vtt/episode_040_large.vtt

WEBVTT

00:00.000 --> 00:03.220
 The following is a conversation with Regina Barzilay.

00:03.220 --> 00:06.700
 She's a professor at MIT and a world class researcher

00:06.700 --> 00:08.340
 in natural language processing

00:08.340 --> 00:12.460
 and applications of deep learning to chemistry and oncology

00:12.460 --> 00:15.340
 or the use of deep learning for early diagnosis,

00:15.340 --> 00:18.300
 prevention and treatment of cancer.

00:18.300 --> 00:21.020
 She has also been recognized for teaching

00:21.020 --> 00:24.700
 of several successful AI related courses at MIT,

00:24.700 --> 00:26.840
 including the popular Introduction

00:26.840 --> 00:28.920
 to Machine Learning course.

00:28.920 --> 00:32.160
 This is the Artificial Intelligence podcast.

00:32.160 --> 00:34.560
 If you enjoy it, subscribe on YouTube,

00:34.560 --> 00:37.840
 give it five stars on iTunes, support it on Patreon

00:37.840 --> 00:39.840
 or simply connect with me on Twitter

00:39.840 --> 00:43.760
 at Lex Friedman spelled F R I D M A N.

00:43.760 --> 00:47.760
 And now here's my conversation with Regina Barzilay.

00:48.840 --> 00:50.320
 In an interview you've mentioned

00:50.320 --> 00:51.960
 that if there's one course you would take,

00:51.960 --> 00:54.600
 it would be a literature course with a friend of yours

00:54.600 --> 00:56.360
 that a friend of yours teaches.

00:56.360 --> 00:59.160
 Just out of curiosity, because I couldn't find anything

00:59.160 --> 01:04.160
 on it, are there books or ideas that had profound impact

01:04.400 --> 01:07.200
 on your life journey, books and ideas perhaps

01:07.200 --> 01:10.800
 outside of computer science and the technical fields?

01:11.780 --> 01:14.680
 I think because I'm spending a lot of my time at MIT

01:14.680 --> 01:18.280
 and previously in other institutions where I was a student,

01:18.280 --> 01:21.040
 I have limited ability to interact with people.

01:21.040 --> 01:22.640
 So a lot of what I know about the world

01:22.640 --> 01:24.220
 actually comes from books.

01:24.220 --> 01:27.240
 And there were quite a number of books

01:27.240 --> 01:31.380
 that had profound impact on me and how I view the world.

01:31.380 --> 01:35.820
 Let me just give you one example of such a book.

01:35.820 --> 01:39.660
 I've maybe a year ago read a book

01:39.660 --> 01:42.500
 called The Emperor of All Melodies.

01:42.500 --> 01:45.740
 It's a book about, it's kind of a history of science book

01:45.740 --> 01:50.740
 on how the treatments and drugs for cancer were developed.

01:50.740 --> 01:54.580
 And that book, despite the fact that I am in the business

01:54.580 --> 01:59.580
 of science, really opened my eyes on how imprecise

01:59.780 --> 02:03.060
 and imperfect the discovery process is

02:03.060 --> 02:05.820
 and how imperfect our current solutions

02:06.980 --> 02:11.060
 and what makes science succeed and be implemented.

02:11.060 --> 02:14.100
 And sometimes it's actually not the strengths of the idea,

02:14.100 --> 02:17.420
 but devotion of the person who wants to see it implemented.

02:17.420 --> 02:19.780
 So this is one of the books that, you know,

02:19.780 --> 02:22.300
 at least for the last year, quite changed the way

02:22.300 --> 02:24.940
 I'm thinking about scientific process

02:24.940 --> 02:26.700
 just from the historical perspective

02:26.700 --> 02:31.700
 and what do I need to do to make my ideas really implemented.

02:33.460 --> 02:36.060
 Let me give you an example of a book

02:36.060 --> 02:39.580
 which is not kind of, which is a fiction book.

02:40.620 --> 02:43.100
 It's a book called Americana.

02:44.420 --> 02:48.780
 And this is a book about a young female student

02:48.780 --> 02:53.260
 who comes from Africa to study in the United States.

02:53.260 --> 02:57.740
 And it describes her past, you know, within her studies

02:57.740 --> 03:02.020
 and her life transformation that, you know,

03:02.020 --> 03:06.540
 in a new country and kind of adaptation to a new culture.

03:06.540 --> 03:11.220
 And when I read this book, I saw myself

03:11.220 --> 03:13.540
 in many different points of it,

03:13.540 --> 03:18.540
 but it also kind of gave me the lens on different events.

03:20.140 --> 03:22.060
 And some of it that I never actually paid attention.

03:22.060 --> 03:24.700
 One of the funny stories in this book

03:24.700 --> 03:29.700
 is how she arrives to her new college

03:30.420 --> 03:32.900
 and she starts speaking in English

03:32.900 --> 03:35.700
 and she had this beautiful British accent

03:35.700 --> 03:39.860
 because that's how she was educated in her country.

03:39.860 --> 03:40.980
 This is not my case.

03:40.980 --> 03:45.460
 And then she notices that the person who talks to her,

03:45.460 --> 03:47.220
 you know, talks to her in a very funny way,

03:47.220 --> 03:48.340
 in a very slow way.

03:48.340 --> 03:51.460
 And she's thinking that this woman is disabled

03:51.460 --> 03:54.500
 and she's also trying to kind of to accommodate her.

03:54.500 --> 03:56.700
 And then after a while, when she finishes her discussion

03:56.700 --> 03:58.580
 with this officer from her college,

03:59.860 --> 04:02.100
 she sees how she interacts with the other students,

04:02.100 --> 04:03.020
 with American students.

04:03.020 --> 04:08.020
 And she discovers that actually she talked to her this way

04:08.020 --> 04:11.020
 because she saw that she doesn't understand English.

04:11.020 --> 04:14.180
 And I thought, wow, this is a funny experience.

04:14.180 --> 04:16.940
 And literally within few weeks,

04:16.940 --> 04:20.820
 I went to LA to a conference

04:20.820 --> 04:23.180
 and I asked somebody in the airport,

04:23.180 --> 04:25.580
 you know, how to find like a cab or something.

04:25.580 --> 04:28.380
 And then I noticed that this person is talking

04:28.380 --> 04:29.220
 in a very strange way.

04:29.220 --> 04:31.100
 And my first thought was that this person

04:31.100 --> 04:34.500
 have some, you know, pronunciation issues or something.

04:34.500 --> 04:36.060
 And I'm trying to talk very slowly to him

04:36.060 --> 04:38.580
 and I was with another professor, Ernst Frankel.

04:38.580 --> 04:42.180
 And he's like laughing because it's funny

04:42.180 --> 04:44.860
 that I don't get that the guy is talking in this way

04:44.860 --> 04:46.060
 because he thinks that I cannot speak.

04:46.060 --> 04:49.100
 So it was really kind of mirroring experience.

04:49.100 --> 04:53.300
 And it led me think a lot about my own experiences

04:53.300 --> 04:56.060
 moving, you know, from different countries.

04:56.060 --> 04:59.300
 So I think that books play a big role

04:59.300 --> 05:01.780
 in my understanding of the world.

05:01.780 --> 05:06.420
 On the science question, you mentioned that

05:06.420 --> 05:09.780
 it made you discover that personalities of human beings

05:09.780 --> 05:12.420
 are more important than perhaps ideas.

05:12.420 --> 05:13.660
 Is that what I heard?

05:13.660 --> 05:15.740
 It's not necessarily that they are more important

05:15.740 --> 05:19.180
 than ideas, but I think that ideas on their own

05:19.180 --> 05:20.460
 are not sufficient.

05:20.460 --> 05:24.660
 And many times, at least at the local horizon,

05:24.660 --> 05:29.140
 it's the personalities and their devotion to their ideas

05:29.140 --> 05:32.980
 is really that locally changes the landscape.

05:32.980 --> 05:37.500
 Now, if you're looking at AI, like let's say 30 years ago,

05:37.500 --> 05:39.180
 you know, dark ages of AI or whatever,

05:39.180 --> 05:42.420
 what is symbolic times, you can use any word.

05:42.420 --> 05:44.660
 You know, there were some people,

05:44.660 --> 05:46.620
 now we're looking at a lot of that work

05:46.620 --> 05:48.780
 and we're kind of thinking this was not really

05:48.780 --> 05:52.220
 maybe a relevant work, but you can see that some people

05:52.220 --> 05:54.900
 managed to take it and to make it so shiny

05:54.900 --> 05:59.260
 and dominate the academic world

05:59.260 --> 06:02.380
 and make it to be the standard.

06:02.380 --> 06:05.180
 If you look at the area of natural language processing,

06:06.420 --> 06:09.140
 it is well known fact that the reason that statistics

06:09.140 --> 06:13.980
 in NLP took such a long time to become mainstream

06:13.980 --> 06:16.860
 because there were quite a number of personalities

06:16.860 --> 06:18.460
 which didn't believe in this idea

06:18.460 --> 06:22.060
 and didn't stop research progress in this area.

06:22.060 --> 06:25.900
 So I do not think that, you know,

06:25.900 --> 06:28.940
 kind of asymptotically maybe personalities matters,

06:28.940 --> 06:33.940
 but I think locally it does make quite a bit of impact

06:33.940 --> 06:36.900
 and it's generally, you know,

06:36.900 --> 06:41.340
 speeds up the rate of adoption of the new ideas.

06:41.340 --> 06:43.500
 Yeah, and the other interesting question

06:43.500 --> 06:46.540
 is in the early days of particular discipline,

06:46.540 --> 06:50.460
 I think you mentioned in that book

06:50.460 --> 06:52.340
 is ultimately a book of cancer.

06:52.340 --> 06:55.100
 It's called The Emperor of All Melodies.

06:55.100 --> 06:58.580
 Yeah, and those melodies included the trying to,

06:58.580 --> 07:00.740
 the medicine, was it centered around?

07:00.740 --> 07:04.900
 So it was actually centered on, you know,

07:04.900 --> 07:07.180
 how people thought of curing cancer.

07:07.180 --> 07:10.660
 Like for me, it was really a discovery how people,

07:10.660 --> 07:14.140
 what was the science of chemistry behind drug development

07:14.140 --> 07:17.220
 that it actually grew up out of dying,

07:17.220 --> 07:19.780
 like coloring industry that people

07:19.780 --> 07:23.780
 who developed chemistry in 19th century in Germany

07:23.780 --> 07:28.140
 and Britain to do, you know, the really new dyes.

07:28.140 --> 07:30.180
 They looked at the molecules and identified

07:30.180 --> 07:32.140
 that they do certain things to cells.

07:32.140 --> 07:34.500
 And from there, the process started.

07:34.500 --> 07:35.740
 And, you know, like historically saying,

07:35.740 --> 07:36.900
 yeah, this is fascinating

07:36.900 --> 07:38.700
 that they managed to make the connection

07:38.700 --> 07:42.300
 and look under the microscope and do all this discovery.

07:42.300 --> 07:44.340
 But as you continue reading about it

07:44.340 --> 07:48.780
 and you read about how chemotherapy drugs

07:48.780 --> 07:50.500
 which were developed in Boston,

07:50.500 --> 07:52.500
 and some of them were developed.

07:52.500 --> 07:57.500
 And Farber, Dr. Farber from Dana Farber,

07:57.500 --> 08:00.460
 you know, how the experiments were done

08:00.460 --> 08:03.340
 that, you know, there was some miscalculation,

08:03.340 --> 08:04.540
 let's put it this way.

08:04.540 --> 08:06.740
 And they tried it on the patients and they just,

08:06.740 --> 08:09.980
 and those were children with leukemia and they died.

08:09.980 --> 08:11.660
 And then they tried another modification.

08:11.660 --> 08:15.020
 You look at the process, how imperfect is this process?

08:15.020 --> 08:17.500
 And, you know, like, if we're again looking back

08:17.500 --> 08:19.180
 like 60 years ago, 70 years ago,

08:19.180 --> 08:20.780
 you can kind of understand it.

08:20.780 --> 08:23.020
 But some of the stories in this book

08:23.020 --> 08:24.620
 which were really shocking to me

08:24.620 --> 08:27.980
 were really happening, you know, maybe decades ago.

08:27.980 --> 08:30.660
 And we still don't have a vehicle

08:30.660 --> 08:35.100
 to do it much more fast and effective and, you know,

08:35.100 --> 08:38.220
 scientific the way I'm thinking computer science scientific.

08:38.220 --> 08:40.420
 So from the perspective of computer science,

08:40.420 --> 08:43.780
 you've gotten a chance to work the application to cancer

08:43.780 --> 08:44.860
 and to medicine in general.

08:44.860 --> 08:48.420
 From a perspective of an engineer and a computer scientist,

08:48.420 --> 08:51.780
 how far along are we from understanding the human body,

08:51.780 --> 08:55.140
 biology of being able to manipulate it

08:55.140 --> 08:57.940
 in a way we can cure some of the maladies,

08:57.940 --> 08:59.740
 some of the diseases?

08:59.740 --> 09:02.220
 So this is very interesting question.

09:03.460 --> 09:06.020
 And if you're thinking as a computer scientist

09:06.020 --> 09:09.820
 about this problem, I think one of the reasons

09:09.820 --> 09:11.900
 that we succeeded in the areas

09:11.900 --> 09:13.980
 we as a computer scientist succeeded

09:13.980 --> 09:16.260
 is because we don't have,

09:16.260 --> 09:18.980
 we are not trying to understand in some ways.

09:18.980 --> 09:22.260
 Like if you're thinking about like eCommerce, Amazon,

09:22.260 --> 09:24.220
 Amazon doesn't really understand you.

09:24.220 --> 09:27.700
 And that's why it recommends you certain books

09:27.700 --> 09:29.580
 or certain products, correct?

09:30.660 --> 09:34.660
 And, you know, traditionally when people

09:34.660 --> 09:36.380
 were thinking about marketing, you know,

09:36.380 --> 09:39.780
 they divided the population to different kind of subgroups,

09:39.780 --> 09:41.740
 identify the features of this subgroup

09:41.740 --> 09:43.140
 and come up with a strategy

09:43.140 --> 09:45.580
 which is specific to that subgroup.

09:45.580 --> 09:47.340
 If you're looking about recommendation system,

09:47.340 --> 09:50.580
 they're not claiming that they're understanding somebody,

09:50.580 --> 09:52.700
 they're just managing to,

09:52.700 --> 09:54.780
 from the patterns of your behavior

09:54.780 --> 09:57.540
 to recommend you a product.

09:57.540 --> 09:59.580
 Now, if you look at the traditional biology,

09:59.580 --> 10:03.180
 and obviously I wouldn't say that I

10:03.180 --> 10:06.180
 at any way, you know, educated in this field,

10:06.180 --> 10:09.300
 but you know what I see, there's really a lot of emphasis

10:09.300 --> 10:10.660
 on mechanistic understanding.

