vtt/episode_040_small.vtt

WEBVTT

00:00.000 --> 00:03.240
 The following is a conversation with Regina Barsley.

00:03.240 --> 00:06.680
 She's a professor at MIT and a world class researcher

00:06.680 --> 00:09.120
 in natural language processing and applications

00:09.120 --> 00:12.440
 of deep learning to chemistry and oncology,

00:12.440 --> 00:16.240
 or the use of deep learning for early diagnosis, prevention,

00:16.240 --> 00:18.320
 and treatment of cancer.

00:18.320 --> 00:21.000
 She has also been recognized for a teaching

00:21.000 --> 00:24.680
 of several successful AI related courses at MIT,

00:24.680 --> 00:27.360
 including the popular introduction to machine

00:27.360 --> 00:28.920
 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 5,000 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 Freedman, spelled F R I D M A N.

00:43.760 --> 00:47.720
 And now here's my conversation with Regina Barsley.

00:48.800 --> 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 --> 00:57.920
 that a friend of yours teaches just out of curiosity

00:57.920 --> 01:00.320
 because I couldn't find anything on it.

01:00.320 --> 01:04.480
 Are there books or ideas that had profound impact

01:04.480 --> 01:07.760
 on your life journey, books and ideas perhaps outside

01:07.760 --> 01:10.880
 of computer science and the technical fields?

01:11.840 --> 01:14.800
 I think because I'm spending a lot of my time at MIT

01:14.800 --> 01:18.360
 and previously in other institutions where I was a student,

01:18.360 --> 01:21.080
 I have a limited ability to interact with people.

01:21.080 --> 01:22.720
 So a lot of what I know about the world

01:22.720 --> 01:24.280
 actually comes from books.

01:25.320 --> 01:27.280
 And there were quite a number of books

01:27.280 --> 01:31.400
 that had profound impact on me and how I view the world.

01:31.400 --> 01:35.840
 Let me just give you one example of such a book.

01:35.840 --> 01:39.680
 I've maybe a year ago read a book

01:39.680 --> 01:42.520
 called The Emperor of All Melodies.

01:42.520 --> 01:45.760
 It's a book about, it's kind of a history of science book

01:45.760 --> 01:50.760
 on how the treatments and drugs for cancer were developed.

01:50.760 --> 01:54.040
 And that book, despite the fact that I am

01:54.040 --> 01:57.040
 in the business of science, really opened my eyes

01:57.040 --> 02:02.040
 on how imprecise and imperfect the discovery process is

02:03.080 --> 02:05.920
 and how imperfect our current solutions.

02:07.000 --> 02:11.080
 And what makes science succeed and be implemented

02:11.080 --> 02:14.120
 and sometimes it's actually not the strengths of the idea

02:14.120 --> 02:17.480
 but devotion of the person who wants to see it implemented.

02:17.480 --> 02:19.800
 So this is one of the books that, you know,

02:19.800 --> 02:22.320
 at least for the last year quite changed the way

02:22.320 --> 02:24.920
 I'm thinking about scientific process

02:24.920 --> 02:26.720
 just from the historical perspective.

02:26.720 --> 02:31.720
 And what do I need to do to make my ideas really implemented?

02:33.440 --> 02:36.040
 Let me give you an example of a book

02:36.040 --> 02:39.560
 which is not kind of, which is a fiction book.

02:40.600 --> 02:43.080
 It's a book called Americana.

02:44.440 --> 02:48.760
 And this is a book about a young female student

02:48.760 --> 02:53.240
 who comes from Africa to study in the United States.

02:53.240 --> 02:57.720
 And it describes her past, you know, within her studies

02:57.720 --> 03:02.000
 and her life transformation that, you know,

03:02.000 --> 03:06.560
 in a new country and kind of adaptation to a new culture.

03:06.560 --> 03:09.520
 And when I read this book,

03:09.520 --> 03:13.520
 I saw myself in many different points of it.

03:14.600 --> 03:19.600
 But it also kind of gave me the lens on different events

03:20.160 --> 03:22.040
 and some events that I never actually paid attention

03:22.040 --> 03:24.720
 to one of the funny stories in this book

03:24.720 --> 03:29.720
 is how she arrives to her new college

03:30.400 --> 03:32.880
 and she starts speaking in English

03:32.880 --> 03:35.680
 and she had this beautiful British accent

03:35.680 --> 03:39.840
 because that's how she was educated in her country.

03:39.840 --> 03:40.960
 This is not my case.

03:40.960 --> 03:45.400
 And then she notices that the person who talks to her,

03:45.400 --> 03:47.160
 you know, talks to her in a very funny way,

03:47.160 --> 03:48.280
 in a very slow way.

03:48.280 --> 03:50.800
 And she's thinking that this woman is disabled

03:50.800 --> 03:54.480
 and she's also trying to kind of to accommodate her.

