00:13:46.160 | When all of a sudden, you know, I walked out of MIT,

00:13:49.000 | which is, you know, when people really do care,

00:13:51.600 | you know, what happened to your Eclair paper,

00:13:54.240 | you know, what is your next publication,

00:13:56.640 | to ACL, to the world where people, you know,

00:13:59.920 | people, you see a lot of sufferings,

00:14:01.920 | and I'm kind of totally shielded on it on daily basis.

00:14:04.880 | And it's like the first time I've seen, like,

00:14:06.840 | real life and real suffering.

00:14:08.680 | And I was thinking,

00:14:10.720 | why are we trying to improve the parts there,

00:14:13.280 | or deal with some trivialities

00:14:16.120 | when we have capacity to really make a change?

00:14:20.720 | And it was really challenging to me,

00:14:23.480 | because on one hand, you know,

00:14:24.600 | I have my graduate students

00:14:25.760 | who really want to do their papers and their work,

00:14:28.720 | and they want to continue to do what they were doing,

00:14:30.840 | which was great.

00:14:31.920 | And then it was me who really kind of reevaluated

00:14:36.320 | what is the importance.

00:14:37.480 | And also at that point,

00:14:38.560 | because I had to take some break,

00:14:40.280 | I look back into, like, my years in science,

00:14:47.520 | and I was thinking, you know, like,

00:14:49.600 | 10 years ago, this was the biggest thing.

00:14:51.680 | I don't know, topic models.

00:14:52.960 | We have, like, millions of papers on topic models

00:14:55.360 | and variation of topics models,

00:14:56.520 | and I was totally, like, irrelevant.

00:14:58.600 | And you start looking at this, you know,

00:15:01.480 | what do you perceive as important

00:15:03.240 | at different point of time,

00:15:04.520 | and how, you know, it fades over time.

00:15:08.920 | And since we have a limited time,

00:15:13.000 | all of us have limited time on Earth,

00:15:14.960 | it's really important to prioritize

00:15:17.840 | things that really matter to you,

00:15:19.760 | maybe matter to you at that particular point,

00:15:22.040 | but it's important to take some time

00:15:24.400 | and understand what matters to you,

00:15:26.960 | which may not necessarily be the same

00:15:28.880 | as what matters to the rest of your scientific community,

00:15:31.720 | and pursue that vision.

00:15:34.600 | - So that moment, did it make you cognizant,

00:15:38.480 | you mentioned suffering,

00:15:39.600 | of just the general amount of suffering in the world.

00:15:44.400 | Is that what you're referring to?

00:15:45.680 | So as opposed to topic models

00:15:47.440 | and specific detailed problems in NLP,

00:15:50.840 | did you start to think about other people

00:15:54.480 | who have been diagnosed with cancer?

00:15:56.960 | Is that the way you started to see the world, perhaps?

00:16:00.040 | - Oh, absolutely, and it actually creates,

00:16:02.440 | because like, for instance, you know,

00:16:04.960 | there is parts of the treatment

00:16:06.040 | where you need to go to the hospital every day,

00:16:08.520 | and you see, you know, the community of people that you see,

00:16:11.600 | and many of them are much worse than I was at the time,

00:16:16.080 | and you all of a sudden see it all.

00:16:20.480 | And people who are happier some day

00:16:23.920 | just because they feel better,

00:16:25.320 | and for people who are in our normal realm,

00:16:28.480 | you take it totally for granted that you feel well,

00:16:30.800 | that if you decide to go running, you can go running,

00:16:32.920 | and you can, you know, you're pretty much free

00:16:35.880 | to do whatever you want with your body.

00:16:37.600 | Like, I saw like a community,

00:16:40.200 | my community became those people.

00:16:42.800 | And I remember one of my friends, Dina Khatabi,

00:16:47.480 | took me to Prudential to buy me a gift for my birthday,

00:16:50.400 | and it was like the first time in months

00:16:52.320 | that I went to kind of to see other people,

00:16:54.960 | and I was like, wow, first of all,

00:16:57.080 | these people, you know, they are happy,

00:16:58.960 | and they are laughing, and they're very different

00:17:00.640 | from these other my people.

00:17:02.600 | And second, I think it's totally crazy,

00:17:04.640 | they're like laughing and wasting their money

00:17:06.360 | on some stupid gifts, and, you know, they may die.

00:17:11.360 | They already may have cancer, and they don't understand it.

00:17:15.960 | So you can really see how the mind changes,

00:17:20.080 | that you can see that, you know,

00:17:22.360 | before that you can, as didn't you know

00:17:23.680 | that you're gonna die, of course I knew,

00:17:25.280 | but it was kind of a theoretical notion,

00:17:28.320 | it wasn't something which was concrete.

00:17:31.040 | And at that point, when you really see it,

00:17:33.880 | and see how little means sometimes the system

00:17:37.640 | has to harm them, you really feel that we need

00:17:41.000 | to take a lot of our brilliance that we have here at MIT

00:17:45.440 | and translate it into something useful.

00:17:48.040 | - Yeah, and useful can have a lot of definitions,

00:17:50.520 | but of course, alleviating suffering,

00:17:52.320 | alleviating trying to cure cancer is a beautiful mission.

00:17:57.320 | So I of course know theoretically the notion of cancer,

00:18:01.920 | but just reading more and more about it,

00:18:04.680 | it's 1.7 million new cancer cases

00:18:08.000 | in the United States every year,

00:18:09.840 | 600,000 cancer-related deaths every year.

00:18:13.480 | So this has a huge impact, United States globally.

00:18:19.340 | When broadly, before we talk about how machine learning,

00:18:24.340 | how MIT can help, when do you think we as a civilization

00:18:29.380 | will cure cancer?

00:18:32.100 | How hard of a problem is it from everything

00:18:34.620 | you've learned from it recently?

00:18:36.220 | - I cannot really assess it.

00:18:39.340 | What I do believe will happen with the advancement

00:18:42.100 | in machine learning, that a lot of types of cancer

00:18:45.940 | we will be able to predict way early

00:18:48.500 | and more effectively utilize existing treatments.

00:18:53.420 | I think, I hope at least, that with all the advancements

00:18:57.540 | in AI and drug discovery, we would be able

00:19:01.180 | to much faster find relevant molecules.

00:19:04.700 | What I'm not sure about is how long it will take

00:19:08.220 | the medical establishment and regulatory bodies

00:19:11.940 | to kind of catch up and to implement it.

00:19:14.780 | And I think this is a very big piece of puzzles

00:19:17.460 | that is currently not addressed.

00:19:20.420 | - That's a really interesting question.

00:19:21.820 | So first, a small detail that I think the answer is yes,

00:19:25.460 | but is cancer one of the diseases that when detected earlier

00:19:30.460 | that significantly improves the outcomes?

00:19:37.120 | 'Cause we will talk about there's the cure

00:19:41.020 | and then there is detection.

00:19:43.020 | And I think where machine learning can really help

00:19:45.180 | is earlier detection.

00:19:46.660 | So does detection help?

00:19:48.540 | - Detection is crucial.

00:19:49.660 | For instance, the vast majority of pancreatic cancer patients

00:19:53.940 | are detected at the stage that they are incurable.

00:19:57.300 | That's why they have such a terrible survival rate.

00:20:02.300 | It's like just few percent over five years.

00:20:07.340 | It's pretty much today a death sentence.

00:20:09.820 | But if you can discover this disease early,

00:20:14.500 | there are mechanisms to treat it.

00:20:16.740 | And in fact, I know a number of people who were diagnosed

00:20:20.740 | and saved just because they had food poisoning.

00:20:23.600 | They had terrible food poisoning.

00:20:25.020 | They went to ER and they got scan.

00:20:28.540 | There were early signs on the scan

00:20:30.660 | and that what saved their lives.

00:20:33.540 | But this wasn't really an accidental case.

00:20:35.820 | So as we become better, we would be able to help

00:20:41.220 | too many more people that are likely to develop diseases.

00:20:46.220 | And I just want to say that as I got more into this field,

00:20:50.980 | I realized that cancer is of course terrible disease,

00:20:53.580 | but there are really the whole slew of terrible diseases

00:20:56.660 | out there like neurodegenerative diseases and others.

00:21:00.780 | So we, of course, a lot of us are fixated on cancer

00:21:04.540 | just because it's so prevalent in our society

00:21:06.500 | and you see these people,

00:21:07.500 | but there are a lot of patients

00:21:08.620 | with neurodegenerative diseases

00:21:10.400 | and the kind of aging diseases

00:21:12.580 | that we still don't have a good solution for.

00:21:17.180 | And we, you know, and I felt as a computer scientist,

00:21:22.180 | we kind of decided that it's other people's job

00:21:25.540 | to treat these diseases

00:21:26.940 | because it's like traditionally people in biology

00:21:30.460 | or in chemistry or MDs are the ones who are thinking

00:21:33.660 | about it and have to kind of start paying attention.

00:21:37.460 | I think that it's really a wrong assumption.

00:21:40.380 | And we all need to join the battle.

00:21:43.020 | - So how, it seems like in cancer specifically,

00:21:46.540 | that there's a lot of ways that machine learning can help.

00:21:49.200 | So what's the role of machine learning

00:21:51.940 | in the diagnosis of cancer?