10:10.660 --> 10:12.540
 And it was very surprising to me

10:12.540 --> 10:13.820
 coming from computer science,

10:13.820 --> 10:17.580
 how much emphasis is on this understanding.

10:17.580 --> 10:20.740
 And given the complexity of the system,

10:20.740 --> 10:23.220
 maybe the deterministic full understanding

10:23.220 --> 10:27.380
 of this process is, you know, beyond our capacity.

10:27.380 --> 10:29.460
 And the same ways in computer science

10:29.460 --> 10:31.540
 when we're doing recognition, when you do recommendation

10:31.540 --> 10:32.780
 and many other areas,

10:32.780 --> 10:35.940
 it's just probabilistic matching process.

10:35.940 --> 10:40.100
 And in some way, maybe in certain cases,

10:40.100 --> 10:42.940
 we shouldn't even attempt to understand

10:42.940 --> 10:45.780
 or we can attempt to understand, but in parallel,

10:45.780 --> 10:48.060
 we can actually do this kind of matchings

10:48.060 --> 10:51.060
 that would help us to find key role

10:51.060 --> 10:54.100
 to do early diagnostics and so on.

10:54.100 --> 10:55.860
 And I know that in these communities,

10:55.860 --> 10:59.060
 it's really important to understand,

10:59.060 --> 11:00.700
 but I'm sometimes wondering, you know,

11:00.700 --> 11:02.940
 what exactly does it mean to understand here?

11:02.940 --> 11:05.500
 Well, there's stuff that works and,

11:05.500 --> 11:07.620
 but that can be, like you said,

11:07.620 --> 11:10.340
 separate from this deep human desire

11:10.340 --> 11:12.700
 to uncover the mysteries of the universe,

11:12.700 --> 11:16.140
 of science, of the way the body works,

11:16.140 --> 11:17.620
 the way the mind works.

11:17.620 --> 11:19.540
 It's the dream of symbolic AI,

11:19.540 --> 11:24.540
 of being able to reduce human knowledge into logic

11:25.220 --> 11:26.900
 and be able to play with that logic

11:26.900 --> 11:28.700
 in a way that's very explainable

11:28.700 --> 11:30.300
 and understandable for us humans.

11:30.300 --> 11:31.780
 I mean, that's a beautiful dream.

11:31.780 --> 11:34.860
 So I understand it, but it seems that

11:34.860 --> 11:37.900
 what seems to work today and we'll talk about it more

11:37.900 --> 11:40.780
 is as much as possible, reduce stuff into data,

11:40.780 --> 11:43.900
 reduce whatever problem you're interested in to data

11:43.900 --> 11:47.060
 and try to apply statistical methods,

11:47.060 --> 11:49.100
 apply machine learning to that.

11:49.100 --> 11:51.140
 On a personal note,

11:51.140 --> 11:54.140
 you were diagnosed with breast cancer in 2014.

11:55.380 --> 11:58.420
 What did facing your mortality make you think about?

11:58.420 --> 12:00.260
 How did it change you?

12:00.260 --> 12:01.860
 You know, this is a great question

12:01.860 --> 12:03.820
 and I think that I was interviewed many times

12:03.820 --> 12:05.740
 and nobody actually asked me this question.

12:05.740 --> 12:09.700
 I think I was 43 at a time.

12:09.700 --> 12:12.860
 And the first time I realized in my life that I may die

12:12.860 --> 12:14.460
 and I never thought about it before.

12:14.460 --> 12:17.260
 And there was a long time since you're diagnosed

12:17.260 --> 12:18.580
 until you actually know what you have

12:18.580 --> 12:20.180
 and how severe is your disease.

12:20.180 --> 12:23.500
 For me, it was like maybe two and a half months.

12:23.500 --> 12:28.340
 And I didn't know where I am during this time

12:28.340 --> 12:30.660
 because I was getting different tests

12:30.660 --> 12:33.380
 and one would say it's bad and I would say, no, it is not.

12:33.380 --> 12:34.900
 So until I knew where I am,

12:34.900 --> 12:36.300
 I really was thinking about

12:36.300 --> 12:38.220
 all these different possible outcomes.

12:38.220 --> 12:39.700
 Were you imagining the worst

12:39.700 --> 12:41.940
 or were you trying to be optimistic or?

12:41.940 --> 12:43.540
 It would be really,

12:43.540 --> 12:47.340
 I don't remember what was my thinking.

12:47.340 --> 12:51.100
 It was really a mixture with many components at the time

12:51.100 --> 12:54.100
 speaking in our terms.

12:54.100 --> 12:59.100
 And one thing that I remember,

12:59.340 --> 13:01.500
 and every test comes and then you're saying,

13:01.500 --> 13:03.300
 oh, it could be this or it may not be this.

13:03.300 --> 13:04.700
 And you're hopeful and then you're desperate.

13:04.700 --> 13:07.660
 So it's like, there is a whole slew of emotions

13:07.660 --> 13:08.700
 that goes through you.

13:09.820 --> 13:14.820
 But what I remember is that when I came back to MIT,

13:15.100 --> 13:17.780
 I was kind of going the whole time through the treatment

13:17.780 --> 13:19.780
 to MIT, but my brain was not really there.

13:19.780 --> 13:21.820
 But when I came back, really finished my treatment

13:21.820 --> 13:23.860
 and I was here teaching and everything,

13:24.900 --> 13:27.060
 I look back at what my group was doing,

13:27.060 --> 13:28.820
 what other groups was doing.

13:28.820 --> 13:30.820
 And I saw these trivialities.

13:30.820 --> 13:33.260
 It's like people are building their careers

13:33.260 --> 13:36.900
 on improving some parts around two or 3% or whatever.

13:36.900 --> 13:38.380
 I was, it's like, seriously,

13:38.380 --> 13:40.740
 I did a work on how to decipher ugaritic,

13:40.740 --> 13:42.860
 like a language that nobody speak and whatever,

13:42.860 --> 13:46.140
 like what is significance?

13:46.140 --> 13:49.020
 When all of a sudden, I walked out of MIT,

13:49.020 --> 13:51.860
 which is when people really do care

13:51.860 --> 13:54.500
 what happened to your ICLR paper,

13:54.500 --> 13:57.900
 what is your next publication to ACL,

13:57.900 --> 14:01.860
 to the world where people, you see a lot of suffering

14:01.860 --> 14:04.900
 that I'm kind of totally shielded on it on daily basis.

14:04.900 --> 14:07.460
 And it's like the first time I've seen like real life

14:07.460 --> 14:08.660
 and real suffering.

14:09.700 --> 14:13.260
 And I was thinking, why are we trying to improve the parser

14:13.260 --> 14:18.260
 or deal with trivialities when we have capacity

14:18.340 --> 14:20.700
 to really make a change?

14:20.700 --> 14:24.620
 And it was really challenging to me because on one hand,

14:24.620 --> 14:27.420
 I have my graduate students really want to do their papers

14:27.420 --> 14:29.860
 and their work, and they want to continue to do

14:29.860 --> 14:31.900
 what they were doing, which was great.

14:31.900 --> 14:36.300
 And then it was me who really kind of reevaluated

14:36.300 --> 14:37.460
 what is the importance.

14:37.460 --> 14:40.260
 And also at that point, because I had to take some break,

14:42.500 --> 14:47.500
 I look back into like my years in science

14:47.740 --> 14:50.460
 and I was thinking, like 10 years ago,

14:50.460 --> 14:52.940
 this was the biggest thing, I don't know, topic models.

14:52.940 --> 14:55.340
 We have like millions of papers on topic models

14:55.340 --> 14:56.500
 and variation of topics models.

14:56.500 --> 14:58.580
 Now it's totally like irrelevant.

14:58.580 --> 15:02.460
 And you start looking at this, what do you perceive

15:02.460 --> 15:04.500
 as important at different point of time

15:04.500 --> 15:08.900
 and how it fades over time.

15:08.900 --> 15:12.980
 And since we have a limited time,

15:12.980 --> 15:14.900
 all of us have limited time on us,

15:14.900 --> 15:18.380
 it's really important to prioritize things

15:18.380 --> 15:20.540
 that really matter to you, maybe matter to you

15:20.540 --> 15:22.020
 at that particular point.

15:22.020 --> 15:24.380
 But it's important to take some time

15:24.380 --> 15:26.940
 and understand what matters to you,

15:26.940 --> 15:28.860
 which may not necessarily be the same

15:28.860 --> 15:31.700
 as what matters to the rest of your scientific community

15:31.700 --> 15:34.580
 and pursue that vision.

15:34.580 --> 15:38.460
 So that moment, did it make you cognizant?

15:38.460 --> 15:42.500
 You mentioned suffering of just the general amount

15:42.500 --> 15:44.340
 of suffering in the world.

15:44.340 --> 15:45.620
 Is that what you're referring to?

15:45.620 --> 15:47.420
 So as opposed to topic models

15:47.420 --> 15:50.780
 and specific detailed problems in NLP,

15:50.780 --> 15:54.460
 did you start to think about other people

15:54.460 --> 15:56.940
 who have been diagnosed with cancer?

15:56.940 --> 16:00.020
 Is that the way you started to see the world perhaps?

16:00.020 --> 16:00.860
 Oh, absolutely.

16:00.860 --> 16:04.980
 And it actually creates, because like, for instance,

16:04.980 --> 16:05.820
 there is parts of the treatment

16:05.820 --> 16:08.500
 where you need to go to the hospital every day

16:08.500 --> 16:11.620
 and you see the community of people that you see

16:11.620 --> 16:16.100
 and many of them are much worse than I was at a time.

16:16.100 --> 16:20.460
 And you all of a sudden see it all.

16:20.460 --> 16:23.940
 And people who are happier someday

16:23.940 --> 16:25.300
 just because they feel better.

16:25.300 --> 16:28.500
 And for people who are in our normal realm,

16:28.500 --> 16:30.820
 you take it totally for granted that you feel well,

16:30.820 --> 16:32.940
 that if you decide to go running, you can go running

16:32.940 --> 16:35.900
 and you're pretty much free

16:35.900 --> 16:37.620
 to do whatever you want with your body.

16:37.620 --> 16:40.180
 Like I saw like a community,

16:40.180 --> 16:42.820
 my community became those people.

16:42.820 --> 16:47.460
 And I remember one of my friends, Dina Katabi,

16:47.460 --> 16:50.420
 took me to Prudential to buy me a gift for my birthday.

16:50.420 --> 16:52.340
 And it was like the first time in months

16:52.340 --> 16:54.980
 that I went to kind of to see other people.

16:54.980 --> 16:58.180
 And I was like, wow, first of all, these people,

16:58.180 --> 16:59.820
 they are happy and they're laughing

16:59.820 --> 17:02.620
 and they're very different from these other my people.

17:02.620 --> 17:04.620
 And second of thing, I think it's totally crazy.

17:04.620 --> 17:06.620
 They're like laughing and wasting their money

17:06.620 --> 17:08.420
 on some stupid gifts.

17:08.420 --> 17:12.540
 And they may die.

17:12.540 --> 17:15.940
 They already may have cancer and they don't understand it.

17:15.940 --> 17:20.060
 So you can really see how the mind changes

17:20.060 --> 17:22.340
 that you can see that,

17:22.340 --> 17:23.180
 before that you can ask,

17:23.180 --> 17:24.380
 didn't you know that you're gonna die?

17:24.380 --> 17:28.340
 Of course I knew, but it was a kind of a theoretical notion.

17:28.340 --> 17:31.060
 It wasn't something which was concrete.

17:31.060 --> 17:33.900
 And at that point, when you really see it

17:33.900 --> 17:38.060
 and see how little means sometimes the system has

17:38.060 --> 17:41.740
 to have them, you really feel that we need to take a lot

17:41.740 --> 17:45.420
 of our brilliance that we have here at MIT

17:45.420 --> 17:48.020
 and translate it into something useful.

17:48.020 --> 17:50.540
 Yeah, and you still couldn't have a lot of definitions,

17:50.540 --> 17:53.620
 but of course, alleviating, suffering, alleviating,

17:53.620 --> 17:57.460
 trying to cure cancer is a beautiful mission.

17:57.460 --> 18:01.940
 So I of course know theoretically the notion of cancer,

18:01.940 --> 18:06.940
 but just reading more and more about it's 1.7 million

18:07.100 --> 18:09.860
 new cancer cases in the United States every year,

18:09.860 --> 18:13.460
 600,000 cancer related deaths every year.

18:13.460 --> 18:18.460
 So this has a huge impact, United States globally.

18:19.340 --> 18:24.340
 When broadly, before we talk about how machine learning,

18:24.340 --> 18:27.180
 how MIT can help,

18:27.180 --> 18:32.100
 when do you think we as a civilization will cure cancer?

18:32.100 --> 18:34.980
 How hard of a problem is it from everything you've learned

18:34.980 --> 18:35.940
 from it recently?

18:37.260 --> 18:39.300
 I cannot really assess it.

18:39.300 --> 18:42.100
 What I do believe will happen with the advancement

18:42.100 --> 18:45.940
 in machine learning is that a lot of types of cancer

18:45.940 --> 18:48.500
 we will be able to predict way early

18:48.500 --> 18:53.420
 and more effectively utilize existing treatments.

18:53.420 --> 18:57.540
 I think, I hope at least that with all the advancements

18:57.540 --> 19:01.180
 in AI and drug discovery, we would be able

19:01.180 --> 19:04.700
 to much faster find relevant molecules.

19:04.700 --> 19:08.220
 What I'm not sure about is how long it will take

19:08.220 --> 19:11.940
 the medical establishment and regulatory bodies

19:11.940 --> 19:14.780
 to kind of catch up and to implement it.

19:14.780 --> 19:17.420
 And I think this is a very big piece of puzzle

19:17.420 --> 19:20.420
 that is currently not addressed.

19:20.420 --> 19:21.780
 That's the really interesting question.

19:21.780 --> 19:25.460
 So first a small detail that I think the answer is yes,

19:25.460 --> 19:30.460
 but is cancer one of the diseases that when detected earlier

19:33.700 --> 19:37.820
 that's a significantly improves the outcomes?

19:37.820 --> 19:41.020
 So like, cause we will talk about there's the cure

19:41.020 --> 19:43.020
 and then there is detection.

19:43.020 --> 19:45.180
 And I think where machine learning can really help

19:45.180 --> 19:46.660
 is earlier detection.

19:46.660 --> 19:48.580
 So does detection help?

19:48.580 --> 19:49.660
 Detection is crucial.

19:49.660 --> 19:53.940
 For instance, the vast majority of pancreatic cancer patients

19:53.940 --> 19:57.300
 are detected at the stage that they are incurable.