03:54.480 --> 03:55.400
 And then after a while,

03:55.400 --> 03:57.560
 when she finishes her discussion with this officer

03:57.560 --> 04:01.040
 from her college, she sees how she interacts

04:01.040 --> 04:03.000
 with the other students, with American students

04:03.000 --> 04:08.000
 and she discovers that actually she talked to her this way

04:09.720 --> 04:12.200
 because she saw that she doesn't understand English.

04:12.200 --> 04:15.080
 And he thought, wow, this is a funny experience.

04:15.080 --> 04:18.000
 And literally within few weeks,

04:18.000 --> 04:21.880
 I went to LA to a conference

04:21.880 --> 04:24.320
 and they asked somebody in an airport, you know,

04:24.320 --> 04:26.560
 how to find like a cab or something.

04:26.560 --> 04:29.360
 And then I noticed that this person is talking

04:29.360 --> 04:30.280
 in a very strange way.

04:30.280 --> 04:32.080
 And my first thought was that this person

04:32.080 --> 04:35.480
 have some, you know, pronunciation issues or something.

04:35.480 --> 04:37.160
 And I'm trying to talk very slowly to him

04:37.160 --> 04:39.640
 and I was with another professor, Ernst Frankel.

04:39.640 --> 04:43.120
 And he's like laughing because it's funny

04:43.120 --> 04:45.720
 that I don't get that the guy is talking in this way

04:45.720 --> 04:47.000
 because he thinks that I cannot speak.

04:47.000 --> 04:50.120
 So it was really kind of mirroring experience

04:50.120 --> 04:54.320
 and it led me think a lot about my own experiences

04:54.320 --> 04:57.000
 moving, you know, from different countries.

04:57.000 --> 05:00.280
 So I think that books play a big role

05:00.280 --> 05:02.720
 in my understanding of the world.

05:02.720 --> 05:06.440
 On the science question, you mentioned that

05:06.440 --> 05:08.760
 it made you discover that personalities

05:08.760 --> 05:12.440
 of human beings are more important than perhaps ideas.

05:12.440 --> 05:13.680
 Is that what I heard?

05:13.680 --> 05:15.760
 It's not necessarily that they are more important

05:15.760 --> 05:19.200
 than ideas, but I think that ideas on their own

05:19.200 --> 05:20.520
 are not sufficient.

05:20.520 --> 05:24.720
 And many times at least at the local horizon

05:24.720 --> 05:29.200
 is the personalities and their devotion to their ideas

05:29.200 --> 05:33.000
 is really that locally changes the landscape.

05:33.000 --> 05:37.520
 Now, if you're looking at AI, like let's say 30 years ago,

05:37.520 --> 05:39.240
 you know, dark ages of AI or whatever,

05:39.240 --> 05:42.480
 what the symbolic times you can use any word,

05:42.480 --> 05:44.720
 you know, there were some people,

05:44.720 --> 05:46.680
 now we are looking at a lot of that work

05:46.680 --> 05:48.840
 and we are kind of thinking this was not really

05:48.840 --> 05:50.760
 maybe a relevant work.

05:50.760 --> 05:53.160
 But you can see that some people managed to take it

05:53.160 --> 05:58.000
 and to make it so shiny and dominate the, you know,

05:58.000 --> 06:02.440
 the academic world and make it to be the standard.

06:02.440 --> 06:05.280
 If you look at the area of natural language processing,

06:06.520 --> 06:09.240
 it is well known fact that the reason the statistics

06:09.240 --> 06:14.080
 in NLP took such a long time to become mainstream

06:14.080 --> 06:16.880
 because there were quite a number of personalities

06:16.880 --> 06:18.480
 which didn't believe in this idea

06:18.480 --> 06:22.080
 and then stop research progress in this area.

06:22.080 --> 06:27.080
 So I do not think that, you know, kind of asymptotically

06:27.360 --> 06:28.960
 maybe personalities matters,

06:28.960 --> 06:33.960
 but I think locally it does make quite a bit of impact

06:33.960 --> 06:38.080
 and it's generally, you know, speed up,

06:38.080 --> 06:41.440
 speeds up the rate of adoption of the new ideas.

06:41.440 --> 06:45.480
 Yeah, and the other interesting question is in the early days

06:45.480 --> 06:50.000
 of particular discipline, I think you mentioned in that book

06:50.000 --> 06:52.400
 was, is ultimately a book of cancer?

06:52.400 --> 06:55.160
 It's called The Emperor of All Melodies.

06:55.160 --> 06:58.640
 Yeah, and those melodies included the trying to,

06:58.640 --> 07:00.680
 the medicine, was it centered around?

07:00.680 --> 07:04.960
 So it was actually centered on, you know,

07:04.960 --> 07:07.240
 how people thought of curing cancer.