00:21:54.160 | - So for many cancers today,

00:21:57.260 | we really don't know what is your likelihood to get cancer.

00:22:02.260 | And for the vast majority of patients,

00:22:06.380 | especially on the younger patients,

00:22:08.020 | it really comes as a surprise.

00:22:09.620 | Like for instance, for breast cancer,

00:22:11.180 | 80% of the patients are first in their families.

00:22:13.900 | It's like me.

00:22:15.400 | And I never saw that I had any increased risk

00:22:18.500 | because, you know, nobody had it in my family.

00:22:20.860 | And for some reason in my head,

00:22:22.300 | it was kind of inherited disease.

00:22:24.840 | But even if I would pay attention,

00:22:28.380 | the models that currently,

00:22:30.220 | there's very simplistic statistical models

00:22:32.420 | that are currently used and in clinical practice,

00:22:34.540 | they really don't give you an answer.

00:22:35.820 | So you don't know.

00:22:37.460 | And the same true for pancreatic cancer,

00:22:40.380 | the same true for non-smoking lung cancer and many others.

00:22:45.380 | So what machine learning can do here

00:22:47.340 | is utilize all this data to tell us early

00:22:51.620 | who is likely to be susceptible

00:22:53.140 | and using all the information that is already there,

00:22:55.980 | be it imaging, be it your other tests,

00:22:59.980 | and, you know, eventually liquid biopsies and others

00:23:04.860 | where the signal itself is not sufficiently strong

00:23:08.180 | for human eye to do good discrimination

00:23:11.300 | because the signal may be weak.

00:23:12.900 | But by combining many sources,

00:23:15.620 | a machine which is trained on large volumes of data

00:23:18.100 | can really detect it early.

00:23:20.700 | And that's what we've seen with breast cancer

00:23:22.500 | and people are reporting it in other diseases as well.

00:23:25.900 | - That really boils down to data, right?

00:23:28.260 | And in the different kinds of sources of data.

00:23:30.980 | And you mentioned regulatory challenges.

00:23:33.720 | So what are the challenges in gathering

00:23:36.500 | large data sets in this space?

00:23:39.220 | - Again, another great question.

00:23:42.660 | So it took me after I decided that I want to work on it

00:23:45.500 | two years to get access to data.

00:23:48.740 | - Any data, like any significant data set?

00:23:50.580 | - Any significant amount.

00:23:51.420 | Like right now in this country,

00:23:53.580 | there is no publicly available data set of modern mammogram

00:23:58.060 | that you can just go on your computer,

00:23:59.540 | sign a document and get it.

00:24:01.860 | It just doesn't exist.

00:24:03.180 | I mean, obviously every hospital

00:24:05.280 | has its own collection of mammograms.

00:24:07.560 | There are data that came out of clinical trials.

00:24:11.300 | But we're talking about you as a computer scientist

00:24:13.220 | who just want to run his or her model

00:24:17.120 | and see how it works.

00:24:19.060 | This data, like ImageNet, doesn't exist.

00:24:22.900 | And there is a set which is called like Florida Dataset,

00:24:27.900 | which is a film mammogram from '90s,

00:24:30.860 | which is totally not representative

00:24:32.460 | of the current developments.

00:24:33.880 | Whatever you're learning on them doesn't scale up.

00:24:35.800 | This is the only resource that is available.

00:24:39.340 | And today there are many agencies

00:24:42.820 | that govern access to data.

00:24:44.460 | Like the hospital holds your data

00:24:46.300 | and the hospital decides whether they would give it

00:24:49.280 | to the researcher to work with his data.

00:24:52.220 | - An individual hospital?

00:24:54.180 | - Yeah, I mean, the hospital may,

00:24:56.120 | assuming that you're doing research collaboration,

00:24:59.220 | you can submit, there is a proper approval process

00:25:03.500 | guided by our RPE.

00:25:05.060 | And if you go through all the processes,

00:25:07.820 | you can eventually get access to the data.

00:25:10.140 | But if you yourself know, our AI community,

00:25:13.540 | there are not that many people

00:25:14.740 | who actually ever got access to data

00:25:16.620 | because it's a very challenging process.

00:25:20.220 | - And sorry, just a quick comment.

00:25:22.780 | MGH or any kind of hospital, are they scanning the data?

00:25:28.100 | Are they digitally storing it?

00:25:29.700 | - Oh, it is already digitally stored.

00:25:31.580 | You don't need to do any extra processing steps.

00:25:34.160 | It's already there in the right format.

00:25:36.340 | Is that right now there are a lot of issues

00:25:39.820 | that govern access to the data

00:25:41.180 | because the hospital is legally responsible for the data.

00:25:46.180 | And they have a lot to lose

00:25:51.100 | if they give the data to the wrong person,

00:25:53.220 | but they may not have a lot to gain if they give it,

00:25:56.540 | as a hospital, as a legal entity is giving it to you.

00:26:00.660 | And the way, you know, what I would imagine

00:26:02.820 | happening in the future is the same thing that happens

00:26:05.300 | when you're getting your driving license.

00:26:06.860 | You can decide whether you want to donate your organs.

00:26:09.900 | You can imagine that whenever a person goes to the hospital,

00:26:12.900 | it should be easy for them to donate their data for research

00:26:18.100 | and it can be different kind of,

00:26:19.500 | do they only give you a test results

00:26:21.340 | or only imaging data or the whole medical record?

00:26:26.780 | Because at the end,

00:26:29.020 | we all will benefit from all this insights.

00:26:33.900 | And it's not like you say, I want to keep my data private,

00:26:36.100 | but I would really love to get it from other people

00:26:38.820 | because other people are thinking the same way.

00:26:40.780 | So if there is a mechanism to do this donation

00:26:45.780 | and the patient has an ability to say

00:26:48.060 | how they want to use their data for research,

00:26:50.860 | it would be really a game changer.

00:26:54.140 | - People, when they think about this problem,

00:26:56.500 | there's a, it depends on the population,

00:26:58.500 | depends on the demographics,

00:27:00.180 | but there's some privacy concerns.

00:27:02.460 | Generally, not just medical data, just any kind of data.

00:27:05.900 | It's what you said, my data, it should belong kind of to me.

00:27:09.660 | I'm worried how it's going to be misused.

00:27:11.660 | How do we alleviate those concerns?

00:27:15.660 | Because that seems like a problem that needs to be,

00:27:19.500 | that problem of trust, of transparency,

00:27:21.660 | needs to be solved before we build large datasets

00:27:25.180 | that help detect cancer,

00:27:27.300 | help save those very people in the future.

00:27:30.220 | - So I think there are two things that could be done.

00:27:31.980 | There is a technical solutions

00:27:34.460 | and there are societal solutions.

00:27:38.220 | So on the technical end,

00:27:40.180 | we today have ability to improve disambiguation,

00:27:46.440 | like for instance, for imaging,

00:27:51.380 | for imaging, you can do it pretty well.

00:27:55.620 | - What's disambiguation?

00:27:56.780 | - And disambiguation, sorry, disambiguation,

00:27:58.540 | removing the identification,

00:27:59.860 | removing the names of the people.

00:28:02.220 | There are other data, like if it is a raw text,

00:28:04.860 | you cannot really achieve 99.9%,

00:28:08.180 | but there are all these techniques,

00:28:10.060 | and actually some of them are developed at MIT,

00:28:12.500 | how you can do learning on the encoded data,

00:28:15.460 | where you locally encode the image,

00:28:17.420 | you train a network which only works

00:28:19.900 | on the encoded images,

00:28:22.420 | and then you send the outcome back to the hospital

00:28:24.940 | and you can open it up.

00:28:26.580 | So those are the technical solution.

00:28:28.020 | There are a lot of people who are working in this space

00:28:30.660 | where the learning happens in the encoded form.

00:28:33.780 | We are still early,

00:28:35.300 | but this is an interesting research area

00:28:39.220 | where I think we'll make more progress.

00:28:41.860 | There is a lot of work in natural language

00:28:45.180 | processing community, how to do the identification better.

00:28:49.560 | But even today, there are already a lot of data

00:28:54.000 | which can be de-identified perfectly,

00:28:55.880 | like your test data, for instance, correct?

00:28:58.720 | Where you can just, you know the name of the patient,

00:29:00.960 | you just want to extract the part with the numbers.

00:29:04.280 | The big problem here is again,

00:29:07.440 | hospitals don't see much incentive

00:29:10.400 | to give this data away on one hand,

00:29:12.600 | and then there is general concern.

00:29:14.200 | Now when I'm talking about societal benefits

00:29:17.680 | and about the education,

00:29:19.640 | the public needs to understand,

00:29:23.640 | and I think that there are situations,

00:29:27.880 | and I still remember myself when I really needed an answer.

00:29:31.560 | I had to make a choice,

00:29:33.360 | and there was no information to make a choice.

00:29:35.280 | You're just guessing.

00:29:36.720 | And at that moment, you feel that your life is at stake,

00:29:41.100 | but you just don't have information to make the choice.

00:29:44.840 | And many times when I give talks,

00:29:48.760 | I get emails from women who say,

00:29:51.320 | you know, I'm in this situation,

00:29:52.840 | can you please run statistics and see what are the outcomes?