19:57.300 --> 20:02.300
 That's why they have such a terrible survival rate.

20:03.740 --> 20:07.300
 It's like just few percent over five years.

20:07.300 --> 20:09.820
 It's pretty much today the sentence.

20:09.820 --> 20:13.620
 But if you can discover this disease early,

20:14.500 --> 20:16.740
 there are mechanisms to treat it.

20:16.740 --> 20:20.740
 And in fact, I know a number of people who were diagnosed

20:20.740 --> 20:23.580
 and saved just because they had food poisoning.

20:23.580 --> 20:25.020
 They had terrible food poisoning.

20:25.020 --> 20:28.540
 They went to ER, they got scan.

20:28.540 --> 20:30.660
 There were early signs on the scan

20:30.660 --> 20:33.540
 and that would save their lives.

20:33.540 --> 20:35.820
 But this wasn't really an accidental case.

20:35.820 --> 20:40.820
 So as we become better, we would be able to help

20:41.260 --> 20:46.260
 to many more people that are likely to develop diseases.

20:46.540 --> 20:51.020
 And I just want to say that as I got more into this field,

20:51.020 --> 20:53.620
 I realized that cancer is of course terrible disease,

20:53.620 --> 20:56.700
 but there are really the whole slew of terrible diseases

20:56.700 --> 21:00.820
 out there like neurodegenerative diseases and others.

21:01.660 --> 21:04.580
 So we, of course, a lot of us are fixated on cancer

21:04.580 --> 21:06.420
 because it's so prevalent in our society.

21:06.420 --> 21:08.540
 And you see these people where there are a lot of patients

21:08.540 --> 21:10.340
 with neurodegenerative diseases

21:10.340 --> 21:12.540
 and the kind of aging diseases

21:12.540 --> 21:17.100
 that we still don't have a good solution for.

21:17.100 --> 21:22.100
 And I felt as a computer scientist,

21:22.860 --> 21:25.460
 we kind of decided that it's other people's job

21:25.460 --> 21:29.340
 to treat these diseases because it's like traditionally

21:29.340 --> 21:32.420
 people in biology or in chemistry or MDs

21:32.420 --> 21:35.340
 are the ones who are thinking about it.

21:35.340 --> 21:37.420
 And after kind of start paying attention,

21:37.420 --> 21:40.340
 I think that it's really a wrong assumption

21:40.340 --> 21:42.940
 and we all need to join the battle.

21:42.940 --> 21:46.460
 So how it seems like in cancer specifically

21:46.460 --> 21:49.140
 that there's a lot of ways that machine learning can help.

21:49.140 --> 21:51.860
 So what's the role of machine learning

21:51.860 --> 21:54.100
 in the diagnosis of cancer?

21:55.260 --> 21:58.700
 So for many cancers today, we really don't know

21:58.700 --> 22:03.460
 what is your likelihood to get cancer.

22:03.460 --> 22:06.300
 And for the vast majority of patients,

22:06.300 --> 22:07.940
 especially on the younger patients,

22:07.940 --> 22:09.580
 it really comes as a surprise.

22:09.580 --> 22:11.140
 Like for instance, for breast cancer,

22:11.140 --> 22:13.860
 80% of the patients are first in their families,

22:13.860 --> 22:15.380
 it's like me.

22:15.380 --> 22:18.460
 And I never saw that I had any increased risk

22:18.460 --> 22:20.820
 because nobody had it in my family.

22:20.820 --> 22:22.300
 And for some reason in my head,

22:22.300 --> 22:24.820
 it was kind of inherited disease.

22:26.580 --> 22:28.380
 But even if I would pay attention,

22:28.380 --> 22:32.420
 the very simplistic statistical models

22:32.420 --> 22:34.540
 that are currently used in clinical practice,

22:34.540 --> 22:37.460
 they really don't give you an answer, so you don't know.

22:37.460 --> 22:40.380
 And the same true for pancreatic cancer,

22:40.380 --> 22:45.380
 the same true for non smoking lung cancer and many others.

22:45.380 --> 22:47.340
 So what machine learning can do here

22:47.340 --> 22:51.620
 is utilize all this data to tell us early

22:51.620 --> 22:53.140
 who is likely to be susceptible

22:53.140 --> 22:55.980
 and using all the information that is already there,

22:55.980 --> 22:59.980
 be it imaging, be it your other tests,

22:59.980 --> 23:04.860
 and eventually liquid biopsies and others,

23:04.860 --> 23:08.180
 where the signal itself is not sufficiently strong

23:08.180 --> 23:11.300
 for human eye to do good discrimination

23:11.300 --> 23:12.940
 because the signal may be weak,

23:12.940 --> 23:15.620
 but by combining many sources,

23:15.620 --> 23:18.100
 machine which is trained on large volumes of data

23:18.100 --> 23:20.700
 can really detect it early.

23:20.700 --> 23:22.500
 And that's what we've seen with breast cancer

23:22.500 --> 23:25.900
 and people are reporting it in other diseases as well.

23:25.900 --> 23:28.260
 That really boils down to data, right?

23:28.260 --> 23:30.980
 And in the different kinds of sources of data.

23:30.980 --> 23:33.740
 And you mentioned regulatory challenges.

23:33.740 --> 23:35.180
 So what are the challenges

23:35.180 --> 23:39.260
 in gathering large data sets in this space?

23:40.860 --> 23:42.660
 Again, another great question.

23:42.660 --> 23:45.500
 So it took me after I decided that I want to work on it

23:45.500 --> 23:48.740
 two years to get access to data.

23:48.740 --> 23:50.580
 Any data, like any significant data set?

23:50.580 --> 23:53.580
 Any significant amount, like right now in this country,

23:53.580 --> 23:57.060
 there is no publicly available data set

23:57.060 --> 23:58.820
 of modern mammograms that you can just go

23:58.820 --> 24:01.860
 on your computer, sign a document and get it.

24:01.860 --> 24:03.180
 It just doesn't exist.

24:03.180 --> 24:06.860
 I mean, obviously every hospital has its own collection

24:06.860 --> 24:07.700
 of mammograms.

24:07.700 --> 24:11.300
 There are data that came out of clinical trials.

24:11.300 --> 24:13.220
 What we're talking about here is a computer scientist

24:13.220 --> 24:17.140
 who just wants to run his or her model

24:17.140 --> 24:19.060
 and see how it works.

24:19.060 --> 24:22.900
 This data, like ImageNet, doesn't exist.

24:22.900 --> 24:27.900
 And there is a set which is called like Florida data set

24:28.620 --> 24:30.860
 which is a film mammogram from 90s

24:30.860 --> 24:32.420
 which is totally not representative

24:32.420 --> 24:33.860
 of the current developments.

24:33.860 --> 24:35.780
 Whatever you're learning on them doesn't scale up.

24:35.780 --> 24:39.300
 This is the only resource that is available.

24:39.300 --> 24:42.780
 And today there are many agencies

24:42.780 --> 24:44.460
 that govern access to data.

24:44.460 --> 24:46.300
 Like the hospital holds your data

24:46.300 --> 24:49.260
 and the hospital decides whether they would give it

24:49.260 --> 24:52.340
 to the researcher to work with this data or not.

24:52.340 --> 24:54.180
 Individual hospital?

24:54.180 --> 24:55.020
 Yeah.

24:55.020 --> 24:57.220
 I mean, the hospital may, you know,

24:57.220 --> 24:59.220
 assuming that you're doing research collaboration,

24:59.220 --> 25:01.980
 you can submit, you know,

25:01.980 --> 25:05.060
 there is a proper approval process guided by RB

25:05.060 --> 25:07.820
 and if you go through all the processes,

25:07.820 --> 25:10.140
 you can eventually get access to the data.

25:10.140 --> 25:13.540
 But if you yourself know our OEI community,

25:13.540 --> 25:16.100
 there are not that many people who actually ever got access

25:16.100 --> 25:20.260
 to data because it's very challenging process.

25:20.260 --> 25:22.780
 And sorry, just in a quick comment,

25:22.780 --> 25:25.780
 MGH or any kind of hospital,

25:25.780 --> 25:28.100
 are they scanning the data?

25:28.100 --> 25:29.740
 Are they digitally storing it?

25:29.740 --> 25:31.580
 Oh, it is already digitally stored.

25:31.580 --> 25:34.180
 You don't need to do any extra processing steps.

25:34.180 --> 25:38.340
 It's already there in the right format is that right now

25:38.340 --> 25:41.180
 there are a lot of issues that govern access to the data

25:41.180 --> 25:46.180
 because the hospital is legally responsible for the data.

25:46.180 --> 25:51.020
 And, you know, they have a lot to lose

25:51.020 --> 25:53.140
 if they give the data to the wrong person,

25:53.140 --> 25:56.460
 but they may not have a lot to gain if they give it

25:56.460 --> 26:00.580
 as a hospital, as a legal entity has given it to you.

26:00.580 --> 26:02.740
 And the way, you know, what I would imagine

26:02.740 --> 26:05.220
 happening in the future is the same thing that happens

26:05.220 --> 26:06.780
 when you're getting your driving license,

26:06.780 --> 26:09.820
 you can decide whether you want to donate your organs.

26:09.820 --> 26:13.100
 You can imagine that whenever a person goes to the hospital,

26:13.100 --> 26:17.540
 they, it should be easy for them to donate their data

26:17.540 --> 26:19.420
 for research and it can be different kind of,

26:19.420 --> 26:22.420
 do they only give you a test results or only mammogram

26:22.420 --> 26:25.900
 or only imaging data or the whole medical record?

26:27.060 --> 26:28.980
 Because at the end,

26:30.540 --> 26:33.860
 we all will benefit from all this insights.

26:33.860 --> 26:36.060
 And it's not like you say, I want to keep my data private,

26:36.060 --> 26:38.780
 but I would really love to get it from other people

26:38.780 --> 26:40.740
 because other people are thinking the same way.

26:40.740 --> 26:45.740
 So if there is a mechanism to do this donation

26:45.740 --> 26:48.020
 and the patient has an ability to say

26:48.020 --> 26:50.820
 how they want to use their data for research,

26:50.820 --> 26:54.100
 it would be really a game changer.

26:54.100 --> 26:56.460
 People, when they think about this problem,

26:56.460 --> 26:58.460
 there's a, it depends on the population,

26:58.460 --> 27:00.140
 depends on the demographics,

27:00.140 --> 27:03.420
 but there's some privacy concerns generally,

27:03.420 --> 27:05.860
 not just medical data, just any kind of data.

27:05.860 --> 27:09.620
 It's what you said, my data, it should belong kind of to me.

27:09.620 --> 27:11.660
 I'm worried how it's going to be misused.

27:12.540 --> 27:15.620
 How do we alleviate those concerns?

27:17.100 --> 27:19.460
 Because that seems like a problem that needs to be,

27:19.460 --> 27:22.980
 that problem of trust, of transparency needs to be solved

27:22.980 --> 27:27.260
 before we build large data sets that help detect cancer,

27:27.260 --> 27:30.180
 help save those very people in the future.

27:30.180 --> 27:31.940
 So I think there are two things that could be done.

27:31.940 --> 27:34.460
 There is a technical solutions

27:34.460 --> 27:38.220
 and there are societal solutions.

27:38.220 --> 27:40.180
 So on the technical end,

27:41.460 --> 27:46.460
 we today have ability to improve disambiguation.

27:48.140 --> 27:49.740
 Like, for instance, for imaging,

27:49.740 --> 27:54.740
 it's, you know, for imaging, you can do it pretty well.

27:55.620 --> 27:56.780
 What's disambiguation?

27:56.780 --> 27:58.540
 And disambiguation, sorry, disambiguation,

27:58.540 --> 27:59.860
 removing the identification,

27:59.860 --> 28:02.220
 removing the names of the people.

28:02.220 --> 28:04.820
 There are other data, like if it is a raw tax,

28:04.820 --> 28:08.180
 you cannot really achieve 99.9%,

28:08.180 --> 28:10.060
 but there are all these techniques

28:10.060 --> 28:12.460
 that actually some of them are developed at MIT,

28:12.460 --> 28:15.460
 how you can do learning on the encoded data

28:15.460 --> 28:17.420
 where you locally encode the image,

28:17.420 --> 28:22.420
 you train a network which only works on the encoded images

28:22.420 --> 28:24.940
 and then you send the outcome back to the hospital

28:24.940 --> 28:26.580
 and you can open it up.

28:26.580 --> 28:28.020
 So those are the technical solutions.

28:28.020 --> 28:30.660
 There are a lot of people who are working in this space

28:30.660 --> 28:33.780
 where the learning happens in the encoded form.

28:33.780 --> 28:36.180
 We are still early,

28:36.180 --> 28:39.260
 but this is an interesting research area

28:39.260 --> 28:41.900
 where I think we'll make more progress.

28:43.340 --> 28:45.620
 There is a lot of work in natural language processing

28:45.620 --> 28:48.620
 community how to do the identification better.

28:50.380 --> 28:54.020
 But even today, there are already a lot of data

28:54.020 --> 28:55.900
 which can be deidentified perfectly,

28:55.900 --> 28:58.780
 like your test data, for instance, correct,

28:58.780 --> 29:00.980
 where you can just, you know the name of the patient,

29:00.980 --> 29:04.300
 you just want to extract the part with the numbers.

29:04.300 --> 29:07.460
 The big problem here is again,

29:08.420 --> 29:10.420
 hospitals don't see much incentive

29:10.420 --> 29:12.660
 to give this data away on one hand

29:12.660 --> 29:14.220
 and then there is general concern.

29:14.220 --> 29:17.700
 Now, when I'm talking about societal benefits

29:17.700 --> 29:19.660
 and about the education,

29:19.660 --> 29:24.340
 the public needs to understand that I think

29:25.700 --> 29:29.420
 that there are situation and I still remember myself

29:29.420 --> 29:33.380
 when I really needed an answer, I had to make a choice.

29:33.380 --> 29:35.220
 There was no information to make a choice,

29:35.220 --> 29:36.660
 you're just guessing.

29:36.660 --> 29:41.060
 And at that moment you feel that your life is at the stake,

29:41.060 --> 29:44.820
 but you just don't have information to make the choice.

29:44.820 --> 29:48.740
 And many times when I give talks,

29:48.740 --> 29:51.300
 I get emails from women who say,

29:51.300 --> 29:52.820
 you know, I'm in this situation,

29:52.820 --> 29:55.940
 can you please run statistic and see what are the outcomes?

29:57.100 --> 30:01.300
 We get almost every week a mammogram that comes by mail

30:01.300 --> 30:03.460
 to my office at MIT, I'm serious.

30:04.380 --> 30:07.860
 That people ask to run because they need to make

30:07.860 --> 30:10.020
 life changing decisions.