07:07.240 --> 07:10.720
 Like for me, it was really a discovery how people,

07:10.720 --> 07:14.200
 what was the science of chemistry behind drug development

07:14.200 --> 07:17.280
 that it actually grew up out of dyeing,

07:17.280 --> 07:21.960
 like coloring industry that people who develop chemistry

07:21.960 --> 07:25.840
 in 19th century in Germany and Britain to do,

07:25.840 --> 07:28.160
 you know, the really new dyes,

07:28.160 --> 07:30.200
 they looked at the molecules and identified

07:30.200 --> 07:32.160
 that they do certain things to cells.

07:32.160 --> 07:34.960
 And from there, the process started and, you know,

07:34.960 --> 07:36.920
 like historians think, yeah, this is fascinating

07:36.920 --> 07:38.720
 that they managed to make the connection

07:38.720 --> 07:42.280
 and look under the microscope and do all this discovery.

07:42.280 --> 07:44.440
 But as you continue reading about it

07:44.440 --> 07:48.760
 and you read about how chemotherapy drugs

07:48.760 --> 07:50.520
 were actually developed in Boston

07:50.520 --> 07:52.480
 and some of them were developed

07:52.480 --> 07:57.480
 and Dr. Farber from Dana Farber,

07:57.480 --> 08:00.480
 you know, how the experiments were done,

08:00.480 --> 08:03.360
 that, you know, there was some miscalculation.

08:03.360 --> 08:04.520
 Let's put it this way.

08:04.520 --> 08:05.960
 And they tried it on the patients

08:05.960 --> 08:08.480
 and they just, and those were children with leukemia

08:08.480 --> 08:11.640
 and they died and they tried another modification.

08:11.640 --> 08:15.040
 You look at the process, how imperfect is this process?

08:15.040 --> 08:17.520
 And, you know, like, if we're again looking back

08:17.520 --> 08:19.200
 like 60 years ago, 70 years ago,

08:19.200 --> 08:20.800
 you can kind of understand it.

08:20.800 --> 08:23.040
 But some of the stories in this book,

08:23.040 --> 08:24.640
 which were really shocking to me,

08:24.640 --> 08:28.040
 were really happening, you know, maybe decades ago.

08:28.040 --> 08:30.680
 And we still don't have a vehicle

08:30.680 --> 08:35.120
 to do it much more fast and effective and, you know,

08:35.120 --> 08:36.280
 scientific the way I'm thinking,

08:36.280 --> 08:38.240
 computer science scientific.

08:38.240 --> 08:40.480
 So from the perspective of computer science,

08:40.480 --> 08:43.800
 you've got a chance to work the application to cancer

08:43.800 --> 08:44.920
 and to medicine in general.

08:44.920 --> 08:48.480
 From a perspective of an engineer and a computer scientist,

08:48.480 --> 08:51.840
 how far along are we from understanding the human body,

08:51.840 --> 08:55.600
 biology, of being able to manipulate it in a way

08:55.600 --> 08:59.760
 we can cure some of the maladies, some of the diseases?

08:59.760 --> 09:02.240
 So this is a very interesting question.

09:03.520 --> 09:06.080
 And if you're thinking as a computer scientist

09:06.080 --> 09:09.880
 about this problem, I think one of the reasons

09:09.880 --> 09:11.920
 that we succeeded in the areas

09:11.920 --> 09:14.040
 we as a computer scientist succeeded

09:14.040 --> 09:16.320
 is because we don't have,

09:16.320 --> 09:19.040
 we are not trying to understand in some ways.

09:19.040 --> 09:22.280
 Like if you're thinking about like eCommerce, Amazon,

09:22.280 --> 09:24.280
 Amazon doesn't really understand you.

09:24.280 --> 09:27.760
 And that's why it recommends you certain books

09:27.760 --> 09:29.600
 or certain products, correct?

09:30.720 --> 09:34.280
 And in, you know, traditionally,

09:34.280 --> 09:36.440
 when people were thinking about marketing, you know,

09:36.440 --> 09:37.840
 they divided the population

09:37.840 --> 09:39.840
 to different kind of subgroups,

09:39.840 --> 09:41.800
 identify the features of the subgroup

09:41.800 --> 09:43.160
 and come up with a strategy

09:43.160 --> 09:45.640
 which is specific to that subgroup.

09:45.640 --> 09:47.120
 If you're looking about recommendations,

09:47.120 --> 09:50.640
 they're not claiming that they're understanding somebody,

09:50.640 --> 09:54.800
 they're just managing from the patterns of your behavior

09:54.800 --> 09:57.600
 to recommend you a product.

09:57.600 --> 09:59.600
 Now, if you look at the traditional biology,

09:59.600 --> 10:04.600
 obviously I wouldn't say that I am at any way,

10:04.920 --> 10:06.200
 you know, educated in this field.

10:06.200 --> 10:07.120
 But, you know, what I see,

10:07.120 --> 10:09.320
 there is really a lot of emphasis

10:09.320 --> 10:10.720
 on mechanistic understanding.