00:29:55.980 | We get almost every week a mammogram

00:30:00.080 | that comes by mail to my office at MIT, I'm serious.

00:30:03.500 | That people ask to run because they need to make

00:30:07.920 | life-changing decisions.

00:30:10.060 | And of course, I'm not planning to open a clinic here,

00:30:13.000 | but we do run and give them the results for their doctors.

00:30:16.640 | But the point that I'm trying to make,

00:30:20.080 | that we all at some point, or our loved ones,

00:30:23.760 | will be in the situation where you need information

00:30:26.640 | to make the best choice.

00:30:28.880 | And if this information is not available,

00:30:31.880 | you would feel vulnerable and unprotected.

00:30:35.120 | And then the question is, you know, what do I care more?

00:30:37.880 | Because at the end, everything is a trade-off, correct?

00:30:40.360 | - Yeah, exactly.

00:30:41.680 | Just out of curiosity,

00:30:43.560 | it seems like one possible solution,

00:30:45.600 | I'd like to see what you think of it,

00:30:47.440 | based on what you just said,

00:30:50.680 | based on wanting to know answers

00:30:52.520 | for when you're yourself in that situation.

00:30:55.060 | Is it possible for patients to own their data

00:30:58.400 | as opposed to hospitals owning their data?

00:31:01.040 | Of course, theoretically, I guess patients own their data,

00:31:04.120 | but can you walk out there with a USB stick

00:31:06.640 | containing everything, or upload it to the cloud,

00:31:10.620 | where a company, you know,

00:31:13.000 | I remember Microsoft had a service,

00:31:15.680 | like I was really excited about,

00:31:17.800 | and Google Health was there.

00:31:19.280 | I tried to give, I was excited about it.

00:31:21.880 | Basically, companies helping you upload your data

00:31:24.780 | to the cloud, so that you can move from hospital to hospital,

00:31:27.960 | from doctor to doctor.

00:31:29.240 | Do you see a promise of that kind of possibility?

00:31:32.680 | - I absolutely think this is, you know,

00:31:34.680 | the right way to exchange the data.

00:31:38.160 | I don't know now who's the biggest player in this field,

00:31:41.700 | but I can clearly see that even for,

00:31:44.620 | even for totally selfish health reasons,

00:31:46.820 | when you are going to a new facility,

00:31:49.300 | and many of us are sent to some specialized treatment,

00:31:52.620 | they don't easily have access to your data.

00:31:55.740 | And today, you know,

00:31:57.940 | we would want to send this mammogram,

00:31:59.420 | need to go to the hospital, find some small office,

00:32:01.780 | which gives them the CD, and they ship us the CD,

00:32:04.820 | so you can imagine we're looking

00:32:06.460 | at kind of decades old mechanism of data exchange.

00:32:10.140 | So I definitely think this is an area

00:32:14.060 | where hopefully all the right regulatory

00:32:18.260 | and technical forces will align,

00:32:20.380 | and we will see it actually implemented.

00:32:23.220 | - It's sad because unfortunately,

00:32:25.740 | and I need to research why that happened,

00:32:28.420 | but I'm pretty sure Google Health

00:32:30.620 | and Microsoft Health Vault, or whatever it's called,

00:32:32.940 | both closed down, which means that there was

00:32:36.100 | either regulatory pressure, or there's not a business case,

00:32:39.100 | or there's challenges from hospitals,

00:32:41.820 | which is very disappointing.

00:32:43.260 | So when you say you don't know

00:32:44.540 | what the biggest players are,

00:32:46.520 | the two biggest that I was aware of closed their doors.

00:32:50.540 | So I'm hoping, I'd love to see why,

00:32:53.140 | and I'd love to see who else can come up.

00:32:54.780 | It seems like one of those Elon Musk style problems

00:32:59.620 | that are obvious needs to be solved,

00:33:01.300 | and somebody needs to step up

00:33:02.380 | and actually do this large scale data,

00:33:05.180 | you know, data collection.

00:33:07.540 | - So I know there is an initiative in Massachusetts,

00:33:09.660 | I think, actually led by the governor,

00:33:11.780 | to try to create this kind of health exchange system,

00:33:15.500 | where at least to help people who kind of,

00:33:17.260 | when you show up in emergency room

00:33:18.700 | and there is no information about

00:33:20.580 | what are your allergies and other things.

00:33:22.620 | So I don't know how far it will go.

00:33:26.180 | But another thing that you said,

00:33:28.220 | and I find it very interesting,

00:33:30.340 | is actually who are the successful players in this space,

00:33:33.820 | and the whole implementation, how does it go?

00:33:37.260 | To me, it is from the anthropological perspective,

00:33:40.300 | it's more fascinating that AI that today goes in healthcare.

00:33:44.140 | You know, we've seen so many attempts

00:33:47.900 | and so very little successes.

00:33:50.380 | And it's interesting to understand that I by no means,

00:33:53.860 | you know, have knowledge to assess it,

00:33:56.680 | why we are in the position where we are.

00:33:59.620 | - Yeah, it's interesting, 'cause data is really fuel

00:34:02.940 | for a lot of successful applications.

00:34:04.980 | And when that data requires regulatory approval,

00:34:08.500 | like the FDA or any kind of approval,

00:34:12.440 | it seems that the computer scientists

00:34:15.740 | are not quite there yet in being able

00:34:17.460 | to play the regulatory game,

00:34:18.900 | understanding the fundamentals of it.

00:34:21.220 | - I think that in many cases,

00:34:23.700 | when even people do have data,

00:34:26.500 | we still don't know what exactly do you need to demonstrate

00:34:31.580 | to change the standard of care.

00:34:33.880 | Like, let me give you an example

00:34:37.220 | related to my breast cancer research.

00:34:41.100 | So in traditional breast cancer risk assessment,

00:34:45.500 | there is something called density,

00:34:47.100 | which determines the likelihood of a woman to get cancer.

00:34:50.500 | And this pretty much says how much white

00:34:52.800 | do you see on the mammogram?

00:34:54.220 | The whiter it is, the more likely the tissue is dense.

00:34:58.980 | And the idea behind density, it's not a bad idea.

00:35:03.660 | In 1967, a radiologist called Wolf

00:35:06.860 | decided to look back at women who were diagnosed

00:35:09.780 | and see what is special in their images.

00:35:12.420 | Can we look back and say that they're likely to develop?

00:35:14.700 | So he come up with some patterns.

00:35:16.180 | It was the best that his human eye can identify.

00:35:20.620 | Then it was kind of formalized and coded

00:35:22.660 | into four categories, and that's what we are using today.

00:35:26.940 | And today, this density assessment

00:35:30.980 | is actually a federal law from 2019,

00:35:34.580 | approved by President Trump

00:35:36.140 | and for the previous FDA commissioner,

00:35:38.780 | where women are supposed to be advised by their providers

00:35:43.620 | if they have high density,

00:35:45.100 | putting them into higher risk category.

00:35:47.260 | And in some states, you can actually get

00:35:50.220 | supplementary screening paid by your insurance

00:35:52.460 | because you're in this category.

00:35:53.700 | Now you can say, how much science do we have behind it?

00:35:56.780 | Whatever, biological science or epidemiological evidence.

00:36:00.860 | So it turns out that between 40 and 50% of women

00:36:05.180 | have dense breast.

00:36:06.660 | So about 40% of patients are coming out of their screening

00:36:11.140 | and somebody tells them, you are in high risk.

00:36:15.060 | Now, what exactly does it mean

00:36:16.900 | if you as half of the population are in high risk?

00:36:19.620 | It's from saying, maybe I'm not,

00:36:22.060 | or what do I really need to do with it?

00:36:23.700 | Because the system doesn't provide me

00:36:27.220 | a lot of the solutions

00:36:28.340 | because there are so many people like me,

00:36:30.180 | we cannot really provide very expensive solutions for them.

00:36:34.620 | And the reason this whole density became this big deal,

00:36:38.740 | it's actually advocated by the patients

00:36:40.820 | who felt very unprotected

00:36:42.500 | because many women went in the mammograms,

00:36:44.900 | which were normal.

00:36:46.260 | And then it turns out that they already had cancer,

00:36:49.460 | quite developed cancer.

00:36:50.580 | So they didn't have a way to know who is really at risk

00:36:54.420 | and what is the likelihood that when the doctor tells you,

00:36:56.300 | you're okay, you are not okay.

00:36:58.060 | So at the time, and it was 15 years ago,

00:37:02.140 | this maybe was the best piece of science that we had.

00:37:06.820 | And it took quite 15, 16 years to make it federal law.

00:37:11.820 | But now this is a standard.

00:37:15.660 | Now with a deep learning model,

00:37:17.620 | we can so much more accurately predict

00:37:19.660 | who is gonna develop breast cancer

00:37:21.620 | just because you are trained on a logical thing.

00:37:23.700 | And instead of describing how much white

00:37:26.060 | and what kind of white machine

00:37:27.380 | can systematically identify the patterns,

00:37:30.140 | which was the original idea

00:37:31.980 | behind the sort of the tradiologist machine

00:37:34.020 | is can do it much more systematically

00:37:35.700 | and predict the risk when you're training the machine

00:37:38.260 | to look at the image

00:37:39.100 | and to say the risk in one, two, five years.