30:10.020 --> 30:12.980
 And of course, I'm not planning to open a clinic here,

30:12.980 --> 30:16.660
 but we do run and give them the results for their doctors.

30:16.660 --> 30:20.100
 But the point that I'm trying to make,

30:20.100 --> 30:23.780
 that we all at some point or our loved ones

30:23.780 --> 30:26.620
 will be in the situation where you need information

30:26.620 --> 30:28.860
 to make the best choice.

30:28.860 --> 30:31.860
 And if this information is not available,

30:31.860 --> 30:35.100
 you would feel vulnerable and unprotected.

30:35.100 --> 30:37.860
 And then the question is, you know, what do I care more?

30:37.860 --> 30:40.380
 Because at the end, everything is a trade off, correct?

30:40.380 --> 30:41.700
 Yeah, exactly.

30:41.700 --> 30:45.580
 Just out of curiosity, it seems like one possible solution,

30:45.580 --> 30:47.420
 I'd like to see what you think of it,

30:49.340 --> 30:50.660
 based on what you just said,

30:50.660 --> 30:52.500
 based on wanting to know answers

30:52.500 --> 30:55.060
 for when you're yourself in that situation.

30:55.060 --> 30:58.420
 Is it possible for patients to own their data

30:58.420 --> 31:01.020
 as opposed to hospitals owning their data?

31:01.020 --> 31:04.100
 Of course, theoretically, I guess patients own their data,

31:04.100 --> 31:06.620
 but can you walk out there with a USB stick

31:07.580 --> 31:10.620
 containing everything or upload it to the cloud?

31:10.620 --> 31:14.500
 Where a company, you know, I remember Microsoft

31:14.500 --> 31:17.820
 had a service, like I try, I was really excited about

31:17.820 --> 31:19.260
 and Google Health was there.

31:19.260 --> 31:21.900
 I tried to give, I was excited about it.

31:21.900 --> 31:24.780
 Basically companies helping you upload your data

31:24.780 --> 31:27.940
 to the cloud so that you can move from hospital to hospital

31:27.940 --> 31:29.260
 from doctor to doctor.

31:29.260 --> 31:32.700
 Do you see a promise of that kind of possibility?

31:32.700 --> 31:34.660
 I absolutely think this is, you know,

31:34.660 --> 31:38.180
 the right way to exchange the data.

31:38.180 --> 31:41.700
 I don't know now who's the biggest player in this field,

31:41.700 --> 31:45.940
 but I can clearly see that even for totally selfish

31:45.940 --> 31:49.300
 health reasons, when you are going to a new facility

31:49.300 --> 31:52.620
 and many of us are sent to some specialized treatment,

31:52.620 --> 31:55.740
 they don't easily have access to your data.

31:55.740 --> 31:59.420
 And today, you know, we might want to send this mammogram,

31:59.420 --> 32:01.780
 need to go to the hospital, find some small office

32:01.780 --> 32:04.820
 which gives them the CD and they ship as a CD.

32:04.820 --> 32:08.340
 So you can imagine we're looking at kind of decades old

32:08.340 --> 32:10.100
 mechanism of data exchange.

32:11.340 --> 32:15.620
 So I definitely think this is an area where hopefully

32:15.620 --> 32:20.380
 all the right regulatory and technical forces will align

32:20.380 --> 32:23.220
 and we will see it actually implemented.

32:23.220 --> 32:27.500
 It's sad because unfortunately, and I need to research

32:27.500 --> 32:30.620
 why that happened, but I'm pretty sure Google Health

32:30.620 --> 32:32.940
 and Microsoft Health Vault or whatever it's called

32:32.940 --> 32:36.100
 both closed down, which means that there was

32:36.100 --> 32:39.100
 either regulatory pressure or there's not a business case

32:39.100 --> 32:41.820
 or there's challenges from hospitals,

32:41.820 --> 32:43.260
 which is very disappointing.

32:43.260 --> 32:46.500
 So when you say you don't know what the biggest players are,

32:46.500 --> 32:50.540
 the two biggest that I was aware of closed their doors.

32:50.540 --> 32:53.140
 So I'm hoping, I'd love to see why

32:53.140 --> 32:54.780
 and I'd love to see who else can come up.

32:54.780 --> 32:59.620
 It seems like one of those Elon Musk style problems

32:59.620 --> 33:01.300
 that are obvious needs to be solved

33:01.300 --> 33:02.980
 and somebody needs to step up and actually do

33:02.980 --> 33:07.540
 this large scale data collection.

33:07.540 --> 33:09.620
 So I know there is an initiative in Massachusetts,

33:09.620 --> 33:11.740
 I think, which you led by the governor

33:11.740 --> 33:15.460
 to try to create this kind of health exchange system

33:15.460 --> 33:17.860
 where at least to help people who kind of when you show up

33:17.860 --> 33:20.220
 in emergency room and there is no information

33:20.220 --> 33:23.540
 about what are your allergies and other things.

33:23.540 --> 33:26.140
 So I don't know how far it will go.

33:26.140 --> 33:28.180
 But another thing that you said

33:28.180 --> 33:30.780
 and I find it very interesting is actually

33:30.780 --> 33:33.780
 who are the successful players in this space

33:33.780 --> 33:37.260
 and the whole implementation, how does it go?

33:37.260 --> 33:40.300
 To me, it is from the anthropological perspective,

33:40.300 --> 33:44.660
 it's more fascinating that AI that today goes in healthcare,

33:44.660 --> 33:50.380
 we've seen so many attempts and so very little successes.

33:50.380 --> 33:54.220
 And it's interesting to understand that I've by no means

33:54.220 --> 33:56.700
 have knowledge to assess it,

33:56.700 --> 33:59.620
 why we are in the position where we are.

33:59.620 --> 34:02.940
 Yeah, it's interesting because data is really fuel

34:02.940 --> 34:04.980
 for a lot of successful applications.

34:04.980 --> 34:08.500
 And when that data acquires regulatory approval,

34:08.500 --> 34:12.940
 like the FDA or any kind of approval,

34:12.940 --> 34:15.740
 it seems that the computer scientists

34:15.740 --> 34:17.460
 are not quite there yet in being able

34:17.460 --> 34:18.900
 to play the regulatory game,

34:18.900 --> 34:21.220
 understanding the fundamentals of it.

34:21.220 --> 34:26.500
 I think that in many cases when even people do have data,

34:26.500 --> 34:31.300
 we still don't know what exactly do you need to demonstrate

34:31.300 --> 34:33.860
 to change the standard of care.

34:35.500 --> 34:37.180
 Like let me give you an example

34:37.180 --> 34:41.100
 related to my breast cancer research.

34:41.100 --> 34:45.500
 So in traditional breast cancer risk assessment,

34:45.500 --> 34:47.140
 there is something called density,

34:47.140 --> 34:50.500
 which determines the likelihood of a woman to get cancer.

34:50.500 --> 34:51.700
 And this pretty much says,

34:51.700 --> 34:54.220
 how much white do you see on the mammogram?

34:54.220 --> 34:58.980
 The whiter it is, the more likely the tissue is dense.

34:58.980 --> 35:03.660
 And the idea behind density, it's not a bad idea.

35:03.660 --> 35:08.100
 In 1967, a radiologist called Wolf decided to look back

35:08.100 --> 35:09.780
 at women who were diagnosed

35:09.780 --> 35:12.420
 and see what is special in their images.

35:12.420 --> 35:14.700
 Can we look back and say that they're likely to develop?

35:14.700 --> 35:16.180
 So he come up with some patterns.

35:16.180 --> 35:20.660
 And it was the best that his human eye can identify.

35:20.660 --> 35:22.060
 Then it was kind of formalized

35:22.060 --> 35:24.220
 and coded into four categories.

35:24.220 --> 35:26.940
 And that's what we are using today.

35:26.940 --> 35:31.020
 And today this density assessment

35:31.020 --> 35:34.620
 is actually a federal law from 2019,

35:34.620 --> 35:36.180
 approved by President Trump

35:36.180 --> 35:40.100
 and for the previous FDA commissioner,

35:40.100 --> 35:43.620
 where women are supposed to be advised by their providers

35:43.620 --> 35:45.100
 if they have high density,

35:45.100 --> 35:47.260
 putting them into higher risk category.

35:47.260 --> 35:49.460
 And in some states,

35:49.460 --> 35:51.260
 you can actually get supplementary screening

35:51.260 --> 35:53.700
 paid by your insurance because you're in this category.

35:53.700 --> 35:56.780
 Now you can say, how much science do we have behind it?

35:56.780 --> 36:00.820
 Whatever, biological science or epidemiological evidence.

36:00.820 --> 36:05.140
 So it turns out that between 40 and 50% of women

36:05.140 --> 36:06.660
 have dense breasts.

36:06.660 --> 36:11.140
 So about 40% of patients are coming out of their screening

36:11.140 --> 36:15.020
 and somebody tells them, you are in high risk.

36:15.020 --> 36:16.860
 Now, what exactly does it mean

36:16.860 --> 36:19.620
 if you as half of the population in high risk?

36:19.620 --> 36:22.060
 It's from saying, maybe I'm not,

36:22.060 --> 36:23.700
 or what do I really need to do with it?

36:23.700 --> 36:27.220
 Because the system doesn't provide me

36:27.220 --> 36:28.340
 a lot of the solutions

36:28.340 --> 36:30.140
 because there are so many people like me,

36:30.140 --> 36:34.620
 we cannot really provide very expensive solutions for them.

36:34.620 --> 36:38.740
 And the reason this whole density became this big deal,

36:38.740 --> 36:40.820
 it's actually advocated by the patients

36:40.820 --> 36:42.500
 who felt very unprotected

36:42.500 --> 36:44.900
 because many women went and did the mammograms

36:44.900 --> 36:46.260
 which were normal.

36:46.260 --> 36:49.460
 And then it turns out that they already had cancer,

36:49.460 --> 36:50.580
 quite developed cancer.

36:50.580 --> 36:54.420
 So they didn't have a way to know who is really at risk

36:54.420 --> 36:56.300
 and what is the likelihood that when the doctor tells you,

36:56.300 --> 36:58.060
 you're okay, you are not okay.

36:58.060 --> 37:02.140
 So at the time, and it was 15 years ago,

37:02.140 --> 37:06.820
 this maybe was the best piece of science that we had.

37:06.820 --> 37:11.820
 And it took quite 15, 16 years to make it federal law.

37:12.180 --> 37:15.660
 But now this is a standard.

37:15.660 --> 37:17.620
 Now with a deep learning model,

37:17.620 --> 37:19.660
 we can so much more accurately predict

37:19.660 --> 37:21.580
 who is gonna develop breast cancer

37:21.580 --> 37:23.700
 just because you're trained on a logical thing.

37:23.700 --> 37:26.060
 And instead of describing how much white

37:26.060 --> 37:27.380
 and what kind of white machine

37:27.380 --> 37:30.140
 can systematically identify the patterns,

37:30.140 --> 37:32.780
 which was the original idea behind the thought

37:32.780 --> 37:33.700
 of the cardiologist,

37:33.700 --> 37:35.740
 machines can do it much more systematically

37:35.740 --> 37:38.260
 and predict the risk when you're training the machine

37:38.260 --> 37:42.140
 to look at the image and to say the risk in one to five years.

37:42.140 --> 37:45.060
 Now you can ask me how long it will take

37:45.060 --> 37:46.460
 to substitute this density,

37:46.460 --> 37:48.620
 which is broadly used across the country

37:48.620 --> 37:53.620
 and really is not helping to bring this new models.

37:54.380 --> 37:56.700
 And I would say it's not a matter of the algorithm.

37:56.700 --> 37:58.780
 Algorithms use already orders of magnitude better

37:58.780 --> 38:00.460
 than what is currently in practice.

38:00.460 --> 38:02.500
 I think it's really the question,

38:02.500 --> 38:04.380
 who do you need to convince?

38:04.380 --> 38:07.460
 How many hospitals do you need to run the experiment?

38:07.460 --> 38:11.500
 What, you know, all this mechanism of adoption

38:11.500 --> 38:15.180
 and how do you explain to patients

38:15.180 --> 38:17.580
 and to women across the country

38:17.580 --> 38:20.460
 that this is really a better measure?

38:20.460 --> 38:22.740
 And again, I don't think it's an AI question.

38:22.740 --> 38:25.940
 We can work more and make the algorithm even better,

38:25.940 --> 38:29.300
 but I don't think that this is the current, you know,

38:29.300 --> 38:32.060
 the barrier, the barrier is really this other piece

38:32.060 --> 38:35.260
 that for some reason is not really explored.

38:35.260 --> 38:36.860
 It's like anthropological piece.

38:36.860 --> 38:39.860
 And coming back to your question about books,

38:39.860 --> 38:42.980
 there is a book that I'm reading.

38:42.980 --> 38:47.980
 It's called American Sickness by Elizabeth Rosenthal.

38:48.260 --> 38:51.580
 And I got this book from my clinical collaborator,

38:51.580 --> 38:53.100
 Dr. Connie Lehman.

38:53.100 --> 38:54.820
 And I said, I know everything that I need to know

38:54.820 --> 38:56.020
 about American health system,

38:56.020 --> 38:59.220
 but you know, every page doesn't fail to surprise me.

38:59.220 --> 39:03.140
 And I think there is a lot of interesting

39:03.140 --> 39:06.860
 and really deep lessons for people like us

39:06.860 --> 39:09.660
 from computer science who are coming into this field

39:09.660 --> 39:13.660
 to really understand how complex is the system of incentives

39:13.660 --> 39:17.660
 in the system to understand how you really need to play

39:17.660 --> 39:18.780
 to drive adoption.

39:19.740 --> 39:21.180
 You just said it's complex,

39:21.180 --> 39:23.980
 but if we're trying to simplify it,

39:23.980 --> 39:27.380
 who do you think most likely would be successful

39:27.380 --> 39:29.540
 if we push on this group of people?

39:29.540 --> 39:30.780
 Is it the doctors?

39:30.780 --> 39:31.820
 Is it the hospitals?

39:31.820 --> 39:34.300
 Is it the governments or policymakers?

39:34.300 --> 39:37.380
 Is it the individual patients, consumers?

39:38.860 --> 39:43.860
 Who needs to be inspired to most likely lead to adoption?

39:45.180 --> 39:47.100
 Or is there no simple answer?

39:47.100 --> 39:48.260
 There's no simple answer,

39:48.260 --> 39:51.980
 but I think there is a lot of good people in medical system

39:51.980 --> 39:55.180
 who do want to make a change.

39:56.460 --> 40:01.460
 And I think a lot of power will come from us as consumers

40:01.540 --> 40:04.260
 because we all are consumers or future consumers

40:04.260 --> 40:06.500
 of healthcare services.