10:10.720 --> 10:13.840
 And it was very surprising to me coming from computer science

10:13.840 --> 10:17.640
 how much emphasis is on this understanding.

10:17.640 --> 10:20.760
 And given the complexity of the system,

10:20.760 --> 10:23.240
 maybe the deterministic full understanding

10:23.240 --> 10:27.400
 of this process is, you know, beyond our capacity.

10:27.400 --> 10:29.480
 And the same way as in computer science,

10:29.480 --> 10:30.520
 when we're doing recognition,

10:30.520 --> 10:32.760
 when we're doing recommendation in many other areas,

10:32.760 --> 10:35.960
 it's just probabilistic matching process.

10:35.960 --> 10:40.080
 And in some way, maybe in certain cases,

10:40.080 --> 10:42.960
 we shouldn't even attempt to understand

10:42.960 --> 10:45.760
 or we can attempt to understand, but in parallel,

10:45.760 --> 10:48.040
 we can actually do this kind of matching

10:48.040 --> 10:52.600
 that would help us to find Qo to do early diagnostics

10:52.600 --> 10:54.120
 and so on.

10:54.120 --> 10:55.840
 And I know that in these communities,

10:55.840 --> 10:59.040
 it's really important to understand,

10:59.040 --> 11:00.680
 but I am sometimes wondering,

11:00.680 --> 11:02.920
 what exactly does it mean to understand here?

11:02.920 --> 11:04.440
 Well, there's stuff that works,

11:04.440 --> 11:07.600
 and, but that can be, like you said,

11:07.600 --> 11:10.320
 separate from this deep human desire

11:10.320 --> 11:12.720
 to uncover the mysteries of the universe,

11:12.720 --> 11:16.160
 of science, of the way the body works,

11:16.160 --> 11:17.600
 the way the mind works.

11:17.600 --> 11:19.560
 It's the dream of symbolic AI,

11:19.560 --> 11:24.560
 of being able to reduce human knowledge into logic

11:25.200 --> 11:26.880
 and be able to play with that logic

11:26.880 --> 11:28.680
 in a way that's very explainable

11:28.680 --> 11:30.280
 and understandable for us humans.

11:30.280 --> 11:31.760
 I mean, that's a beautiful dream.

11:31.760 --> 11:34.840
 So I understand it, but it seems that

11:34.840 --> 11:37.880
 what seems to work today, and we'll talk about it more,

11:37.880 --> 11:40.760
 is as much as possible, reduce stuff into data,

11:40.760 --> 11:43.880
 reduce whatever problem you're interested in to data

11:43.880 --> 11:47.040
 and try to apply statistical methods,

11:47.040 --> 11:49.080
 apply machine learning to that.

11:49.080 --> 11:51.120
 On a personal note,

11:51.120 --> 11:54.160
 you were diagnosed with breast cancer in 2014.

11:55.400 --> 11:58.400
 Would it facing your mortality make you think about

11:58.400 --> 12:00.200
 how did it change you?

12:00.200 --> 12:01.800
 You know, this is a great question.

12:01.800 --> 12:03.800
 And I think that I was interviewed many times

12:03.800 --> 12:05.680
 and nobody actually asked me this question.

12:05.680 --> 12:09.640
 I think I was 43 at a time.

12:09.640 --> 12:12.800
 And the first time I realized in my life that I may die.

12:12.800 --> 12:14.400
 And I never thought about it before.

12:14.400 --> 12:16.200
 And yeah, and there was a long time

12:16.200 --> 12:17.920
 since you diagnosed until you actually know

12:17.920 --> 12:20.120
 what you have and how severe is your disease.

12:20.120 --> 12:23.480
 For me, it was like maybe two and a half months.

12:23.480 --> 12:28.280
 And I didn't know where I am during this time

12:28.280 --> 12:30.640
 because I was getting different tests.

12:30.640 --> 12:32.200
 And one would say, it's bad.

12:32.200 --> 12:33.360
 And I would say, no, it is not.

12:33.360 --> 12:34.840
 So until I knew where I am,

12:34.840 --> 12:36.280
 I really was thinking about

12:36.280 --> 12:38.200
 all these different possible outcomes.

12:38.200 --> 12:39.680
 Were you imagining the worst?

12:39.680 --> 12:41.920
 Or were you trying to be optimistic?

12:41.920 --> 12:45.600
 It would be really, I don't remember, you know,

12:45.600 --> 12:47.360
 what was my thinking?

12:47.360 --> 12:50.880
 It was really a mixture with many components at the time

12:50.880 --> 12:54.080
 at speaking, you know, in our terms.

12:54.080 --> 12:59.320
 And one thing that I remember,

12:59.320 --> 13:01.480
 and you know, every test comes and then you think,

13:01.480 --> 13:03.280
 oh, it could be this or it may not be this.

13:03.280 --> 13:04.680
 And you're hopeful and then you're desperate.