00:37:42.140 | Now you can ask me how long it will take

00:37:45.020 | to substitute this density,

00:37:46.460 | which is broadly used across the country.

00:37:48.620 | And I really it's not helping to bring this new models.

00:37:53.620 | And I would say it's not a matter of the algorithm.

00:37:56.700 | Algorithms already orders of magnitude better

00:37:58.780 | than what is currently in practice.

00:38:00.460 | I think it's really the question,

00:38:02.500 | who do you need to convince?

00:38:04.380 | How many hospitals do you need to run the experiment?

00:38:07.540 | What, you know, all this mechanism of adoption

00:38:11.660 | and how do you explain to patients

00:38:15.180 | and to women across the country

00:38:17.620 | that this is really a better measure?

00:38:20.460 | And again, I don't think it's an AI question.

00:38:22.740 | We can work more and make the algorithm even better,

00:38:25.940 | but I don't think that this is the current,

00:38:28.140 | you know, the barrier.

00:38:29.980 | The barrier is really this other piece

00:38:32.060 | that for some reason is not really explored.

00:38:35.260 | It's like anthropological piece.

00:38:36.860 | And coming back to your question about books,

00:38:39.860 | there is a book that I'm reading.

00:38:42.980 | It's called "American Sickness" by Elizabeth Rosenthal.

00:38:47.980 | And I got this book from my clinical collaborator,

00:38:51.580 | Dr. Connie Lehman.

00:38:53.100 | And I said, I know everything that I need to know

00:38:54.820 | about American health system,

00:38:56.020 | but you know, every page doesn't fail to surprise me.

00:38:59.220 | And I think that there is a lot of interesting

00:39:03.140 | and really deep lessons for people like us

00:39:06.860 | from computer science who are coming into this field

00:39:09.620 | to really understand how complex

00:39:11.540 | is the system of incentives in the system

00:39:14.820 | to understand how you really need to play

00:39:17.660 | to drive adoption.

00:39:18.780 | - You just said it's complex,

00:39:21.160 | but if we're trying to simplify it,

00:39:23.980 | who do you think most likely would be successful

00:39:27.380 | if we push on this group of people?

00:39:29.540 | Is it the doctors?

00:39:30.740 | Is it the hospitals?

00:39:31.820 | Is it the governments or policy makers?

00:39:34.300 | Is it the individual patients, consumers?

00:39:38.860 | Who needs to be inspired to most likely lead to adoption?

00:39:43.860 | Or is there no simple answer?

00:39:47.100 | - There's no simple answer,

00:39:48.260 | but I think there is a lot of good people in medical system

00:39:51.980 | who do want, you know, to make a change.

00:39:55.200 | And I think a lot of power will come from us as consumers,

00:40:01.480 | because we all are consumers or future consumers

00:40:04.260 | of healthcare services.

00:40:06.500 | And I think we can do so much more

00:40:11.500 | in explaining the potential and not in the hype terms

00:40:15.540 | and not saying that we now cured all Alzheimer

00:40:17.900 | and, you know, I'm really sick of reading

00:40:19.500 | these kind of articles which make these claims,

00:40:22.100 | but really to show with some examples

00:40:24.780 | what this implementation does and how it changes the care.

00:40:29.060 | Because I can't imagine,

00:40:30.020 | it doesn't matter what kind of politician it is,

00:40:32.620 | you know, we all are susceptible to these diseases.

00:40:35.220 | There is no one who is free.

00:40:37.740 | And eventually, you know, we all are humans

00:40:41.060 | and we're looking for a way to alleviate the suffering.

00:40:44.860 | And this is one possible way

00:40:47.260 | where we currently are underutilizing,

00:40:49.300 | which I think can help.

00:40:50.940 | - So it sounds like the biggest problems are outside of AI

00:40:55.100 | in terms of the biggest impact at this point.

00:40:57.960 | But are there any open problems

00:41:00.420 | in the application of ML to oncology in general?

00:41:03.780 | So improving the detection or any other creative methods,

00:41:07.540 | whether it's on the detection segmentations

00:41:09.620 | or the vision perception side

00:41:11.780 | or some other clever inference.

00:41:16.260 | Yeah, what in general in your view

00:41:18.500 | are the open problems in this space?

00:41:20.340 | - Yeah, I just want to mention that beside detection,

00:41:22.420 | another area where I am kind of quite active

00:41:24.820 | and I think it's really an increasingly important area

00:41:28.580 | in healthcare is drug design.

00:41:30.960 | - Absolutely.

00:41:33.100 | - Because, you know, it's fine if you detect something early

00:41:36.900 | but you still need to get drugs

00:41:41.100 | and new drugs for these conditions.

00:41:43.860 | And today, all of the drug design, ML is non-existent there.

00:41:48.300 | We don't have any drug that was developed by the ML model

00:41:52.980 | or even not developed, but at least even knew

00:41:56.220 | that ML model plays some significant role.

00:41:59.260 | I think this area was all the new ability

00:42:03.300 | to generate molecules with desired properties

00:42:05.780 | to do in silica screening is really a big open area.

00:42:10.780 | To be totally honest with you,

00:42:12.740 | when we are doing diagnostics and imaging,

00:42:14.940 | primarily taking the ideas that were developed

00:42:17.260 | for other areas and you applying them with some adaptation,

00:42:20.460 | the area of drug design is really technically interesting

00:42:25.460 | and exciting area.

00:42:27.980 | You need to work a lot with graphs

00:42:30.380 | and capture various 3D properties.

00:42:34.620 | There are lots and lots of opportunities

00:42:37.420 | to be technically creative.

00:42:39.820 | And I think there are a lot of open questions in this area.

00:42:44.820 | You know, we're already getting a lot of successes

00:42:48.820 | even with the kind of the first generation of these models,

00:42:52.700 | but there is much more new creative things that you can do.

00:42:56.500 | And what's very nice to see is that actually

00:42:59.300 | the more powerful, the more interesting models

00:43:04.180 | actually do do better.

00:43:05.460 | So there is a place to innovate

00:43:10.100 | in machine learning in this area.

00:43:12.520 | And some of these techniques are really unique

00:43:16.660 | to let's say to graph generation and other things.

00:43:19.700 | - Just to interrupt really quick, I'm sorry.

00:43:23.980 | Graph generation or graphs, drug discovery in general,

00:43:28.980 | how do you discover a drug?

00:43:31.980 | Is this chemistry?

00:43:33.340 | Is this trying to predict different chemical reactions?

00:43:37.500 | Or is it some kind of, what do graphs even represent

00:43:41.180 | in this space?

00:43:42.020 | - Oh, sorry, sorry.

00:43:43.980 | - And what's a drug?

00:43:45.300 | - Okay, so let's say you're thinking

00:43:47.100 | there are many different types of drugs,

00:43:48.500 | but let's say you're gonna talk about small molecules

00:43:50.540 | because I think today the majority of drugs

00:43:52.820 | are small molecules.

00:43:53.660 | So small molecule is a graph.

00:43:55.020 | The molecule is just where the node in the graph is an atom

00:44:00.020 | and then you have the bond.

00:44:01.500 | So it's really a graph representation

00:44:03.220 | if you're looking at it in 2D, correct?

00:44:05.540 | You can do it 3D, but let's say,

00:44:07.460 | let's keep it simple and stick in 2D.

00:44:09.560 | So pretty much my understanding today

00:44:14.740 | how it is done at scale in the companies

00:44:17.740 | you're without machine learning,

00:44:20.220 | you have high throughput screening.

00:44:22.100 | So you know that you are interested

00:44:23.740 | to get certain biological activity of the compound.

00:44:26.540 | So you scan a lot of compounds,

00:44:28.860 | like maybe hundreds of thousands,

00:44:30.700 | some really big number of compounds.

00:44:33.020 | You identify some compounds which have the right activity

00:44:36.100 | and then at this point, the chemists come

00:44:39.260 | and they're trying to now to optimize this original heat

00:44:44.260 | to different properties that you want it to be,

00:44:46.340 | maybe soluble, you want it to decrease toxicity,

00:44:49.100 | you want it to decrease the side effects.

00:44:51.660 | - Are those, sorry again to interrupt,

00:44:54.060 | can that be done in simulation

00:44:55.500 | or just by looking at the molecules

00:44:57.700 | or do you need to actually run reactions

00:44:59.860 | in real labs with laptops and stuff?

00:45:02.500 | - So when you do high throughput screening,

00:45:04.020 | you really do screening, it's in the lab.

00:45:07.060 | It's really the lab screening.

00:45:09.140 | You screen the molecules, correct?

00:45:10.980 | - I don't know what screening is.

00:45:12.620 | - The screening is just check them for certain property.

00:45:15.100 | - Like in the physical space, in the physical world,

00:45:17.300 | like actually there's a machine probably

00:45:18.780 | that's doing some, that's actually running the reaction.

00:45:21.460 | - Actually running the reactions, yeah.

00:45:22.900 | So there is a process where you can run

00:45:25.420 | and that's why it's called high throughput,

00:45:27.100 | it become cheaper and faster to do it

00:45:30.060 | on very big number of molecules.