40:06.500 --> 40:11.500
 And I think we can do so much more

40:12.060 --> 40:15.500
 in explaining the potential and not in the hype terms

40:15.500 --> 40:17.900
 and not saying that we now killed all Alzheimer

40:17.900 --> 40:20.500
 and I'm really sick of reading this kind of articles

40:20.500 --> 40:22.100
 which make these claims,

40:22.100 --> 40:24.780
 but really to show with some examples

40:24.780 --> 40:29.060
 what this implementation does and how it changes the care.

40:29.060 --> 40:30.020
 Because I can't imagine,

40:30.020 --> 40:33.220
 it doesn't matter what kind of politician it is,

40:33.220 --> 40:35.220
 we all are susceptible to these diseases.

40:35.220 --> 40:37.740
 There is no one who is free.

40:37.740 --> 40:41.060
 And eventually, we all are humans

40:41.060 --> 40:44.860
 and we're looking for a way to alleviate the suffering.

40:44.860 --> 40:47.260
 And this is one possible way

40:47.260 --> 40:49.300
 where we currently are under utilizing,

40:49.300 --> 40:50.940
 which I think can help.

40:51.860 --> 40:55.100
 So it sounds like the biggest problems are outside of AI

40:55.100 --> 40:57.980
 in terms of the biggest impact at this point.

40:57.980 --> 41:00.420
 But are there any open problems

41:00.420 --> 41:03.780
 in the application of ML to oncology in general?

41:03.780 --> 41:07.540
 So improving the detection or any other creative methods,

41:07.540 --> 41:09.620
 whether it's on the detection segmentations

41:09.620 --> 41:11.780
 or the vision perception side

41:11.780 --> 41:16.260
 or some other clever of inference?

41:16.260 --> 41:19.620
 Yeah, what in general in your view are the open problems

41:19.620 --> 41:20.460
 in this space?

41:20.460 --> 41:22.460
 Yeah, I just want to mention that beside detection,

41:22.460 --> 41:24.820
 not the area where I am kind of quite active

41:24.820 --> 41:28.580
 and I think it's really an increasingly important area

41:28.580 --> 41:30.940
 in healthcare is drug design.

41:32.260 --> 41:33.100
 Absolutely.

41:33.100 --> 41:36.900
 Because it's fine if you detect something early,

41:36.900 --> 41:41.100
 but you still need to get drugs

41:41.100 --> 41:43.860
 and new drugs for these conditions.

41:43.860 --> 41:46.740
 And today, all of the drug design,

41:46.740 --> 41:48.300
 ML is non existent there.

41:48.300 --> 41:52.980
 We don't have any drug that was developed by the ML model

41:52.980 --> 41:54.900
 or even not developed,

41:54.900 --> 41:57.060
 but at least even knew that ML model

41:57.060 --> 41:59.260
 plays some significant role.

41:59.260 --> 42:03.300
 I think this area with all the new ability

42:03.300 --> 42:05.780
 to generate molecules with desired properties

42:05.780 --> 42:10.780
 to do in silica screening is really a big open area.

42:11.460 --> 42:12.740
 To be totally honest with you,

42:12.740 --> 42:14.900
 when we are doing diagnostics and imaging,

42:14.900 --> 42:17.260
 primarily taking the ideas that were developed

42:17.260 --> 42:20.460
 for other areas and you applying them with some adaptation,

42:20.460 --> 42:25.460
 the area of drug design is really technically interesting

42:26.820 --> 42:27.980
 and exciting area.

42:27.980 --> 42:30.380
 You need to work a lot with graphs

42:30.380 --> 42:34.580
 and capture various 3D properties.

42:34.580 --> 42:37.420
 There are lots and lots of opportunities

42:37.420 --> 42:39.820
 to be technically creative.

42:39.820 --> 42:44.820
 And I think there are a lot of open questions in this area.

42:46.820 --> 42:48.820
 We're already getting a lot of successes

42:48.820 --> 42:52.700
 even with kind of the first generation of these models,

42:52.700 --> 42:56.500
 but there is much more new creative things that you can do.

42:56.500 --> 42:59.260
 And what's very nice to see is that actually

42:59.260 --> 43:04.180
 the more powerful, the more interesting models

43:04.180 --> 43:05.460
 actually do do better.

43:05.460 --> 43:10.460
 So there is a place to innovate in machine learning

43:11.300 --> 43:12.540
 in this area.

43:13.900 --> 43:16.820
 And some of these techniques are really unique to,

43:16.820 --> 43:19.620
 let's say, to graph generation and other things.

43:19.620 --> 43:20.820
 So...

43:20.820 --> 43:23.980
 What, just to interrupt really quick, I'm sorry,

43:23.980 --> 43:28.980
 graph generation or graphs, drug discovery in general,

43:30.620 --> 43:31.940
 how do you discover a drug?

43:31.940 --> 43:33.340
 Is this chemistry?

43:33.340 --> 43:37.500
 Is this trying to predict different chemical reactions?

43:37.500 --> 43:39.660
 Or is it some kind of...

43:39.660 --> 43:42.100
 What do graphs even represent in this space?

43:42.100 --> 43:43.980
 Oh, sorry, sorry.

43:43.980 --> 43:45.340
 And what's a drug?

43:45.340 --> 43:47.140
 Okay, so let's say you're thinking

43:47.140 --> 43:48.540
 there are many different types of drugs,

43:48.540 --> 43:50.580
 but let's say you're gonna talk about small molecules

43:50.580 --> 43:52.860
 because I think today the majority of drugs

43:52.860 --> 43:53.700
 are small molecules.

43:53.700 --> 43:55.020
 So small molecule is a graph.

43:55.020 --> 43:59.180
 The molecule is just where the node in the graph

43:59.180 --> 44:01.500
 is an atom and then you have the bonds.

44:01.500 --> 44:03.220
 So it's really a graph representation.

44:03.220 --> 44:05.540
 If you look at it in 2D, correct,

44:05.540 --> 44:07.460
 you can do it 3D, but let's say,

44:07.460 --> 44:09.540
 let's keep it simple and stick in 2D.

44:11.500 --> 44:14.740
 So pretty much my understanding today,

44:14.740 --> 44:18.620
 how it is done at scale in the companies,

44:18.620 --> 44:20.220
 without machine learning,

44:20.220 --> 44:22.100
 you have high throughput screening.

44:22.100 --> 44:23.740
 So you know that you are interested

44:23.740 --> 44:26.540
 to get certain biological activity of the compound.

44:26.540 --> 44:28.860
 So you scan a lot of compounds,

44:28.860 --> 44:30.700
 like maybe hundreds of thousands,

44:30.700 --> 44:32.980
 some really big number of compounds.

44:32.980 --> 44:36.060
 You identify some compounds which have the right activity

44:36.060 --> 44:39.220
 and then at this point, the chemists come

44:39.220 --> 44:43.220
 and they're trying to now to optimize

44:43.220 --> 44:45.340
 this original heat to different properties

44:45.340 --> 44:47.180
 that you want it to be maybe soluble,

44:47.180 --> 44:49.060
 you want it to decrease toxicity,

44:49.060 --> 44:51.620
 you want it to decrease the side effects.

44:51.620 --> 44:54.020
 Are those, sorry again to interrupt,

44:54.020 --> 44:55.500
 can that be done in simulation

44:55.500 --> 44:57.700
 or just by looking at the molecules

44:57.700 --> 44:59.820
 or do you need to actually run reactions

44:59.820 --> 45:02.460
 in real labs with lab coats and stuff?

45:02.460 --> 45:04.020
 So when you do high throughput screening,

45:04.020 --> 45:06.100
 you really do screening.

45:06.100 --> 45:07.020
 It's in the lab.

45:07.020 --> 45:09.140
 It's really the lab screening.

45:09.140 --> 45:10.980
 You screen the molecules, correct?

45:10.980 --> 45:12.580
 I don't know what screening is.

45:12.580 --> 45:15.060
 The screening is just check them for certain property.

45:15.060 --> 45:17.260
 Like in the physical space, in the physical world,

45:17.260 --> 45:18.740
 like actually there's a machine probably

45:18.740 --> 45:21.420
 that's actually running the reaction.

45:21.420 --> 45:22.900
 Actually running the reactions, yeah.

45:22.900 --> 45:25.420
 So there is a process where you can run

45:25.420 --> 45:26.660
 and that's why it's called high throughput

45:26.660 --> 45:29.580
 that it become cheaper and faster

45:29.580 --> 45:33.820
 to do it on very big number of molecules.

45:33.820 --> 45:35.820
 You run the screening,

45:35.820 --> 45:40.300
 you identify potential good starts

45:40.300 --> 45:42.340
 and then when the chemists come in

45:42.340 --> 45:44.060
 who have done it many times

45:44.060 --> 45:46.180
 and then they can try to look at it and say,

45:46.180 --> 45:48.260
 how can you change the molecule

45:48.260 --> 45:51.780
 to get the desired profile

45:51.780 --> 45:53.460
 in terms of all other properties?

45:53.460 --> 45:56.500
 So maybe how do I make it more bioactive and so on?

45:56.500 --> 45:59.460
 And there the creativity of the chemists

45:59.460 --> 46:03.980
 really is the one that determines the success

46:03.980 --> 46:07.460
 of this design because again,

46:07.460 --> 46:09.300
 they have a lot of domain knowledge

46:09.300 --> 46:12.900
 of what works, how do you decrease the CCD and so on

46:12.900 --> 46:15.020
 and that's what they do.

46:15.020 --> 46:17.860
 So all the drugs that are currently

46:17.860 --> 46:20.220
 in the FDA approved drugs

46:20.220 --> 46:22.140
 or even drugs that are in clinical trials,

46:22.140 --> 46:27.100
 they are designed using these domain experts

46:27.100 --> 46:30.060
 which goes through this combinatorial space

46:30.060 --> 46:31.940
 of molecules or graphs or whatever

46:31.940 --> 46:35.140
 and find the right one or adjust it to be the right ones.

46:35.140 --> 46:38.060
 It sounds like the breast density heuristic

46:38.060 --> 46:40.460
 from 67 to the same echoes.

46:40.460 --> 46:41.820
 It's not necessarily that.

46:41.820 --> 46:45.380
 It's really driven by deep understanding.

46:45.380 --> 46:46.820
 It's not like they just observe it.

46:46.820 --> 46:48.540
 I mean, they do deeply understand chemistry

46:48.540 --> 46:50.460
 and they do understand how different groups

46:50.460 --> 46:53.140
 and how does it changes the properties.

46:53.140 --> 46:56.660
 So there is a lot of science that gets into it

46:56.660 --> 46:58.740
 and a lot of kind of simulation,

46:58.740 --> 47:00.940
 how do you want it to behave?

47:01.900 --> 47:03.900
 It's very, very complex.

47:03.900 --> 47:06.140
 So they're quite effective at this design, obviously.

47:06.140 --> 47:08.420
 Now effective, yeah, we have drugs.

47:08.420 --> 47:10.780
 Like depending on how do you measure effective,

47:10.780 --> 47:13.940
 if you measure it in terms of cost, it's prohibitive.

47:13.940 --> 47:15.820
 If you measure it in terms of times,

47:15.820 --> 47:18.420
 we have lots of diseases for which we don't have any drugs

47:18.420 --> 47:20.060
 and we don't even know how to approach

47:20.060 --> 47:23.460
 and don't need to mention few drugs

47:23.460 --> 47:27.140
 or neurodegenerative disease drugs that fail.

47:27.140 --> 47:32.140
 So there are lots of trials that fail in later stages,

47:32.180 --> 47:35.180
 which is really catastrophic from the financial perspective.

47:35.180 --> 47:39.540
 So is it the effective, the most effective mechanism?

47:39.540 --> 47:42.700
 Absolutely no, but this is the only one that currently works.

47:44.300 --> 47:47.900
 And I was closely interacting

47:47.900 --> 47:49.260
 with people in pharmaceutical industry.

47:49.260 --> 47:51.340
 I was really fascinated on how sharp

47:51.340 --> 47:55.260
 and what a deep understanding of the domain do they have.

47:55.260 --> 47:57.020
 It's not observation driven.

47:57.020 --> 48:00.220
 There is really a lot of science behind what they do.

48:00.220 --> 48:02.300
 But if you ask me, can machine learning change it,

48:02.300 --> 48:05.300
 I firmly believe yes,

48:05.300 --> 48:07.860
 because even the most experienced chemists

48:07.860 --> 48:11.100
 cannot hold in their memory and understanding

48:11.100 --> 48:12.500
 everything that you can learn

48:12.500 --> 48:15.420
 from millions of molecules and reactions.

48:17.220 --> 48:19.900
 And the space of graphs is a totally new space.

48:19.900 --> 48:22.060
 I mean, it's a really interesting space

48:22.060 --> 48:23.980
 for machine learning to explore, graph generation.

48:23.980 --> 48:26.260
 Yeah, so there are a lot of things that you can do here.

48:26.260 --> 48:28.740
 So we do a lot of work.

48:28.740 --> 48:31.620
 So the first tool that we started with

48:31.620 --> 48:36.300
 was the tool that can predict properties of the molecules.

48:36.300 --> 48:39.420
 So you can just give the molecule and the property.

48:39.420 --> 48:41.340
 It can be by activity property,

48:41.340 --> 48:44.300
 or it can be some other property.

48:44.300 --> 48:46.460
 And you train the molecules

48:46.460 --> 48:50.020
 and you can now take a new molecule

48:50.020 --> 48:52.180
 and predict this property.

48:52.180 --> 48:54.860
 Now, when people started working in this area,

48:54.860 --> 48:55.980
 it is something very simple.

48:55.980 --> 48:58.580
 They do kind of existing fingerprints,

48:58.580 --> 49:00.740
 which is kind of handcrafted features of the molecule.

49:00.740 --> 49:02.980
 When you break the graph to substructures

49:02.980 --> 49:05.980
 and then you run it in a feed forward neural network.

49:05.980 --> 49:08.500
 And what was interesting to see that clearly,

49:08.500 --> 49:11.020
 this was not the most effective way to proceed.

49:11.020 --> 49:14.060
 And you need to have much more complex models

49:14.060 --> 49:16.300
 that can induce a representation,

49:16.300 --> 49:19.220
 which can translate this graph into the embeddings

49:19.220 --> 49:21.300
 and do these predictions.

49:21.300 --> 49:23.220
 So this is one direction.

49:23.220 --> 49:25.260
 Then another direction, which is kind of related

49:25.260 --> 49:29.180
 is not only to stop by looking at the embedding itself,

49:29.180 --> 49:32.780
 but actually modify it to produce better molecules.

49:32.780 --> 49:36.020
 So you can think about it as machine translation

49:36.020 --> 49:38.140
 that you can start with a molecule

49:38.140 --> 49:40.580
 and then there is an improved version of molecule.