13:04.680 --> 13:06.600
 So it's like, there is a whole, you know,

13:06.600 --> 13:09.800
 slew of emotions that goes through you.

13:09.800 --> 13:15.120
 But what I remember is that when I came back to MIT,

13:15.120 --> 13:17.160
 I was kind of going the whole time

13:17.160 --> 13:18.280
 through the treatment to MIT,

13:18.280 --> 13:19.760
 but my brain was not really there.

13:19.760 --> 13:21.800
 But when I came back really, finished my treatment

13:21.800 --> 13:24.560
 and I was here teaching and everything.

13:24.560 --> 13:27.080
 You know, I look back at what my group was doing,

13:27.080 --> 13:28.840
 what other groups was doing,

13:28.840 --> 13:30.840
 and I saw these three realities.

13:30.840 --> 13:33.240
 It's like people are building their careers

13:33.240 --> 13:35.520
 on improving some parts around two or three percent

13:35.520 --> 13:36.880
 or whatever.

13:36.880 --> 13:38.400
 I was, it's like, seriously,

13:38.400 --> 13:40.760
 I did a work on how to decipher eukaryotic,

13:40.760 --> 13:42.880
 like a language that nobody speak and whatever,

13:42.880 --> 13:46.160
 like what is significance?

13:46.160 --> 13:49.000
 When I was sad, you know, I walked out of MIT,

13:49.000 --> 13:51.600
 which is, you know, when people really do care,

13:51.600 --> 13:54.280
 you know, what happened to your iClear paper,

13:54.280 --> 13:56.680
 you know, what is your next publication,

13:56.680 --> 13:59.960
 to ACL, to the world where people, you know,

13:59.960 --> 14:01.880
 people, you see a lot of suffering

14:01.880 --> 14:04.920
 that I'm kind of totally shielded on it on a daily basis.

14:04.920 --> 14:07.520
 And it's like the first time I've seen like real life

14:07.520 --> 14:08.720
 and real suffering.

14:09.760 --> 14:13.280
 And I was thinking, why are we trying to improve the parser

14:13.280 --> 14:16.120
 or deal with some trivialities

14:16.120 --> 14:20.760
 when we have capacity to really make a change?

14:20.760 --> 14:23.520
 And it was really challenging to me

14:23.520 --> 14:24.640
 because on one hand, you know,

14:24.640 --> 14:26.720
 I have my graduate students who really want to do

14:26.720 --> 14:28.760
 their papers and their work

14:28.760 --> 14:30.880
 and they want to continue to do what they were doing,

14:30.880 --> 14:31.960
 which was great.

14:31.960 --> 14:36.360
 And then it was me who really kind of reevaluated

14:36.360 --> 14:38.600
 what is importance and also at that point

14:38.600 --> 14:40.280
 because I had to take some break.

14:42.560 --> 14:47.560
 I look back into like my years in science

14:47.760 --> 14:49.080
 and I was thinking, you know,

14:49.080 --> 14:51.720
 like 10 years ago, this was the biggest thing.

14:51.720 --> 14:53.000
 I don't know, topic models.

14:53.000 --> 14:55.400
 We have like millions of papers on topic models

14:55.400 --> 14:56.720
 and variation of topics models now.

14:56.720 --> 14:58.640
 It's totally like irrelevant.

14:58.640 --> 15:01.520
 And you start looking at this, you know,

15:01.520 --> 15:03.280
 what do you perceive as important

15:03.280 --> 15:04.560
 at different point of time

15:04.560 --> 15:08.960
 and how, you know, it fades over time.

15:08.960 --> 15:13.040
 And since we have a limited time,

15:13.040 --> 15:14.960
 all of us have limited time on us.

15:14.960 --> 15:18.440
 It's really important to prioritize things

15:18.440 --> 15:19.760
 that really matter to you,

15:19.760 --> 15:22.040
 maybe matter to you at that particular point,

15:22.040 --> 15:24.400
 but it's important to take some time

15:24.400 --> 15:26.960
 and understand what matters to you,

15:26.960 --> 15:28.880
 which may not necessarily be the same

15:28.880 --> 15:31.720
 as what matters to the rest of your scientific community

15:31.720 --> 15:34.600
 and pursue that vision.

15:34.600 --> 15:38.480
 And so though that moment, did it make you cognizant?

15:38.480 --> 15:42.520
 You mentioned suffering of just the general amount

15:42.520 --> 15:44.360
 of suffering in the world.

15:44.360 --> 15:45.640
 Is that what you're referring to?

15:45.640 --> 15:49.480
 So as opposed to topic models and specific detail problems

15:49.480 --> 15:54.480
 in NLP, did you start to think about other people

15:54.480 --> 15:57.040
 who have been diagnosed with cancer?

15:57.040 --> 15:58.360
 Is that the way you saw the,

15:58.360 --> 16:00.040
 started to see the world perhaps?

16:00.040 --> 16:00.880
 Oh, absolutely.