00:45:33.820 | You run the screening, you identify potential,

00:45:37.660 | potential good starts and then when the chemists come in

00:45:42.340 | who have done it many times

00:45:44.060 | and then they can try to look at it and say,

00:45:46.180 | how can you change the molecule

00:45:48.260 | to get the desired profile in terms of all other properties?

00:45:53.260 | So maybe how do I make it more bioactive and so on?

00:45:56.500 | And there, the creativity of the chemists

00:45:59.460 | really is the one that determines the success of this design

00:46:04.460 | because again, they have a lot of domain knowledge

00:46:09.300 | of what works, how do you decrease the CCD and so on

00:46:12.900 | and that's what they do.

00:46:15.020 | So all the drugs that are currently in the FDA approved

00:46:19.780 | drugs or even drugs that are in clinical trials,

00:46:22.140 | they are designed using these domain experts

00:46:27.140 | which goes through this combinatorial space

00:46:30.060 | of molecules or graphs or whatever

00:46:31.980 | and find the right one or adjust it to be the right ones.

00:46:35.180 | - Sounds like the breast density heuristic

00:46:38.060 | from '67, the same echoes.

00:46:40.500 | - It's not necessarily that, it's really driven

00:46:44.260 | by deep understanding, it's not like they just observe it.

00:46:46.860 | I mean, they do deeply understand chemistry

00:46:48.580 | and they do understand how different groups

00:46:50.460 | and how does it change the properties.

00:46:53.140 | So there is a lot of science that gets into it

00:46:56.660 | and a lot of kind of simulation,

00:46:58.740 | how do you want it to behave?

00:47:00.940 | It's very, very complex.

00:47:03.900 | - So they're quite effective at this design, obviously.

00:47:06.140 | - Now, effective, yeah, we have drugs.

00:47:08.420 | Like depending on how do you measure effective?

00:47:10.780 | If you measure it in terms of cost, it's prohibitive.

00:47:13.940 | If you measure it in terms of times,

00:47:15.780 | we have lots of diseases for which we don't have any drugs

00:47:18.380 | and we don't even know how to approach

00:47:20.020 | and don't need to mention few drugs

00:47:23.420 | or neurodegenerative disease drugs that fail.

00:47:27.100 | So there are lots of trials that fail in later stages

00:47:32.100 | which is really catastrophic from the financial perspective.

00:47:35.140 | So is it the effective, the most effective mechanism?

00:47:39.500 | Absolutely no, but this is the only one

00:47:41.420 | that currently works.

00:47:43.700 | - And I was closely interacting

00:47:47.340 | with people in pharmaceutical industry.

00:47:48.660 | I was really fascinated on how sharp

00:47:50.780 | and what a deep understanding of the domain do they have.

00:47:54.700 | It's not observation driven.

00:47:56.460 | There is really a lot of science behind what they do.

00:47:59.620 | But if you ask me, can machine learning change it?

00:48:01.700 | I firmly believe yes,

00:48:04.700 | because even the most experienced chemists

00:48:07.260 | cannot hold in their memory and understanding

00:48:10.500 | everything that you can learn from millions

00:48:13.700 | of molecules and reactions.

00:48:15.460 | - And the space of graphs is a totally new space.

00:48:19.940 | I mean, it's a really interesting space

00:48:22.100 | for machine learning to explore, graph generation.

00:48:24.020 | - Yeah, so there are a lot of things that you can do here.

00:48:26.300 | So we do a lot of work.

00:48:28.780 | So the first tool that we started with

00:48:31.660 | was the tool that can predict properties of the molecules.

00:48:36.340 | So you can just give the molecule and the property.

00:48:39.460 | It can be bioactivity property

00:48:41.340 | or it can be some other property.

00:48:44.300 | And you train the molecules

00:48:46.500 | and you can now take a new molecule

00:48:50.060 | and predict this property.

00:48:52.220 | Now, when people started working in this area,

00:48:54.940 | it is something very simple.

00:48:56.020 | They do kind of existing fingerprints,

00:48:58.620 | which is kind of handcrafted features of the molecule

00:49:00.780 | when you break the graph to substructures

00:49:03.020 | and then you run, I don't know, feed forward neural network.

00:49:06.020 | And what was interesting to see that clearly,

00:49:08.540 | this was not the most effective way to proceed.

00:49:11.060 | And you need to have much more complex models

00:49:14.140 | that can induce the representation,

00:49:16.340 | which can translate this graph into the embeddings

00:49:19.260 | and do these predictions.

00:49:21.340 | So this is one direction.

00:49:23.260 | Then another direction, which is kind of related

00:49:25.340 | is not only to stop by looking at the embedding itself,

00:49:29.260 | but actually modify it to produce better molecules.

00:49:32.860 | So you can think about it as machine translation,

00:49:36.060 | that you can start with a molecule

00:49:38.220 | and then there is an improved version of molecule.

00:49:40.660 | And you can again, within code,

00:49:42.460 | translate it into the hidden space

00:49:43.940 | and then learn how to modify it to improve

00:49:46.140 | the in some ways version of the molecules.

00:49:49.420 | So that's, it's kind of really exciting.

00:49:52.700 | We already have seen that the property prediction

00:49:54.780 | works pretty well.

00:49:56.220 | And now we are generating molecules

00:49:59.820 | and there is actually labs

00:50:01.900 | which are manufacturing this molecule.

00:50:04.260 | So we'll see where it will get us.

00:50:06.420 | - Okay, that's really exciting.

00:50:07.820 | That's a lot of promise.

00:50:08.940 | Speaking of machine translation and embeddings,

00:50:11.900 | you have done a lot of really great research in NLP,

00:50:16.180 | natural language processing.

00:50:17.580 | Can you tell me your journey through NLP?

00:50:21.540 | What ideas, problems, approaches were you working on?

00:50:25.100 | Were you fascinated with?

00:50:26.500 | Did you explore before this magic of deep learning

00:50:31.380 | re-emerged and after?

00:50:34.060 | - So when I started my work in NLP,

00:50:35.940 | it was in '97.

00:50:38.140 | This was very interesting time.

00:50:39.420 | It was exactly the time that I came to ACL

00:50:42.540 | and the time I could barely understand English.

00:50:46.100 | But it was exactly like the transition point

00:50:48.460 | because half of the papers were really,

00:50:51.420 | you know, rule-based approaches

00:50:53.460 | where people took more kind of heavy linguistic approaches

00:50:56.140 | for small domains and try to build up from there.

00:51:00.020 | And then there were the first generation of papers

00:51:02.180 | which were corpus-based papers.

00:51:04.460 | And they were very simple in our terms

00:51:06.420 | when you collect some statistics

00:51:07.900 | and do prediction based on them.

00:51:10.060 | But I found it really fascinating that, you know,

00:51:12.340 | one community can think so very differently

00:51:15.620 | about the problem.

00:51:19.260 | And I remember my first paper that I wrote,

00:51:22.860 | it didn't have a single formula.

00:51:24.500 | It didn't have evaluation.

00:51:25.740 | It just had examples of outputs.

00:51:28.340 | And this was a standard of the field at the time.

00:51:32.060 | In some ways, I mean, people maybe just started

00:51:34.940 | emphasizing their empirical evaluation,

00:51:37.860 | but for many applications like summarization,

00:51:40.100 | you just show some examples of outputs.

00:51:42.780 | And then increasingly, you can see that

00:51:44.660 | how the statistical approaches dominated the field.

00:51:48.300 | And we've seen, you know, increased performance

00:51:52.100 | across many basic tasks.

00:51:56.020 | The sad part of the story may be that

00:51:59.300 | if you look again through this journey,

00:52:01.580 | we see that the role of linguistics

00:52:05.100 | in some ways greatly diminishes.

00:52:07.460 | And I think that you really need to look

00:52:11.580 | through the whole proceeding to find one or two papers

00:52:14.540 | which make some interesting linguistic references.

00:52:17.260 | - You mean today.

00:52:18.460 | - Today, today.

00:52:19.740 | This was definitely--

00:52:20.580 | - Things like syntactic trees,

00:52:21.620 | just even basically against our conversation

00:52:24.420 | about human understanding of language,

00:52:27.540 | which I guess what linguistics would be,

00:52:30.300 | structured, hierarchical, representing language

00:52:34.300 | in a way that's human explainable, understandable,

00:52:37.140 | is missing today.

00:52:39.340 | - I don't know if it is,

00:52:41.140 | what is explainable and understandable.

00:52:43.620 | In the end, you know, we perform functions,

00:52:45.940 | and it's okay to have a machine which performs a function.

00:52:50.180 | Like when you're thinking about your calculator, correct?

00:52:53.220 | Your calculator can do calculation very different

00:52:56.100 | from you would do the calculation,

00:52:57.620 | but it's very effective in it.

00:52:58.860 | And this is fine if we can achieve certain tasks

00:53:02.540 | with high accuracy,

00:53:04.460 | it doesn't necessarily mean that it has to understand

00:53:07.180 | in the same way as we understand.

00:53:09.260 | In some ways, it's even naive to request

00:53:11.260 | because you have so many other sources of information

00:53:14.940 | that are absent when you are training your system.

00:53:17.900 | So it's okay as it delivers it.