49:40.580 --> 49:42.860
 And you can again, with encoder translate it

49:42.860 --> 49:45.380
 into the hidden space and then learn how to modify it

49:45.380 --> 49:49.340
 to improve the in some ways version of the molecules.

49:49.340 --> 49:52.620
 So that's, it's kind of really exciting.

49:52.620 --> 49:54.740
 We already have seen that the property prediction

49:54.740 --> 49:56.140
 works pretty well.

49:56.140 --> 49:59.780
 And now we are generating molecules

49:59.780 --> 50:01.820
 and there is actually labs

50:01.820 --> 50:04.180
 which are manufacturing this molecule.

50:04.180 --> 50:06.340
 So we'll see where it will get us.

50:06.340 --> 50:07.780
 Okay, that's really exciting.

50:07.780 --> 50:08.860
 There's a lot of promise.

50:08.860 --> 50:11.820
 Speaking of machine translation and embeddings,

50:11.820 --> 50:15.580
 I think you have done a lot of really great research

50:15.580 --> 50:17.540
 in NLP, natural language processing.

50:19.260 --> 50:21.540
 Can you tell me your journey through NLP?

50:21.540 --> 50:25.100
 What ideas, problems, approaches were you working on?

50:25.100 --> 50:28.180
 Were you fascinated with, did you explore

50:28.180 --> 50:33.180
 before this magic of deep learning reemerged and after?

50:34.020 --> 50:37.180
 So when I started my work in NLP, it was in 97.

50:38.180 --> 50:39.460
 This was very interesting time.

50:39.460 --> 50:42.620
 It was exactly the time that I came to ACL.

50:43.500 --> 50:46.140
 And at the time I could barely understand English,

50:46.140 --> 50:48.500
 but it was exactly like the transition point

50:48.500 --> 50:53.500
 because half of the papers were really rule based approaches

50:53.500 --> 50:56.180
 where people took more kind of heavy linguistic approaches

50:56.180 --> 51:00.060
 for small domains and try to build up from there.

51:00.060 --> 51:02.220
 And then there were the first generation of papers

51:02.220 --> 51:04.500
 which were corpus based papers.

51:04.500 --> 51:06.420
 And they were very simple in our terms

51:06.420 --> 51:07.900
 when you collect some statistics

51:07.900 --> 51:10.020
 and do prediction based on them.

51:10.020 --> 51:13.100
 And I found it really fascinating that one community

51:13.100 --> 51:18.100
 can think so very differently about the problem.

51:19.220 --> 51:22.820
 And I remember my first paper that I wrote,

51:22.820 --> 51:24.460
 it didn't have a single formula.

51:24.460 --> 51:25.740
 It didn't have evaluation.

51:25.740 --> 51:28.340
 It just had examples of outputs.

51:28.340 --> 51:32.020
 And this was a standard of the field at the time.

51:32.020 --> 51:35.860
 In some ways, I mean, people maybe just started emphasizing

51:35.860 --> 51:38.940
 the empirical evaluation, but for many applications

51:38.940 --> 51:42.780
 like summarization, you just show some examples of outputs.

51:42.780 --> 51:45.460
 And then increasingly you can see that how

51:45.460 --> 51:48.300
 the statistical approaches dominated the field

51:48.300 --> 51:52.100
 and we've seen increased performance

51:52.100 --> 51:56.020
 across many basic tasks.

51:56.020 --> 52:00.420
 The sad part of the story maybe that if you look again

52:00.420 --> 52:05.100
 through this journey, we see that the role of linguistics

52:05.100 --> 52:07.460
 in some ways greatly diminishes.

52:07.460 --> 52:11.580
 And I think that you really need to look

52:11.580 --> 52:14.540
 through the whole proceeding to find one or two papers

52:14.540 --> 52:17.260
 which make some interesting linguistic references.

52:17.260 --> 52:18.100
 It's really big.

52:18.100 --> 52:18.920
 Today, yeah.

52:18.920 --> 52:19.760
 Today, today.

52:19.760 --> 52:20.600
 This was definitely one of the.

52:20.600 --> 52:23.140
 Things like syntactic trees, just even basically

52:23.140 --> 52:26.180
 against our conversation about human understanding

52:26.180 --> 52:30.300
 of language, which I guess what linguistics would be

52:30.300 --> 52:34.300
 structured, hierarchical representing language

52:34.300 --> 52:37.140
 in a way that's human explainable, understandable

52:37.140 --> 52:39.500
 is missing today.

52:39.500 --> 52:42.380
 I don't know if it is, what is explainable

52:42.380 --> 52:43.620
 and understandable.

52:43.620 --> 52:47.360
 In the end, we perform functions and it's okay

52:47.360 --> 52:50.140
 to have machine which performs a function.

52:50.140 --> 52:53.200
 Like when you're thinking about your calculator, correct?

52:53.200 --> 52:56.100
 Your calculator can do calculation very different

52:56.100 --> 52:57.620
 from you would do the calculation,

52:57.620 --> 52:58.860
 but it's very effective in it.

52:58.860 --> 53:02.560
 And this is fine if we can achieve certain tasks

53:02.560 --> 53:05.760
 with high accuracy, doesn't necessarily mean

53:05.760 --> 53:09.300
 that it has to understand it the same way as we understand.

53:09.300 --> 53:11.260
 In some ways, it's even naive to request

53:11.260 --> 53:14.940
 because you have so many other sources of information

53:14.940 --> 53:17.900
 that are absent when you are training your system.

53:17.900 --> 53:19.220
 So it's okay.

53:19.220 --> 53:20.060
 Is it delivered?

53:20.060 --> 53:21.500
 And I would tell you one application

53:21.500 --> 53:22.780
 that is really fascinating.

53:22.780 --> 53:25.060
 In 97, when it came to ACL, there were some papers

53:25.060 --> 53:25.900
 on machine translation.

53:25.900 --> 53:27.440
 They were like primitive.

53:27.440 --> 53:31.060
 Like people were trying really, really simple.

53:31.060 --> 53:34.260
 And the feeling, my feeling was that, you know,

53:34.260 --> 53:36.260
 to make real machine translation system,

53:36.260 --> 53:39.580
 it's like to fly at the moon and build a house there

53:39.580 --> 53:41.580
 and the garden and live happily ever after.

53:41.580 --> 53:42.600
 I mean, it's like impossible.

53:42.600 --> 53:46.740
 I never could imagine that within, you know, 10 years,

53:46.740 --> 53:48.540
 we would already see the system working.

53:48.540 --> 53:51.420
 And now, you know, nobody is even surprised

53:51.420 --> 53:54.420
 to utilize the system on daily basis.

53:54.420 --> 53:56.220
 So this was like a huge, huge progress,

53:56.220 --> 53:57.860
 saying that people for very long time

53:57.860 --> 54:00.820
 tried to solve using other mechanisms.

54:00.820 --> 54:03.220
 And they were unable to solve it.

54:03.220 --> 54:06.140
 That's why coming back to your question about biology,

54:06.140 --> 54:10.800
 that, you know, in linguistics, people try to go this way

54:10.800 --> 54:13.500
 and try to write the syntactic trees

54:13.500 --> 54:17.500
 and try to abstract it and to find the right representation.

54:17.500 --> 54:22.240
 And, you know, they couldn't get very far

54:22.240 --> 54:26.580
 with this understanding while these models using,

54:26.580 --> 54:29.640
 you know, other sources actually capable

54:29.640 --> 54:31.680
 to make a lot of progress.

54:31.680 --> 54:33.960
 Now, I'm not naive to think

54:33.960 --> 54:36.780
 that we are in this paradise space in NLP.

54:36.780 --> 54:38.580
 And sure as you know,

54:38.580 --> 54:40.860
 that when we slightly change the domain

54:40.860 --> 54:42.620
 and when we decrease the amount of training,

54:42.620 --> 54:44.740
 it can do like really bizarre and funny thing.

54:44.740 --> 54:46.500
 But I think it's just a matter

54:46.500 --> 54:48.540
 of improving generalization capacity,

54:48.540 --> 54:51.500
 which is just a technical question.

54:51.500 --> 54:54.340
 Wow, so that's the question.

54:54.340 --> 54:57.720
 How much of language understanding can be solved

54:57.720 --> 54:59.180
 with deep neural networks?

54:59.180 --> 55:03.740
 In your intuition, I mean, it's unknown, I suppose.

55:03.740 --> 55:07.660
 But as we start to creep towards romantic notions

55:07.660 --> 55:10.620
 of the spirit of the Turing test

55:10.620 --> 55:14.220
 and conversation and dialogue

55:14.220 --> 55:18.340
 and something that maybe to me or to us,

55:18.340 --> 55:21.620
 so the humans feels like it needs real understanding.

55:21.620 --> 55:23.500
 How much can that be achieved

55:23.500 --> 55:27.180
 with these neural networks or statistical methods?

55:27.180 --> 55:32.180
 So I guess I am very much driven by the outcomes.

55:33.340 --> 55:35.420
 Can we achieve the performance

55:35.420 --> 55:40.420
 which would be satisfactory for us for different tasks?

55:40.700 --> 55:43.020
 Now, if you again look at machine translation system,

55:43.020 --> 55:46.020
 which are trained on large amounts of data,

55:46.020 --> 55:48.780
 they really can do a remarkable job

55:48.780 --> 55:51.300
 relatively to where they've been a few years ago.

55:51.300 --> 55:54.620
 And if you project into the future,

55:54.620 --> 55:59.380
 if it will be the same speed of improvement, you know,

55:59.380 --> 56:00.220
 this is great.

56:00.220 --> 56:01.060
 Now, does it bother me

56:01.060 --> 56:04.860
 that it's not doing the same translation as we are doing?

56:04.860 --> 56:06.620
 Now, if you go to cognitive science,

56:06.620 --> 56:09.460
 we still don't really understand what we are doing.

56:10.460 --> 56:11.860
 I mean, there are a lot of theories

56:11.860 --> 56:13.840
 and there's obviously a lot of progress and studying,

56:13.840 --> 56:17.540
 but our understanding what exactly goes on in our brains

56:17.540 --> 56:21.020
 when we process language is still not crystal clear

56:21.020 --> 56:25.460
 and precise that we can translate it into machines.

56:25.460 --> 56:29.220
 What does bother me is that, you know,

56:29.220 --> 56:31.700
 again, that machines can be extremely brittle

56:31.700 --> 56:33.980
 when you go out of your comfort zone

56:33.980 --> 56:36.060
 of when there is a distributional shift

56:36.060 --> 56:37.300
 between training and testing.

56:37.300 --> 56:39.020
 And it have been years and years,

56:39.020 --> 56:41.320
 every year when I teach an LP class,

56:41.320 --> 56:43.560
 now show them some examples of translation

56:43.560 --> 56:47.300
 from some newspaper in Hebrew or whatever, it was perfect.

56:47.300 --> 56:51.300
 And then I have a recipe that Tomi Yakel's system

56:51.300 --> 56:53.900
 sent me a while ago and it was written in Finnish

56:53.900 --> 56:55.720
 of Karelian pies.

56:55.720 --> 56:59.280
 And it's just a terrible translation.

56:59.280 --> 57:01.460
 You cannot understand anything what it does.

57:01.460 --> 57:04.180
 It's not like some syntactic mistakes, it's just terrible.

57:04.180 --> 57:07.020
 And year after year, I tried and will translate

57:07.020 --> 57:08.980
 and year after year, it does this terrible work

57:08.980 --> 57:10.980
 because I guess, you know, the recipes

57:10.980 --> 57:14.580
 are not a big part of their training repertoire.

57:14.580 --> 57:19.020
 So, but in terms of outcomes, that's a really clean,

57:19.020 --> 57:20.240
 good way to look at it.

57:21.100 --> 57:23.140
 I guess the question I was asking is,

57:24.060 --> 57:27.700
 do you think, imagine a future,

57:27.700 --> 57:30.540
 do you think the current approaches can pass

57:30.540 --> 57:32.460
 the Turing test in the way,

57:34.700 --> 57:37.060
 in the best possible formulation of the Turing test?

57:37.060 --> 57:39.460
 Which is, would you wanna have a conversation

57:39.460 --> 57:42.340
 with a neural network for an hour?

57:42.340 --> 57:45.820
 Oh God, no, no, there are not that many people

57:45.820 --> 57:48.380
 that I would want to talk for an hour, but.

57:48.380 --> 57:51.500
 There are some people in this world, alive or not,

57:51.500 --> 57:53.260
 that you would like to talk to for an hour.

57:53.260 --> 57:56.700
 Could a neural network achieve that outcome?

57:56.700 --> 57:58.860
 So I think it would be really hard to create

57:58.860 --> 58:02.300
 a successful training set, which would enable it

58:02.300 --> 58:04.980
 to have a conversation, a contextual conversation

58:04.980 --> 58:05.820
 for an hour.

58:05.820 --> 58:08.140
 Do you think it's a problem of data, perhaps?

58:08.140 --> 58:09.940
 I think in some ways it's not a problem of data,

58:09.940 --> 58:13.620
 it's a problem both of data and the problem of

58:13.620 --> 58:15.780
 the way we're training our systems,

58:15.780 --> 58:18.060
 their ability to truly, to generalize,

58:18.060 --> 58:19.300
 to be very compositional.

58:19.300 --> 58:23.220
 In some ways it's limited in the current capacity,

58:23.220 --> 58:27.980
 at least we can translate well,

58:27.980 --> 58:32.540
 we can find information well, we can extract information.

58:32.540 --> 58:35.180
 So there are many capacities in which it's doing very well.

58:35.180 --> 58:38.000
 And you can ask me, would you trust the machine

58:38.000 --> 58:39.820
 to translate for you and use it as a source?

58:39.820 --> 58:42.580
 I would say absolutely, especially if we're talking about

58:42.580 --> 58:45.660
 newspaper data or other data which is in the realm

58:45.660 --> 58:47.900
 of its own training set, I would say yes.

58:48.900 --> 58:52.900
 But having conversations with the machine,

58:52.900 --> 58:56.460
 it's not something that I would choose to do.

58:56.460 --> 58:59.420
 But I would tell you something, talking about Turing tests

58:59.420 --> 59:02.940
 and about all this kind of ELISA conversations,

59:02.940 --> 59:05.540
 I remember visiting Tencent in China

59:05.540 --> 59:07.620
 and they have this chat board and they claim

59:07.620 --> 59:10.780
 there is really humongous amount of the local population

59:10.780 --> 59:12.940
 which for hours talks to the chat board.

59:12.940 --> 59:15.340
 To me it was, I cannot believe it,

59:15.340 --> 59:18.000
 but apparently it's documented that there are some people

59:18.000 --> 59:20.760
 who enjoy this conversation.

59:20.760 --> 59:24.540
 And it brought to me another MIT story

59:24.540 --> 59:26.980
 about ELISA and Weisenbaum.