16:00.880 --> 16:04.480
 And it actually creates because like, for instance,

16:04.480 --> 16:06.080
 you know, there's parts of the treatment

16:06.080 --> 16:08.520
 where you need to go to the hospital every day

16:08.520 --> 16:11.640
 and you see, you know, the community of people that you see

16:11.640 --> 16:16.080
 and many of them are much worse than I was at a time

16:16.080 --> 16:20.480
 and you're all of a sudden see it all.

16:20.480 --> 16:23.920
 And people who are happier someday

16:23.920 --> 16:25.320
 just because they feel better.

16:25.320 --> 16:28.480
 And for people who are in our normal realm,

16:28.480 --> 16:30.800
 you take it totally for granted that you feel well,

16:30.800 --> 16:32.920
 that if you decide to go running, you can go running

16:32.920 --> 16:36.120
 and you can, you know, you're pretty much free to do

16:36.120 --> 16:37.600
 whatever you want with your body.

16:37.600 --> 16:40.200
 Like I saw like a community,

16:40.200 --> 16:42.840
 my community became those people.

16:42.840 --> 16:46.760
 And I remember one of my friends,

16:46.760 --> 16:48.920
 Dina Katabi took me to Prudential

16:48.920 --> 16:50.480
 to buy me a gift for my birthday.

16:50.480 --> 16:52.360
 And it was like the first time in months

16:52.360 --> 16:55.000
 that I went to kind of to see other people.

16:55.000 --> 16:56.640
 And I was like, wow.

16:56.640 --> 16:58.200
 First of all, these people, you know,

16:58.200 --> 16:59.880
 they're happy and they're laughing

16:59.880 --> 17:02.680
 and they're very different from this other my people.

17:02.680 --> 17:04.680
 And second of thing, are they totally crazy?

17:04.680 --> 17:06.400
 They're like laughing and wasting their money

17:06.400 --> 17:08.480
 on some stupid gifts.

17:08.480 --> 17:12.560
 And, you know, they may die.

17:12.560 --> 17:14.280
 They already may have cancer

17:14.280 --> 17:16.000
 and they don't understand it.

17:16.000 --> 17:20.120
 So you can really see how the mind changes

17:20.120 --> 17:22.400
 that you can see that, you know,

17:22.400 --> 17:23.240
 before that you can ask,

17:23.240 --> 17:24.400
 didn't you know that you're gonna die?

17:24.400 --> 17:28.360
 Of course I knew, but it was kind of a theoretical notion.

17:28.360 --> 17:31.080
 It wasn't something which was concrete.

17:31.080 --> 17:33.920
 And at that point when you really see it

17:33.920 --> 17:37.680
 and see how little means sometime the system

17:37.680 --> 17:40.480
 has to help them, you really feel

17:40.480 --> 17:43.960
 that we need to take a lot of our brilliance

17:43.960 --> 17:45.480
 that we have here at MIT

17:45.480 --> 17:48.040
 and translate it into something useful.

17:48.040 --> 17:48.880
 Yeah.

17:48.880 --> 17:50.560
 And you still couldn't have a lot of definitions,

17:50.560 --> 17:53.640
 but of course alleviating, suffering, alleviating,

17:53.640 --> 17:57.480
 trying to cure cancer is a beautiful mission.

17:57.480 --> 18:01.320
 So I, of course, know the theoretically the notion

18:01.320 --> 18:04.200
 of cancer, but just reading more and more

18:04.200 --> 18:08.040
 about it's the 1.7 million new cancer cases

18:08.040 --> 18:09.880
 in the United States every year,

18:09.880 --> 18:13.520
 600,000 cancer related deaths every year.

18:13.520 --> 18:17.600
 So this has a huge impact, United States globally.

18:19.360 --> 18:24.360
 When broadly, before we talk about how machine learning,

18:24.400 --> 18:28.680
 how MIT can help, when do you think

18:28.680 --> 18:32.160
 we as a civilization will cure cancer?

18:32.160 --> 18:34.640
 How hard of a problem is it from everything

18:34.640 --> 18:36.240
 you've learned from it recently?

18:37.280 --> 18:39.320
 I cannot really assess it.

18:39.320 --> 18:42.120
 What I do believe will happen with the advancement

18:42.120 --> 18:45.960
 in machine learning that a lot of types of cancer

18:45.960 --> 18:48.480
 we will be able to predict way early

18:48.480 --> 18:53.400
 and more effectively utilize existing treatments.

18:53.400 --> 18:57.520
 I think, I hope at least that with all the advancements

18:57.520 --> 19:01.200
 in AI and drug discovery, we would be able

19:01.200 --> 19:04.680
 to much faster find relevant molecules.

19:04.680 --> 19:08.240
 What I'm not sure about is how long it will take

19:08.240 --> 19:11.920
 the medical establishment and regulatory bodies

19:11.920 --> 19:14.800
 to kind of catch up and to implement it.