00:53:20.020 | And I will tell you one application

00:53:21.500 | that is really fascinating.

00:53:22.780 | In '97, when it came to ACL,

00:53:24.260 | there were some papers on machine translation.

00:53:25.860 | They were like primitive,

00:53:27.420 | like people were trying really, really simple.

00:53:31.020 | And the feeling, my feeling was that, you know,

00:53:34.220 | to make real machine translation system,

00:53:36.220 | it's like to fly in the moon and build a house there

00:53:39.540 | and a garden and live happily ever after.

00:53:41.540 | I mean, it's like impossible.

00:53:42.580 | I never could imagine that within, you know, 10 years,

00:53:46.700 | we would already see the system working.

00:53:48.500 | And now, you know, nobody is even surprised

00:53:51.380 | to utilize the system on daily basis.

00:53:54.420 | So this was like a huge, huge progress

00:53:56.220 | in the sense that people for a very long time

00:53:57.860 | tried to solve using other mechanisms

00:54:00.820 | and they were unable to solve it.

00:54:03.180 | That's why coming back to a question about biology,

00:54:06.140 | that in linguistics, people try to go this way

00:54:10.780 | and try to write the syntactic trees

00:54:13.500 | and try to obstruct it and to find the right representation.

00:54:17.860 | And, you know, they couldn't get very far

00:54:22.220 | with this understanding while these models,

00:54:25.980 | using, you know, other sources,

00:54:28.740 | actually capable to make a lot of progress.

00:54:31.700 | Now, I'm not naive to think

00:54:33.940 | that we are in this paradise space in NLP.

00:54:36.820 | And I'm sure as you know,

00:54:38.580 | that when we slightly change the domain

00:54:40.860 | and when we decrease the amount of training,

00:54:42.620 | it can do like really bizarre and funny thing.

00:54:44.740 | But I think it's just a matter

00:54:46.500 | of improving generalization capacity,

00:54:48.540 | which is just a technical question.

00:54:51.500 | - Wow, so that's the question.

00:54:54.300 | How much of language understanding

00:54:57.020 | can be solved with deep neural networks?

00:54:59.180 | In your intuition, I mean, it's unknown, I suppose.

00:55:03.740 | But as we start to creep towards romantic notions

00:55:07.620 | of the spirit of the Turing test

00:55:10.620 | and conversation and dialogue

00:55:14.220 | and something that maybe to me or to us silly humans

00:55:19.020 | feels like it needs real understanding,

00:55:21.620 | how much can that be achieved

00:55:23.500 | with these neural networks or statistical methods?

00:55:27.180 | - So I guess I am very much driven by the outcomes.

00:55:33.060 | Can we achieve the performance,

00:55:35.420 | which would be satisfactory for us for different tasks?

00:55:40.420 | Now, if you again look at machine translation systems,

00:55:43.100 | which are, you know, trained on large amounts of data,

00:55:46.020 | they really can do a remarkable job

00:55:48.780 | relatively to where they've been a few years ago.

00:55:51.340 | And if you project into the future,

00:55:54.580 | if it will be the same speed of improvement,

00:55:56.940 | you know, this is great.

00:56:00.020 | Now, does it bother me that it's not doing

00:56:01.900 | the same translation as we are doing?

00:56:04.820 | Now, if you go to cognitive science,

00:56:06.580 | we still don't really understand what we are doing.

00:56:09.420 | I mean, there are a lot of theories

00:56:11.820 | and there is obviously a lot of progress in studying,

00:56:13.820 | but our understanding what exactly goes on,

00:56:16.380 | you know, in our brains when we process language

00:56:18.820 | is still not crystal clear and precise

00:56:21.740 | that we can translate it into machines.

00:56:25.420 | What does bother me is that, you know, again,

00:56:29.740 | that machines can be extremely brittle

00:56:31.660 | when you go out of your comfort zone

00:56:33.940 | of when there is a distributional shift

00:56:36.020 | between training and testing.

00:56:37.260 | And it have been years and years,

00:56:38.980 | every year when I teach an LP class,

00:56:41.300 | you know, I show them some examples of translation

00:56:43.540 | from some newspaper in Hebrew or whatever, it was perfect.

00:56:47.260 | And then I have a recipe that Tommy Yakala's sister

00:56:51.300 | sent me a while ago and it was written in Finnish

00:56:53.900 | of Carilion pies.

00:56:55.700 | And it's just a terrible translation.

00:56:59.260 | You cannot understand anything, what it does.

00:57:01.460 | It's not like some syntactic mistakes, it's just terrible.

00:57:04.300 | And year after year I try it and it will translate,

00:57:07.020 | and year after year it does this terrible work

00:57:08.980 | because I guess, you know, the recipes

00:57:10.980 | are not a big part of the training repertoire.

00:57:15.460 | - So, but in terms of outcomes,

00:57:18.020 | that's a really clean, good way to look at it.

00:57:21.140 | I guess the question I was asking is,

00:57:23.200 | do you think, imagine a future,

00:57:27.740 | do you think the current approaches

00:57:29.820 | can pass the Turing test in the way,

00:57:32.500 | in the best possible formulation of the Turing test?

00:57:37.060 | Which is, would you wanna have a conversation

00:57:39.500 | with a neural network for an hour?

00:57:42.380 | - Oh God, no.

00:57:44.780 | There are not that many people

00:57:45.820 | that I would wanna talk for an hour.

00:57:48.020 | But-- - There are some people

00:57:49.700 | in this world, alive or not,

00:57:51.500 | that you would like to talk to for an hour.

00:57:53.260 | Could a neural network achieve that outcome?

00:57:56.700 | - So I think it would be really hard

00:57:58.180 | to create a successful training set

00:58:01.100 | which would enable it to have a conversation,

00:58:03.820 | a contextual conversation for an hour.

00:58:06.700 | - So you think it's a problem of data, perhaps?

00:58:08.140 | - I think in some ways it's a problem of data.

00:58:09.940 | It's a problem both of data and the problem

00:58:12.460 | of the way we're training our systems,

00:58:15.660 | their ability to truly to generalize,

00:58:18.060 | to be very compositional.

00:58:19.300 | In some ways it's limited.

00:58:20.460 | You know, in the current capacity, at least,

00:58:24.140 | you know, we can translate well,

00:58:27.980 | we can, you know, find information well,

00:58:31.340 | we can extract information.

00:58:32.540 | So there are many capacities in which it's doing very well.

00:58:35.180 | And you can ask me, would you trust the machine

00:58:38.000 | to translate for you and use it as a source?

00:58:39.780 | I would say absolutely,

00:58:40.820 | especially if we're talking about newspaper data

00:58:43.540 | or other data, which is in the realm of its own training set,

00:58:46.740 | I would say yes.

00:58:47.880 | But, you know, having conversations with the machine,

00:58:52.900 | it's not something that I would choose to do.

00:58:56.460 | But you know, I would tell you something,

00:58:58.140 | talking about Turing tests

00:58:59.420 | and about all this kind of ELISA conversations,

00:59:02.940 | I remember visiting Tencent in China

00:59:05.540 | and they have this chat board

00:59:06.940 | and they claim that it's like really humongous amount

00:59:09.520 | of the local population, which like for hours

00:59:11.820 | talks to the chat board.

00:59:12.900 | To me it was, I cannot believe it,

00:59:15.320 | but apparently it's like documented

00:59:17.100 | that there are some people who enjoy this conversation.

00:59:20.760 | And you know, it brought to me another MIT story

00:59:24.540 | about ELISA and Weizenbaum.

00:59:26.940 | I don't know if you're familiar with this story.

00:59:29.340 | So Weizenbaum was a professor at MIT

00:59:31.020 | and when he developed this ELISA,

00:59:32.580 | which was just doing string matching,

00:59:34.620 | very trivial, like restating of what you said

00:59:38.540 | with very few rules, no syntax.

00:59:41.260 | Apparently there were secretaries at MIT

00:59:43.740 | that would sit for hours and converse

00:59:46.560 | with this trivial thing.

00:59:48.180 | And at the time there was no beautiful interfaces,

00:59:50.180 | so you actually need to go through the pain

00:59:51.820 | of communicating.

00:59:53.540 | And Weizenbaum himself was so horrified by this phenomena

00:59:56.900 | that people can believe enough to the machine

00:59:59.260 | that you just need to give them the hint

01:00:00.820 | that machine understands you and you can complete the rest.

01:00:03.940 | Then he kind of stopped this research

01:00:05.460 | and went into kind of trying to understand

01:00:08.660 | what this artificial intelligence can do to our brains.

01:00:11.460 | So my point is, you know, how much,

01:00:15.400 | it's not how good is the technology,

01:00:19.340 | it's how ready we are to believe

01:00:22.620 | that it delivers the good that we are trying to get.

01:00:25.580 | - That's a really beautiful way to put it.

01:00:27.220 | I, by the way, I'm not horrified by that possibility,

01:00:29.800 | but inspired by it because, I mean,

01:00:34.780 | human connection, whether it's through language

01:00:37.060 | or through love, it seems like it's very amenable

01:00:42.060 | to machine learning and the rest is just

01:00:46.060 | challenges of psychology.

01:00:49.340 | Like you said, the secretaries who enjoy spending hours.