59:26.980 --> 59:29.340
 I don't know if you're familiar with the story.

59:29.340 --> 59:31.020
 So Weisenbaum was a professor at MIT

59:31.020 --> 59:32.580
 and when he developed this ELISA,

59:32.580 --> 59:34.620
 which was just doing string matching,

59:34.620 --> 59:38.540
 very trivial, like restating of what you said

59:38.540 --> 59:41.260
 with very few rules, no syntax.

59:41.260 --> 59:43.740
 Apparently there were secretaries at MIT

59:43.740 --> 59:48.180
 that would sit for hours and converse with this trivial thing

59:48.180 --> 59:50.180
 and at the time there was no beautiful interfaces

59:50.180 --> 59:51.820
 so you actually need to go through the pain

59:51.820 --> 59:53.540
 of communicating.

59:53.540 --> 59:56.940
 And Weisenbaum himself was so horrified by this phenomenon

59:56.940 --> 59:59.300
 that people can believe enough to the machine

59:59.300 --> 1:00:00.820
 that you just need to give them the hint

1:00:00.820 --> 1:00:03.940
 that machine understands you and you can complete the rest

1:00:03.940 --> 1:00:05.420
 that he kind of stopped this research

1:00:05.420 --> 1:00:08.660
 and went into kind of trying to understand

1:00:08.660 --> 1:00:11.480
 what this artificial intelligence can do to our brains.

1:00:12.740 --> 1:00:14.380
 So my point is, you know,

1:00:14.380 --> 1:00:19.300
 how much, it's not how good is the technology,

1:00:19.300 --> 1:00:22.620
 it's how ready we are to believe

1:00:22.620 --> 1:00:25.580
 that it delivers the goods that we are trying to get.

1:00:25.580 --> 1:00:27.200
 That's a really beautiful way to put it.

1:00:27.200 --> 1:00:29.800
 I, by the way, I'm not horrified by that possibility,

1:00:29.800 --> 1:00:33.140
 but inspired by it because,

1:00:33.140 --> 1:00:35.920
 I mean, human connection,

1:00:35.920 --> 1:00:38.220
 whether it's through language or through love,

1:00:39.860 --> 1:00:44.860
 it seems like it's very amenable to machine learning

1:00:44.900 --> 1:00:49.340
 and the rest is just challenges of psychology.

1:00:49.340 --> 1:00:52.460
 Like you said, the secretaries who enjoy spending hours.

1:00:52.460 --> 1:00:55.020
 I would say I would describe most of our lives

1:00:55.020 --> 1:00:58.020
 as enjoying spending hours with those we love

1:00:58.020 --> 1:01:00.820
 for very silly reasons.

1:01:00.820 --> 1:01:02.780
 All we're doing is keyword matching as well.

1:01:02.780 --> 1:01:05.100
 So I'm not sure how much intelligence

1:01:05.100 --> 1:01:08.140
 we exhibit to each other with the people we love

1:01:08.140 --> 1:01:09.820
 that we're close with.

1:01:09.820 --> 1:01:12.660
 So it's a very interesting point

1:01:12.660 --> 1:01:16.020
 of what it means to pass the Turing test with language.

1:01:16.020 --> 1:01:16.860
 I think you're right.

1:01:16.860 --> 1:01:18.220
 In terms of conversation,

1:01:18.220 --> 1:01:20.180
 I think machine translation

1:01:21.420 --> 1:01:24.420
 has very clear performance and improvement, right?

1:01:24.420 --> 1:01:28.020
 What it means to have a fulfilling conversation

1:01:28.020 --> 1:01:32.660
 is very person dependent and context dependent

1:01:32.660 --> 1:01:33.580
 and so on.

1:01:33.580 --> 1:01:36.340
 That's, yeah, it's very well put.

1:01:36.340 --> 1:01:40.740
 But in your view, what's a benchmark in natural language,

1:01:40.740 --> 1:01:43.640
 a test that's just out of reach right now,

1:01:43.640 --> 1:01:46.020
 but we might be able to, that's exciting.

1:01:46.020 --> 1:01:49.100
 Is it in perfecting machine translation

1:01:49.100 --> 1:01:51.900
 or is there other, is it summarization?

1:01:51.900 --> 1:01:52.740
 What's out there just out of reach?

1:01:52.740 --> 1:01:55.820
 I think it goes across specific application.

1:01:55.820 --> 1:01:59.500
 It's more about the ability to learn from few examples

1:01:59.500 --> 1:02:03.300
 for real, what we call few short learning and all these cases

1:02:03.300 --> 1:02:05.940
 because the way we publish these papers today,

1:02:05.940 --> 1:02:09.900
 we say, if we have like naively, we get 55,

1:02:09.900 --> 1:02:12.500
 but now we had a few example and we can move to 65.

1:02:12.500 --> 1:02:13.540
 None of these methods

1:02:13.540 --> 1:02:15.980
 actually are realistically doing anything useful.

1:02:15.980 --> 1:02:18.540
 You cannot use them today.

1:02:18.540 --> 1:02:23.540
 And the ability to be able to generalize and to move

1:02:25.460 --> 1:02:28.940
 or to be autonomous in finding the data

1:02:28.940 --> 1:02:30.300
 that you need to learn,

1:02:31.340 --> 1:02:34.280
 to be able to perfect new tasks or new language,

1:02:35.300 --> 1:02:38.100
 this is an area where I think we really need

1:02:39.200 --> 1:02:43.020
 to move forward to and we are not yet there.

1:02:43.020 --> 1:02:45.060
 Are you at all excited,

1:02:45.060 --> 1:02:46.540
 curious by the possibility

1:02:46.540 --> 1:02:48.520
 of creating human level intelligence?

1:02:49.900 --> 1:02:52.540
 Is this, cause you've been very in your discussion.

1:02:52.540 --> 1:02:54.340
 So if we look at oncology,

1:02:54.340 --> 1:02:58.100
 you're trying to use machine learning to help the world

1:02:58.100 --> 1:02:59.700
 in terms of alleviating suffering.

1:02:59.700 --> 1:03:02.340
 If you look at natural language processing,

1:03:02.340 --> 1:03:05.300
 you're focused on the outcomes of improving practical things

1:03:05.300 --> 1:03:06.820
 like machine translation.

1:03:06.820 --> 1:03:09.880
 But human level intelligence is a thing

1:03:09.880 --> 1:03:13.800
 that our civilization has dreamed about creating,

1:03:13.800 --> 1:03:15.740
 super human level intelligence.

1:03:15.740 --> 1:03:16.940
 Do you think about this?

1:03:16.940 --> 1:03:19.040
 Do you think it's at all within our reach?

1:03:20.380 --> 1:03:22.660
 So as you said yourself, Elie,

1:03:22.660 --> 1:03:26.140
 talking about how do you perceive

1:03:26.140 --> 1:03:28.980
 our communications with each other,

1:03:28.980 --> 1:03:31.940
 that we're matching keywords and certain behaviors

1:03:31.940 --> 1:03:33.020
 and so on.

1:03:33.020 --> 1:03:36.860
 So at the end, whenever one assesses,

1:03:36.860 --> 1:03:38.680
 let's say relations with another person,

1:03:38.680 --> 1:03:41.460
 you have separate kind of measurements and outcomes

1:03:41.460 --> 1:03:43.620
 inside your head that determine

1:03:43.620 --> 1:03:45.860
 what is the status of the relation.

1:03:45.860 --> 1:03:48.580
 So one way, this is this classical level,

1:03:48.580 --> 1:03:49.600
 what is the intelligence?

1:03:49.600 --> 1:03:51.860
 Is it the fact that now we are gonna do the same way

1:03:51.860 --> 1:03:52.940
 as human is doing,

1:03:52.940 --> 1:03:55.500
 when we don't even understand what the human is doing?

1:03:55.500 --> 1:03:59.100
 Or we now have an ability to deliver these outcomes,

1:03:59.100 --> 1:04:01.300
 but not in one area, not in NLP,

1:04:01.300 --> 1:04:03.940
 not just to translate or just to answer questions,

1:04:03.940 --> 1:04:05.380
 but across many, many areas

1:04:05.380 --> 1:04:08.100
 that we can achieve the functionalities

1:04:08.100 --> 1:04:11.060
 that humans can achieve with their ability to learn

1:04:11.060 --> 1:04:12.380
 and do other things.

1:04:12.380 --> 1:04:15.500
 I think this is, and this we can actually measure

1:04:15.500 --> 1:04:17.560
 how far we are.

1:04:17.560 --> 1:04:21.580
 And that's what makes me excited that we,

1:04:21.580 --> 1:04:23.780
 in my lifetime, at least so far what we've seen,

1:04:23.780 --> 1:04:25.840
 it's like tremendous progress

1:04:25.840 --> 1:04:28.700
 across these different functionalities.

1:04:28.700 --> 1:04:32.260
 And I think it will be really exciting

1:04:32.260 --> 1:04:35.540
 to see where we will be.

1:04:35.540 --> 1:04:39.300
 And again, one way to think about it,

1:04:39.300 --> 1:04:41.820
 there are machines which are improving their functionality.

1:04:41.820 --> 1:04:44.940
 Another one is to think about us with our brains,

1:04:44.940 --> 1:04:46.420
 which are imperfect,

1:04:46.420 --> 1:04:51.420
 how they can be accelerated by this technology

1:04:51.420 --> 1:04:55.900
 as it becomes stronger and stronger.

1:04:55.900 --> 1:04:57.260
 Coming back to another book

1:04:57.260 --> 1:05:01.060
 that I love, Flowers for Algernon.

1:05:01.060 --> 1:05:02.100
 Have you read this book?

1:05:02.100 --> 1:05:02.940
 Yes.

1:05:02.940 --> 1:05:05.700
 So there is this point that the patient gets

1:05:05.700 --> 1:05:07.980
 this miracle cure, which changes his brain.

1:05:07.980 --> 1:05:11.020
 And all of a sudden they see life in a different way

1:05:11.020 --> 1:05:13.300
 and can do certain things better,

1:05:13.300 --> 1:05:14.860
 but certain things much worse.

1:05:14.860 --> 1:05:19.860
 So you can imagine this kind of computer augmented cognition

1:05:22.400 --> 1:05:24.800
 where it can bring you that now in the same way

1:05:24.800 --> 1:05:28.120
 as the cars enable us to get to places

1:05:28.120 --> 1:05:30.080
 where we've never been before,

1:05:30.080 --> 1:05:31.640
 can we think differently?

1:05:31.640 --> 1:05:33.600
 Can we think faster?

1:05:33.600 --> 1:05:36.680
 And we already see a lot of it happening

1:05:36.680 --> 1:05:38.260
 in how it impacts us,

1:05:38.260 --> 1:05:42.200
 but I think we have a long way to go there.

1:05:42.200 --> 1:05:45.040
 So that's sort of artificial intelligence

1:05:45.040 --> 1:05:47.280
 and technology affecting our,

1:05:47.280 --> 1:05:50.440
 augmenting our intelligence as humans.

1:05:50.440 --> 1:05:55.440
 Yesterday, a company called Neuralink announced,

1:05:55.520 --> 1:05:56.800
 they did this whole demonstration.

1:05:56.800 --> 1:05:57.980
 I don't know if you saw it.

1:05:57.980 --> 1:06:01.000
 It's, they demonstrated brain computer,

1:06:01.000 --> 1:06:02.680
 brain machine interface,

1:06:02.680 --> 1:06:06.360
 where there's like a sewing machine for the brain.

1:06:06.360 --> 1:06:11.120
 Do you, you know, a lot of that is quite out there

1:06:11.120 --> 1:06:14.040
 in terms of things that some people would say

1:06:14.040 --> 1:06:16.340
 are impossible, but they're dreamers

1:06:16.340 --> 1:06:18.080
 and want to engineer systems like that.

1:06:18.080 --> 1:06:20.360
 Do you see, based on what you just said,

1:06:20.360 --> 1:06:23.820
 a hope for that more direct interaction with the brain?

1:06:25.120 --> 1:06:27.040
 I think there are different ways.

1:06:27.040 --> 1:06:29.000
 One is a direct interaction with the brain.

1:06:29.000 --> 1:06:30.900
 And again, there are lots of companies

1:06:30.900 --> 1:06:32.280
 that work in this space

1:06:32.280 --> 1:06:35.080
 and I think there will be a lot of developments.

1:06:35.080 --> 1:06:36.600
 But I'm just thinking that many times

1:06:36.600 --> 1:06:39.080
 we are not aware of our feelings,

1:06:39.080 --> 1:06:41.400
 of motivation, what drives us.

1:06:41.400 --> 1:06:44.200
 Like, let me give you a trivial example, our attention.

1:06:45.520 --> 1:06:47.260
 There are a lot of studies that demonstrate

1:06:47.260 --> 1:06:49.200
 that it takes a while to a person to understand

1:06:49.200 --> 1:06:51.080
 that they are not attentive anymore.

1:06:51.080 --> 1:06:52.160
 And we know that there are people

1:06:52.160 --> 1:06:54.520
 who really have strong capacity to hold attention.

1:06:54.520 --> 1:06:57.080
 There are other end of the spectrum people with ADD

1:06:57.080 --> 1:06:58.800
 and other issues that they have problem

1:06:58.800 --> 1:07:00.760
 to regulate their attention.

1:07:00.760 --> 1:07:03.520
 Imagine to yourself that you have like a cognitive aid

1:07:03.520 --> 1:07:06.280
 that just alerts you based on your gaze,

1:07:06.280 --> 1:07:09.280
 that your attention is now not on what you are doing.

1:07:09.280 --> 1:07:10.560
 And instead of writing a paper,

1:07:10.560 --> 1:07:12.760
 you're now dreaming of what you're gonna do in the evening.

1:07:12.760 --> 1:07:16.360
 So even this kind of simple measurement things,

1:07:16.360 --> 1:07:17.840
 how they can change us.

1:07:17.840 --> 1:07:22.400
 And I see it even in simple ways with myself.

1:07:22.400 --> 1:07:26.480
 I have my zone app that I got in MIT gym.

1:07:26.480 --> 1:07:28.800
 It kind of records, you know, how much did you run

1:07:28.800 --> 1:07:29.800
 and you have some points

1:07:29.800 --> 1:07:32.880
 and you can get some status, whatever.

1:07:32.880 --> 1:07:35.840
 Like, I said, what is this ridiculous thing?

1:07:35.840 --> 1:07:38.800
 Who would ever care about some status in some app?

1:07:38.800 --> 1:07:39.640
 Guess what?

1:07:39.640 --> 1:07:41.560
 So to maintain the status,

1:07:41.560 --> 1:07:44.640
 you have to do set a number of points every month.

1:07:44.640 --> 1:07:48.040
 And not only is that I do it every single month

1:07:48.040 --> 1:07:50.560
 for the last 18 months,

1:07:50.560 --> 1:07:54.160
 it went to the point that I was injured.