19:14.800 --> 19:17.400
 And I think this is a very big piece of puzzle

19:17.400 --> 19:20.440
 that is currently not addressed.

19:20.440 --> 19:21.800
 That's the really interesting question.

19:21.800 --> 19:25.480
 So first, a small detail that I think the answer is yes,

19:25.480 --> 19:30.480
 but is cancer one of the diseases

19:30.480 --> 19:33.400
 that when detected earlier,

19:33.400 --> 19:36.800
 that's a significantly improves the outcomes.

19:38.320 --> 19:40.720
 Cause we will talk about, there's the cure

19:40.720 --> 19:42.720
 and then there is detection.

19:42.720 --> 19:44.880
 And I think one machine learning can really help

19:44.880 --> 19:46.360
 is earlier detection.

19:46.360 --> 19:48.280
 So does detection help?

19:48.280 --> 19:49.400
 Detection is crucial.

19:49.400 --> 19:53.640
 For instance, the vast majority of pancreatic cancer patients

19:53.640 --> 19:57.040
 are detected at the stage that they are incurable.

19:57.040 --> 20:02.040
 That's why they have such a terrible survival rate.

20:03.680 --> 20:07.200
 It's like just a few percent over five years.

20:07.200 --> 20:09.640
 It's pretty much today a death sentence.

20:09.640 --> 20:13.400
 But if you can discover this disease early,

20:14.320 --> 20:16.600
 there are mechanisms to treat it.

20:16.600 --> 20:20.560
 And in fact, I know a number of people who were diagnosed

20:20.560 --> 20:23.440
 and saved just because they had food poisoning.

20:23.440 --> 20:24.840
 They had terrible food poisoning.

20:24.840 --> 20:28.480
 They went to ER, they got scan.

20:28.480 --> 20:30.560
 There were early signs on the scan

20:30.560 --> 20:33.440
 and that would save their lives.

20:33.440 --> 20:35.720
 But this wasn't really an accidental case.

20:35.720 --> 20:37.520
 So as we become better,

20:38.520 --> 20:42.720
 we would be able to help too many more people

20:42.720 --> 20:46.400
 that are likely to develop diseases.

20:46.400 --> 20:50.840
 And I just want to say that as I got more into this field,

20:50.840 --> 20:53.440
 I realized that cancer is of course a terrible disease

20:53.440 --> 20:55.600
 when there are really the whole slew

20:55.600 --> 20:58.240
 of terrible diseases out there,

20:58.240 --> 21:01.560
 like neurodegenerative diseases and others.

21:01.560 --> 21:04.480
 So we, of course, a lot of us are fixated on cancer

21:04.480 --> 21:06.440
 just because it's so prevalent in our society.

21:06.440 --> 21:08.560
 And you see these people when there are a lot of patients

21:08.560 --> 21:10.320
 with neurodegenerative diseases

21:10.320 --> 21:12.520
 and the kind of aging diseases

21:12.520 --> 21:17.120
 that we still don't have a good solution for.

21:17.120 --> 21:22.120
 And I felt as a computer scientist,

21:22.120 --> 21:24.920
 we kind of decided that it's other people's job

21:24.920 --> 21:26.360
 to treat these diseases

21:27.360 --> 21:29.760
 because it's like traditionally people in biology

21:29.760 --> 21:33.720
 or in chemistry or MDs are the ones who are thinking about it.

21:34.720 --> 21:36.720
 And after kind of start paying attention,

21:36.720 --> 21:39.720
 I think that it's really a wrong assumption

21:39.720 --> 21:42.280
 and we all need to join the battle.

21:42.280 --> 21:45.880
 So how it seems like in cancer specifically

21:45.880 --> 21:48.480
 that there's a lot of ways that machine learning can help.

21:48.480 --> 21:51.240
 So what's the role of machine learning

21:51.240 --> 21:54.160
 and machine learning in the diagnosis of cancer?

21:55.320 --> 21:57.280
 So for many cancers today,

21:57.280 --> 22:02.280
 we really don't know what is your likelihood to get cancer.

22:03.520 --> 22:06.360
 And for the vast majority of patients,

22:06.360 --> 22:08.000
 especially on the younger patients,

22:08.000 --> 22:09.640
 it really comes as a surprise.

22:09.640 --> 22:11.200
 Like for instance, for breast cancer,

22:11.200 --> 22:13.920
 80% of the patients are first in their families,

22:13.920 --> 22:15.440
 it's like me.

22:15.440 --> 22:18.520
 And I never saw that I had any increased risk

22:18.520 --> 22:20.880
 because nobody had it in my family

22:20.880 --> 22:22.360
 and for some reason in my head,

22:22.360 --> 22:24.880
 it was kind of inherited disease.