01:00:52.460 | I would say I would describe most of our lives

01:00:55.020 | as enjoying spending hours with those we love

01:00:58.060 | for very silly reasons.

01:01:00.820 | All we're doing is keyword matching as well.

01:01:02.780 | So I'm not sure how much intelligence we exhibit

01:01:05.660 | to each other with the people we love

01:01:08.140 | that we're close with.

01:01:09.820 | So it's a very interesting point

01:01:12.660 | of what it means to pass the Turing test with language.

01:01:16.020 | I think you're right in terms of conversation.

01:01:18.220 | I think machine translation has very clear performance

01:01:23.140 | and improvement, right?

01:01:24.420 | What it means to have a fulfilling conversation

01:01:28.020 | is very, very person dependent and context dependent

01:01:32.660 | and so on.

01:01:33.580 | That's, yeah, it's very well put.

01:01:36.340 | But in your view, what's a benchmark in natural language,

01:01:40.740 | a test that's just out of reach right now,

01:01:43.640 | but we might be able to, that's exciting?

01:01:46.060 | Is it in perfecting machine translation

01:01:49.100 | or is there other, is it summarization?

01:01:51.920 | What's out there?

01:01:52.760 | - I think it goes across specific application.

01:01:55.820 | It's more about the ability to learn from few examples

01:01:59.500 | for real, what we call few short learning

01:02:01.460 | and all these cases, because the way we publish

01:02:04.980 | these papers today, we say, if we have,

01:02:07.540 | like naively we get 55, but now we had a few example

01:02:11.340 | and we can move to 65.

01:02:12.500 | None of these methods actually are realistically

01:02:14.760 | doing anything useful.

01:02:15.980 | You cannot use them today.

01:02:18.540 | And the ability to be able to generalize

01:02:23.540 | and to move or to be autonomous in finding the data

01:02:28.940 | that you need to learn, to be able to perfect new task

01:02:33.180 | or new language, this is an area where I think

01:02:37.220 | we really need to move forward to,

01:02:40.940 | and we are not yet there.

01:02:43.020 | - Are you at all excited, curious by the possibility

01:02:46.540 | of creating human level intelligence?

01:02:48.500 | Is this, 'cause you've been very, in your discussion,

01:02:52.540 | so if we look at oncology, you're trying to use

01:02:56.360 | machine learning to help the world

01:02:58.100 | in terms of alleviating suffering.

01:02:59.660 | If you look at natural language processing,

01:03:02.340 | you're focused on the outcomes of improving

01:03:04.500 | practical things like machine translation.

01:03:06.820 | But human level intelligence is a thing

01:03:09.860 | that our civilization has dreamed about creating,

01:03:13.800 | super human level intelligence.

01:03:15.740 | Do you think about this?

01:03:16.940 | Do you think it's at all within our reach?

01:03:19.040 | - So as you said yourself earlier, talking about,

01:03:25.220 | how do you perceive our communications with each other,

01:03:28.980 | that we're matching keywords and certain behaviors

01:03:31.940 | and so on.

01:03:32.780 | And then whenever one assesses, let's say,

01:03:37.220 | relations with another person, you have separate

01:03:39.900 | kind of measurements and outcomes inside your head

01:03:42.420 | that determine what is the status of the relation.

01:03:45.860 | So one way, this is this classical level,

01:03:48.620 | what is the intelligence?

01:03:49.600 | Is it the fact that now we are gonna do the same way

01:03:51.860 | as human is doing when we don't even understand

01:03:53.980 | what the human is doing?

01:03:55.460 | Or we now have an ability to deliver these outcomes,

01:03:59.100 | but not in one area, not in NLP,

01:04:01.260 | not just to translate or just to answer questions,

01:04:03.940 | but across many, many areas that we can achieve

01:04:06.920 | the functionalities that humans can achieve

01:04:09.740 | with the ability to learn and do other things.

01:04:12.380 | I think this is, and this we can actually measure

01:04:15.500 | how far we are.

01:04:17.560 | And that's what makes me excited that we,

01:04:21.580 | in my lifetime, at least so far what we've seen,

01:04:23.780 | it's like tremendous progress across

01:04:26.260 | with these different functionalities.

01:04:28.740 | And I think it will be really exciting to see

01:04:33.660 | where we will be.

01:04:35.540 | And again, one way to think about it,

01:04:39.340 | there are machines which are improving their functionality.

01:04:41.860 | Another one is to think about us with our brains,

01:04:44.940 | which are imperfect, how they can be accelerated

01:04:49.060 | by this technology as it becomes stronger and stronger.

01:04:54.060 | Coming back to another book that I love,

01:04:58.580 | "Flowers for Algernon."

01:05:00.900 | Have you read this book?

01:05:02.100 | - Yes.

01:05:02.940 | - So there is this point that the patient gets

01:05:05.700 | this miracle cure which changes his brain,

01:05:07.980 | and all of a sudden they see life in a different way

01:05:11.020 | and can do certain things better,

01:05:13.300 | but certain things much worse.

01:05:16.460 | So you can imagine this kind of computer augmented cognition

01:05:21.460 | where it can bring you that now in the same way

01:05:24.780 | as the cars enable us to get to places

01:05:28.100 | where we've never been before,

01:05:30.040 | can we think differently?

01:05:31.580 | Can we think faster?

01:05:33.580 | And we already see a lot of it happening

01:05:36.660 | in how it impacts us,

01:05:38.220 | but I think we have a long way to go there.

01:05:42.180 | - So that's sort of artificial intelligence

01:05:44.980 | and technology augmenting our intelligence as humans.

01:05:49.980 | Yesterday, a company called Neuralink announced,

01:05:55.420 | they did this whole demonstration,

01:05:56.780 | I don't know if you saw it.

01:05:58.900 | They demonstrated brain, computer, brain machine interface,

01:06:02.660 | where there's like a sewing machine for the brain.

01:06:06.100 | A lot of that is quite out there

01:06:11.100 | in terms of things that some people would say

01:06:14.020 | are impossible, but they're dreamers

01:06:16.300 | and want to engineer systems like that.

01:06:18.060 | Do you see, based on what you just said,

01:06:20.340 | a hope for that more direct interaction with the brain?

01:06:23.780 | - I think there are different ways.

01:06:27.020 | One is a direct interaction with the brain,

01:06:28.980 | and again, there are lots of companies

01:06:30.860 | that work in this space,

01:06:32.240 | and I think there will be a lot of developments.

01:06:35.060 | But I'm just thinking that many times

01:06:36.540 | we are not aware of our feelings of motivation

01:06:39.860 | of what drives us.

01:06:41.420 | Like let me give you a trivial example, our attention.

01:06:44.120 | There are a lot of studies that demonstrate

01:06:47.260 | that it takes a while to a person to understand

01:06:49.220 | that they are not attentive anymore.

01:06:51.060 | And we know that there are people

01:06:52.180 | who really have strong capacity to hold attention.

01:06:54.540 | There are another end of the spectrum,

01:06:55.980 | people with ADD and other issues

01:06:57.980 | that they have problem to regulate their attention.

01:07:00.740 | Imagine to yourself that you have like a cognitive aid

01:07:03.540 | that just alerts you based on your gaze,

01:07:06.260 | that your attention is now not on what you are doing,

01:07:09.300 | and instead of writing a paper,

01:07:10.580 | you're now dreaming of what you're gonna do in the evening.

01:07:12.780 | So even this kind of simple measurement things,

01:07:16.420 | how they can change us,

01:07:18.060 | and I see it even in simple ways with myself.

01:07:22.460 | I have my zone up that I got in MIT gym,

01:07:26.540 | it kind of records how much did you run,

01:07:28.820 | and you have some points,

01:07:29.860 | and you can get some status, whatever.

01:07:32.560 | I said, "What is this ridiculous thing?

01:07:35.860 | "Who would ever care about some status in some app?"

01:07:38.880 | Guess what, so to maintain the status,

01:07:41.620 | you have to do set a number of points every month.

01:07:44.700 | And not only is it I do it every single month

01:07:48.100 | for the last 18 months,

01:07:50.620 | it went to the point that I was injured.

01:07:54.220 | And when I could run again,

01:07:56.220 | in two days, I did like some humongous amount of running

01:08:02.500 | just to complete the points.

01:08:04.180 | It was like really not safe.

01:08:05.980 | It's like, I'm not gonna lose my status

01:08:08.500 | because I want to get there.

01:08:10.240 | So you can already see that this direct measurement

01:08:13.320 | and the feedback is,

01:08:14.880 | you know, we're looking at video games

01:08:16.320 | and see why the addiction aspect of it,

01:08:18.680 | but you can imagine that the same idea

01:08:20.460 | can be expanded to many other areas of our life

01:08:23.640 | when we really can get feedback.

01:08:25.960 | And imagine in your case in relations,

01:08:28.480 | when we are doing keyword matching,

01:08:31.200 | imagine that the person who is generating the keywords,

01:08:36.120 | that person gets direct feedback

01:08:37.740 | before the whole thing explodes.

01:08:39.580 | Is it maybe at this happy point,

01:08:41.980 | we are going in the wrong direction.

01:08:44.020 | Maybe it will be really behavior-modifying moment.