1:07:54.160 --> 1:07:56.160
 And when I could run again,

1:07:58.120 --> 1:08:02.560
 in two days, I did like some humongous amount of running

1:08:02.560 --> 1:08:04.080
 just to complete the points.

1:08:04.080 --> 1:08:05.920
 It was like really not safe.

1:08:05.920 --> 1:08:08.440
 It was like, I'm not gonna lose my status

1:08:08.440 --> 1:08:10.240
 because I want to get there.

1:08:10.240 --> 1:08:13.320
 So you can already see that this direct measurement

1:08:13.320 --> 1:08:15.160
 and the feedback is, you know,

1:08:15.160 --> 1:08:16.320
 we're looking at video games

1:08:16.320 --> 1:08:18.720
 and see why, you know, the addiction aspect of it,

1:08:18.720 --> 1:08:21.200
 but you can imagine that the same idea can be expanded

1:08:21.200 --> 1:08:23.640
 to many other areas of our life.

1:08:23.640 --> 1:08:25.960
 When we really can get feedback

1:08:25.960 --> 1:08:28.480
 and imagine in your case in relations,

1:08:29.880 --> 1:08:31.240
 when we are doing keyword matching,

1:08:31.240 --> 1:08:36.120
 imagine that the person who is generating the keywords,

1:08:36.120 --> 1:08:37.720
 that person gets direct feedback

1:08:37.720 --> 1:08:39.560
 before the whole thing explodes.

1:08:39.560 --> 1:08:42.000
 Is it maybe at this happy point,

1:08:42.000 --> 1:08:44.000
 we are going in the wrong direction.

1:08:44.000 --> 1:08:48.040
 Maybe it will be really a behavior modifying moment.

1:08:48.040 --> 1:08:51.360
 So yeah, it's a relationship management too.

1:08:51.360 --> 1:08:54.200
 So yeah, that's a fascinating whole area

1:08:54.200 --> 1:08:56.120
 of psychology actually as well,

1:08:56.120 --> 1:08:58.240
 of seeing how our behavior has changed

1:08:58.240 --> 1:09:01.840
 with basically all human relations now have

1:09:01.840 --> 1:09:06.200
 other nonhuman entities helping us out.

1:09:06.200 --> 1:09:09.440
 So you teach a large,

1:09:09.440 --> 1:09:12.600
 a huge machine learning course here at MIT.

1:09:14.000 --> 1:09:15.360
 I can ask you a million questions,

1:09:15.360 --> 1:09:17.560
 but you've seen a lot of students.

1:09:17.560 --> 1:09:20.920
 What ideas do students struggle with the most

1:09:20.920 --> 1:09:23.920
 as they first enter this world of machine learning?

1:09:23.920 --> 1:09:26.520
 Actually, this year was the first time

1:09:26.520 --> 1:09:28.480
 I started teaching a small machine learning class.

1:09:28.480 --> 1:09:31.160
 And it came as a result of what I saw

1:09:31.160 --> 1:09:34.640
 in my big machine learning class that Tomi Yakel and I built

1:09:34.640 --> 1:09:36.640
 maybe six years ago.

1:09:38.040 --> 1:09:40.360
 What we've seen that as this area become more

1:09:40.360 --> 1:09:43.440
 and more popular, more and more people at MIT

1:09:43.440 --> 1:09:45.360
 want to take this class.

1:09:45.360 --> 1:09:48.320
 And while we designed it for computer science majors,

1:09:48.320 --> 1:09:50.760
 there were a lot of people who really are interested

1:09:50.760 --> 1:09:52.600
 to learn it, but unfortunately,

1:09:52.600 --> 1:09:55.720
 their background was not enabling them

1:09:55.720 --> 1:09:57.200
 to do well in the class.

1:09:57.200 --> 1:09:59.360
 And many of them associated machine learning

1:09:59.360 --> 1:10:01.360
 with the word struggle and failure,

1:10:02.480 --> 1:10:04.640
 primarily for non majors.

1:10:04.640 --> 1:10:06.840
 And that's why we actually started a new class

1:10:06.840 --> 1:10:10.800
 which we call machine learning from algorithms to modeling,

1:10:10.800 --> 1:10:15.000
 which emphasizes more the modeling aspects of it

1:10:15.000 --> 1:10:20.000
 and focuses on, it has majors and non majors.

1:10:20.000 --> 1:10:23.480
 So we kind of try to extract the relevant parts

1:10:23.480 --> 1:10:25.560
 and make it more accessible,

1:10:25.560 --> 1:10:27.800
 because the fact that we're teaching 20 classifiers

1:10:27.800 --> 1:10:29.240
 in standard machine learning class,

1:10:29.240 --> 1:10:32.200
 it's really a big question to really need it.

1:10:32.200 --> 1:10:34.520
 But it was interesting to see this

1:10:34.520 --> 1:10:36.480
 from first generation of students,

1:10:36.480 --> 1:10:39.080
 when they came back from their internships

1:10:39.080 --> 1:10:42.320
 and from their jobs,

1:10:42.320 --> 1:10:45.560
 what different and exciting things they can do.

1:10:45.560 --> 1:10:47.600
 I would never think that you can even apply

1:10:47.600 --> 1:10:50.800
 machine learning to, some of them are like matching,

1:10:50.800 --> 1:10:53.480
 the relations and other things like variety.

1:10:53.480 --> 1:10:56.080
 Everything is amenable as the machine learning.

1:10:56.080 --> 1:10:58.320
 That actually brings up an interesting point

1:10:58.320 --> 1:11:00.680
 of computer science in general.

1:11:00.680 --> 1:11:03.520
 It almost seems, maybe I'm crazy,

1:11:03.520 --> 1:11:06.520
 but it almost seems like everybody needs to learn

1:11:06.520 --> 1:11:08.160
 how to program these days.

1:11:08.160 --> 1:11:11.400
 If you're 20 years old, or if you're starting school,

1:11:11.400 --> 1:11:14.200
 even if you're an English major,

1:11:14.200 --> 1:11:19.200
 it seems like programming unlocks so much possibility

1:11:20.480 --> 1:11:21.880
 in this world.

1:11:21.880 --> 1:11:25.000
 So when you interacted with those non majors,

1:11:25.000 --> 1:11:30.000
 is there skills that they were simply lacking at the time

1:11:30.280 --> 1:11:33.000
 that you wish they had and that they learned

1:11:33.000 --> 1:11:34.680
 in high school and so on?

1:11:34.680 --> 1:11:37.520
 Like how should education change

1:11:37.520 --> 1:11:41.320
 in this computerized world that we live in?

1:11:41.320 --> 1:11:44.320
 I think because I knew that there is a Python component

1:11:44.320 --> 1:11:47.000
 in the class, their Python skills were okay

1:11:47.000 --> 1:11:49.160
 and the class isn't really heavy on programming.

1:11:49.160 --> 1:11:52.400
 They primarily kind of add parts to the programs.

1:11:52.400 --> 1:11:55.440
 I think it was more of the mathematical barriers

1:11:55.440 --> 1:11:58.200
 and the class, again, with the design on the majors

1:11:58.200 --> 1:12:01.200
 was using the notation, like big O for complexity

1:12:01.200 --> 1:12:04.520
 and others, people who come from different backgrounds

1:12:04.520 --> 1:12:05.800
 just don't have it in the lexical,

1:12:05.800 --> 1:12:09.120
 so necessarily very challenging notion,

1:12:09.120 --> 1:12:11.480
 but they were just not aware.

1:12:12.360 --> 1:12:16.240
 So I think that kind of linear algebra and probability,

1:12:16.240 --> 1:12:19.120
 the basics, the calculus, multivariate calculus,

1:12:19.120 --> 1:12:20.840
 things that can help.

1:12:20.840 --> 1:12:23.520
 What advice would you give to students

1:12:23.520 --> 1:12:25.280
 interested in machine learning,

1:12:25.280 --> 1:12:29.240
 interested, you've talked about detecting,

1:12:29.240 --> 1:12:31.360
 curing cancer, drug design,

1:12:31.360 --> 1:12:34.520
 if they want to get into that field, what should they do?

1:12:36.320 --> 1:12:39.040
 Get into it and succeed as researchers

1:12:39.040 --> 1:12:42.080
 and entrepreneurs.

1:12:43.320 --> 1:12:45.240
 The first good piece of news is that right now

1:12:45.240 --> 1:12:47.400
 there are lots of resources

1:12:47.400 --> 1:12:50.160
 that are created at different levels

1:12:50.160 --> 1:12:54.800
 and you can find online in your school classes

1:12:54.800 --> 1:12:57.560
 which are more mathematical, more applied and so on.

1:12:57.560 --> 1:13:01.320
 So you can find a kind of a preacher

1:13:01.320 --> 1:13:02.760
 which preaches in your own language

1:13:02.760 --> 1:13:04.520
 where you can enter the field

1:13:04.520 --> 1:13:06.720
 and you can make many different types of contribution

1:13:06.720 --> 1:13:09.640
 depending of what is your strengths.

1:13:10.760 --> 1:13:13.720
 And the second point, I think it's really important

1:13:13.720 --> 1:13:18.160
 to find some area which you really care about

1:13:18.160 --> 1:13:20.240
 and it can motivate your learning

1:13:20.240 --> 1:13:22.640
 and it can be for somebody curing cancer

1:13:22.640 --> 1:13:25.360
 or doing self driving cars or whatever,

1:13:25.360 --> 1:13:29.680
 but to find an area where there is data

1:13:29.680 --> 1:13:31.320
 where you believe there are strong patterns

1:13:31.320 --> 1:13:33.600
 and we should be doing it and we're still not doing it

1:13:33.600 --> 1:13:35.280
 or you can do it better

1:13:35.280 --> 1:13:39.680
 and just start there and see where it can bring you.

1:13:40.800 --> 1:13:45.600
 So you've been very successful in many directions in life,

1:13:46.480 --> 1:13:48.840
 but you also mentioned Flowers of Argonon.

1:13:51.200 --> 1:13:53.840
 And I think I've read or listened to you mention somewhere

1:13:53.840 --> 1:13:55.360
 that researchers often get lost

1:13:55.360 --> 1:13:56.720
 in the details of their work.

1:13:56.720 --> 1:14:00.240
 This is per our original discussion with cancer and so on

1:14:00.240 --> 1:14:02.200
 and don't look at the bigger picture,

1:14:02.200 --> 1:14:05.320
 bigger questions of meaning and so on.

1:14:05.320 --> 1:14:07.440
 So let me ask you the impossible question

1:14:08.640 --> 1:14:11.560
 of what's the meaning of this thing,

1:14:11.560 --> 1:14:16.560
 of life, of your life, of research.

1:14:16.720 --> 1:14:21.440
 Why do you think we descendant of great apes

1:14:21.440 --> 1:14:24.480
 are here on this spinning ball?

1:14:26.800 --> 1:14:30.320
 You know, I don't think that I have really a global answer.

1:14:30.320 --> 1:14:32.800
 You know, maybe that's why I didn't go to humanities

1:14:33.760 --> 1:14:36.480
 and I didn't take humanities classes in my undergrad.

1:14:39.480 --> 1:14:43.560
 But the way I'm thinking about it,

1:14:43.560 --> 1:14:48.200
 each one of us inside of them have their own set of,

1:14:48.200 --> 1:14:51.120
 you know, things that we believe are important.

1:14:51.120 --> 1:14:53.360
 And it just happens that we are busy

1:14:53.360 --> 1:14:56.240
 with achieving various goals, busy listening to others

1:14:56.240 --> 1:15:00.960
 and to kind of try to conform and to be part of the crowd,

1:15:00.960 --> 1:15:03.680
 that we don't listen to that part.

1:15:04.600 --> 1:15:09.600
 And, you know, we all should find some time to understand

1:15:09.600 --> 1:15:11.840
 what is our own individual missions.

1:15:11.840 --> 1:15:14.080
 And we may have very different missions

1:15:14.080 --> 1:15:18.200
 and to make sure that while we are running 10,000 things,

1:15:18.200 --> 1:15:21.920
 we are not, you know, missing out

1:15:21.920 --> 1:15:26.800
 and we're putting all the resources to satisfy

1:15:26.800 --> 1:15:28.440
 our own mission.

1:15:28.440 --> 1:15:32.400
 And if I look over my time, when I was younger,

1:15:32.400 --> 1:15:35.000
 most of these missions, you know,

1:15:35.000 --> 1:15:38.600
 I was primarily driven by the external stimulus,

1:15:38.600 --> 1:15:41.520
 you know, to achieve this or to be that.

1:15:41.520 --> 1:15:46.520
 And now a lot of what I do is driven by really thinking

1:15:47.640 --> 1:15:51.360
 what is important for me to achieve independently

1:15:51.360 --> 1:15:55.160
 of the external recognition.

1:15:55.160 --> 1:16:00.080
 And, you know, I don't mind to be viewed in certain ways.

1:16:01.400 --> 1:16:05.760
 The most important thing for me is to be true to myself,

1:16:05.760 --> 1:16:07.520
 to what I think is right.

1:16:07.520 --> 1:16:08.680
 How long did it take?

1:16:08.680 --> 1:16:13.240
 How hard was it to find the you that you have to be true to?

1:16:14.160 --> 1:16:15.520
 So it takes time.

1:16:15.520 --> 1:16:17.760
 And even now, sometimes, you know,

1:16:17.760 --> 1:16:20.880
 the vanity and the triviality can take, you know.

1:16:20.880 --> 1:16:22.560
 At MIT.

1:16:22.560 --> 1:16:25.080
 Yeah, it can everywhere, you know,

1:16:25.080 --> 1:16:26.960
 it's just the vanity at MIT is different,

1:16:26.960 --> 1:16:28.160
 the vanity in different places,

1:16:28.160 --> 1:16:30.920
 but we all have our piece of vanity.

1:16:30.920 --> 1:16:35.920
 But I think actually for me, many times the place

1:16:38.720 --> 1:16:43.720
 to get back to it is, you know, when I'm alone

1:16:43.800 --> 1:16:45.800
 and also when I read.

1:16:45.800 --> 1:16:47.760
 And I think by selecting the right books,

1:16:47.760 --> 1:16:52.760
 you can get the right questions and learn from what you read.

1:16:54.880 --> 1:16:58.080
 So, but again, it's not perfect.

1:16:58.080 --> 1:17:02.040
 Like vanity sometimes dominates.

1:17:02.040 --> 1:17:04.800
 Well, that's a beautiful way to end.

1:17:04.800 --> 1:17:06.400
 Thank you so much for talking today.

1:17:06.400 --> 1:17:07.240
 Thank you.

1:17:07.240 --> 1:17:08.080
 That was fun.

1:17:08.080 --> 1:17:28.080
 That was fun.