01:08:48.020 | - So yeah, it's relationship management too.

01:08:51.340 | So yeah, that's a fascinating whole area

01:08:54.220 | of psychology actually as well,

01:08:56.140 | of seeing how our behavior has changed

01:08:58.260 | with basically all human relations

01:09:00.860 | now have other non-human entities helping us out.

01:09:06.240 | - So you teach a large,

01:09:09.480 | a huge machine learning course here at MIT.

01:09:12.640 | I could ask you a million questions,

01:09:15.360 | but you've seen a lot of students.

01:09:17.600 | What ideas do students struggle with the most

01:09:20.940 | as they first enter this world of machine learning?

01:09:23.940 | - Actually this year was the first time

01:09:28.040 | I started teaching a small machine learning class

01:09:30.120 | and it came as a result of what I saw

01:09:32.880 | in my big machine learning class at Tomiakala

01:09:35.720 | and I built maybe six years ago.

01:09:38.360 | What we've seen that as this area

01:09:41.500 | become more and more popular,

01:09:42.900 | more and more people at MIT want to take this class.

01:09:47.000 | And while we designed it for computer science majors,

01:09:50.000 | there were a lot of people

01:09:51.080 | who really are interested to learn it,

01:09:52.920 | but unfortunately their background

01:09:55.720 | was not enabling them to do well in the class.

01:09:58.820 | And many of them associated machine learning

01:10:01.040 | with the word struggle and failure.

01:10:04.420 | Primarily for non-majors.

01:10:06.460 | And that's why we actually started a new class

01:10:08.720 | which we call machine learning from algorithms to modeling,

01:10:12.700 | which emphasizes more the modeling aspects of it

01:10:16.820 | and focuses on, it has majors and non-majors.

01:10:21.780 | So we kind of try to extract the relevant parts

01:10:25.380 | and make it more accessible

01:10:27.460 | because the fact that we're teaching 20 classifiers

01:10:29.740 | in standard machine learning class

01:10:31.100 | is really a big question we really needed.

01:10:34.180 | - But it was interesting to see this

01:10:36.380 | from first generation of students,

01:10:38.300 | when they came back from their internships

01:10:40.900 | and from their jobs,

01:10:43.940 | what different and exciting things they can do

01:10:47.460 | that I would never think

01:10:48.300 | that you can even apply machine learning to.

01:10:51.100 | Some of them are like matching their relations

01:10:53.740 | and other things like variety of different applications.

01:10:55.780 | - Everything is amenable to machine learning.

01:10:58.020 | That actually brings up an interesting point

01:11:00.260 | of computer science in general.

01:11:02.580 | It almost seems, maybe I'm crazy,

01:11:05.420 | but it almost seems like everybody needs to learn

01:11:08.420 | how to program these days.

01:11:10.060 | If you're 20 years old or if you're starting school,

01:11:13.340 | even if you're an English major,

01:11:15.900 | it seems like programming unlocks

01:11:19.500 | so much possibility in this world.

01:11:21.860 | So when you interacted with those non-majors,

01:11:24.980 | is there skills that they were simply lacking at the time

01:11:30.260 | that you wish they had

01:11:32.020 | and that they learned in high school and so on?

01:11:34.660 | Like how should education change

01:11:37.460 | in this computerized world that we live in?

01:11:41.260 | - Same because I knew

01:11:42.100 | that there is a Python component in the class.

01:11:44.780 | Their Python skills were okay

01:11:47.020 | and the class is not really heavy on programming.

01:11:49.140 | They primarily kind of add parts to the programs.

01:11:52.420 | I think it was more of the mathematical barriers

01:11:55.420 | and the class, again, with a design on the majors

01:11:58.220 | was using the notation like big O for complexity

01:12:01.180 | and others, people who come from different backgrounds

01:12:04.540 | just don't have it in the lexical.

01:12:05.780 | It's not necessarily very challenging notion,

01:12:09.100 | but they were just not aware.

01:12:11.460 | So I think that kind of linear algebra and probability,

01:12:16.220 | the basics, the calculus,

01:12:17.620 | multivariate calculus are things that can help.

01:12:20.820 | - What advice would you give to students

01:12:23.540 | interested in machine learning,

01:12:25.260 | interested, you've talked about

01:12:28.460 | detecting, curing cancer, drug design.

01:12:31.380 | If they want to get into that field, what should they do?

01:12:34.520 | Get into it and succeed as researchers and entrepreneurs.

01:12:40.620 | - The first good piece of news is right now

01:12:45.220 | there are lots of resources

01:12:47.380 | that are created at different levels

01:12:50.140 | and you can find online or in your school classes

01:12:54.780 | which are more mathematical, more applied and so on.

01:12:57.540 | So you can find a kind of a preacher

01:13:01.300 | which preaches your own language

01:13:02.740 | where you can enter the field

01:13:04.500 | and you can make many different types of contribution

01:13:06.700 | depending of what is your strengths.

01:13:09.580 | And the second point, I think it's really important

01:13:13.660 | to find some area which you really care about

01:13:18.100 | and it can motivate your learning.

01:13:20.180 | And it can be for somebody curing cancer

01:13:22.540 | or doing self-driving cars or whatever

01:13:25.340 | but to find an area where there is data,

01:13:29.620 | where you believe there are strong patterns

01:13:31.300 | and we should be doing it and we're still not doing it

01:13:33.540 | or you can do it better and just start there

01:13:37.820 | and see where it can bring you.

01:13:39.660 | - So you've been very successful in many directions in life

01:13:45.460 | but you also mentioned "Flowers of Argonaut"

01:13:51.140 | and I think I've read or listened to you

01:13:53.100 | mention somewhere that researchers often get lost

01:13:55.300 | in the details of their work.

01:13:56.680 | This is per our original discussion with cancer and so on

01:14:00.180 | and don't look at the bigger picture,

01:14:02.140 | bigger questions of meaning and so on.

01:14:04.280 | So let me ask you the impossible question

01:14:07.380 | of what's the meaning of this thing,

01:14:11.540 | of life, of your life, of research.

01:14:16.660 | Why do you think we descendant of great apes

01:14:21.420 | are here on this spinning ball?

01:14:24.460 | - You know, I don't think that I have really

01:14:29.100 | a global answer, you know, maybe that's why

01:14:31.140 | I didn't go to humanities

01:14:32.780 | and I didn't take humanities classes in my undergrad.

01:14:36.440 | But the way I'm thinking about it,

01:14:43.540 | each one of us inside of them have their own set of,

01:14:48.220 | you know, things that we believe are important

01:14:51.140 | and it just happens that we are busy

01:14:53.380 | with achieving various goal, busy listening to others

01:14:56.260 | and to kind of try to conform and to be part of the crowd

01:14:59.500 | that we don't listen to that part.

01:15:03.740 | And, you know, we all should find some time to understand

01:15:09.620 | what is our own individual missions

01:15:11.860 | and we may have very different missions

01:15:14.100 | and to make sure that while we are running 10,000 things,

01:15:18.220 | we are not, you know, missing out

01:15:21.940 | and we're putting all the resources

01:15:24.420 | to satisfy our own mission.

01:15:28.500 | And if I look over my time, when I was younger,

01:15:32.460 | most of these missions, you know,

01:15:35.060 | I was primarily driven by the external stimulus,

01:15:38.620 | you know, to achieve this or to be that.

01:15:41.540 | And now a lot of what I do is driven by really thinking

01:15:46.540 | what is important for me to achieve independently

01:15:51.340 | of the external recognition.

01:15:55.140 | And, you know, I don't mind to be viewed in certain ways.

01:16:00.100 | The most important thing for me is to be true to myself,

01:16:05.740 | to what I think is right.

01:16:07.500 | - How long did it take, how hard was it

01:16:09.900 | to find the you that you have to be true to?

01:16:13.220 | - So it takes time and even now,

01:16:16.780 | sometimes, you know, the vanity and the triviality

01:16:19.420 | can take, you know. - At MIT.

01:16:21.660 | - Yeah, it can everywhere, you know,

01:16:25.100 | it's just the vanity at MIT is different,

01:16:26.940 | the vanity in different places,

01:16:28.140 | but we all have our piece of vanity.

01:16:30.940 | But I think actually, for me, many times,

01:16:37.820 | the place to get back to it is, you know,

01:16:41.700 | when I'm alone and also when I read.

01:16:45.820 | And I think by selecting the right books,

01:16:47.740 | you can get the right questions

01:16:49.940 | and learn from what you read.

01:16:53.700 | So, but again, it's not perfect, like,

01:16:58.460 | vanity sometimes dominates. - Nothing is.

01:17:02.020 | Well, that's a beautiful way to end.

01:17:04.780 | Thank you so much for talking today.

01:17:06.340 | - Thank you. - That was fun.

01:17:07.860 | - Oh, it was fun.

01:17:08.700 | (upbeat music)

01:17:11.300 | (upbeat music)

01:17:13.900 | (upbeat music)

01:17:16.500 | (upbeat music)

01:17:19.100 | (upbeat music)

01:17:21.700 | (upbeat music)

01:17:24.300 | [BLANK_AUDIO]

Regina Barzilay: Deep Learning for Cancer Diagnosis and Treatment | Lex Fridman Podcast #40

Chapters