22:26.640 --> 22:28.440
 But even if I would pay attention,

22:28.440 --> 22:30.280
 the models that currently,

22:30.280 --> 22:32.440
 there's very simplistic statistical models

22:32.440 --> 22:34.600
 that are currently used in clinical practice

22:34.600 --> 22:37.520
 that really don't give you an answer, so you don't know.

22:37.520 --> 22:40.440
 And the same true for pancreatic cancer,

22:40.440 --> 22:45.440
 the same true for non smoking lung cancer and many others.

22:45.440 --> 22:47.400
 So what machine learning can do here

22:47.400 --> 22:51.680
 is utilize all this data to tell us Ellie,

22:51.680 --> 22:53.200
 who is likely to be susceptible

22:53.200 --> 22:56.040
 and using all the information that is already there,

22:56.040 --> 23:00.040
 be it imaging, be it your other tests

23:00.040 --> 23:04.880
 and eventually liquid biopsies and others,

23:04.880 --> 23:08.240
 where the signal itself is not sufficiently strong

23:08.240 --> 23:11.360
 for human eye to do good discrimination

23:11.360 --> 23:12.960
 because the signal may be weak,

23:12.960 --> 23:15.680
 but by combining many sources,

23:15.680 --> 23:18.120
 machine which is trained on large volumes of data

23:18.120 --> 23:20.680
 can really detect it Ellie

23:20.680 --> 23:22.480
 and that's what we've seen with breast cancer

23:22.480 --> 23:25.920
 and people are reporting it in other diseases as well.

23:25.920 --> 23:28.240
 That really boils down to data, right?

23:28.240 --> 23:30.960
 And in the different kinds of sources of data.

23:30.960 --> 23:33.720
 And you mentioned regulatory challenges.

23:33.720 --> 23:35.160
 So what are the challenges

23:35.160 --> 23:39.240
 in gathering large data sets in the space?

23:40.840 --> 23:42.640
 Again, another great question.

23:42.640 --> 23:44.320
 So it took me after I decided

23:44.320 --> 23:48.720
 that I want to work on it two years to get access to data.

23:48.720 --> 23:51.360
 And you did, like any significant amount.

23:51.360 --> 23:53.560
 Like right now in this country,

23:53.560 --> 23:58.040
 there is no publicly available data set of modern mammograms

23:58.040 --> 23:59.560
 that you can just go on your computer,

23:59.560 --> 24:01.840
 sign a document and get it.

24:01.840 --> 24:03.320
 It just doesn't exist.

24:03.320 --> 24:06.880
 I mean, obviously every hospital has its own collection

24:06.880 --> 24:10.160
 of mammograms, there are data that came out

24:10.160 --> 24:11.320
 of clinical trials.

24:11.320 --> 24:13.200
 What we're talking about here is a computer scientist

24:13.200 --> 24:17.120
 who just want to run his or her model

24:17.120 --> 24:19.040
 and see how it works.

24:19.040 --> 24:22.880
 This data, like ImageNet, doesn't exist.

24:22.880 --> 24:27.880
 And there is an set which is called like Florid data set

24:28.640 --> 24:30.880
 which is a film mammogram from 90s

24:30.880 --> 24:32.440
 which is totally not representative

24:32.440 --> 24:33.880
 of the current developments,

24:33.880 --> 24:35.800
 whatever you're learning on them doesn't scale up.

24:35.800 --> 24:39.320
 This is the only resource that is available.

24:39.320 --> 24:44.320
 And today there are many agencies that govern access to data,

24:44.440 --> 24:46.280
 like the hospital holds your data

24:46.280 --> 24:49.240
 and the hospital decides whether they would give it

24:49.240 --> 24:52.320
 to the researcher to work with this data or not.

24:52.320 --> 24:54.160
 In the individual hospital?

24:54.160 --> 24:57.160
 Yeah, I mean, the hospital may, you know,

24:57.160 --> 24:59.200
 assuming that you're doing research collaboration,

24:59.200 --> 25:01.960
 you can submit, you know,

25:01.960 --> 25:05.040
 there is a proper approval process guided by IRB

25:05.040 --> 25:07.800
 and if you go through all the processes,

25:07.800 --> 25:10.120
 you can eventually get access to the data.

25:10.120 --> 25:13.520
 But if you yourself know our AI community,

25:13.520 --> 25:14.680
 there are not that many people

25:14.680 --> 25:16.560
 who actually ever got access to data

25:16.560 --> 25:20.200
 because it's very challenging process.

25:20.200 --> 25:22.720
 And sorry, just in a quick comment,

25:22.720 --> 25:25.760
 MGH or any kind of hospital,

25:25.760 --> 25:28.280
 are they scanning the data?

25:28.280 --> 25:29.680
 Are they digitally storing it?

25:29.680 --> 25:31.560
 Oh, it is already digitally stored.

25:31.560 --> 25:34.120
 You don't need to do any extra processing steps.

25:34.120 --> 25:36.320
 It's already there in the right format.

25:36.320 --> 25:39.800
 Is that right now there are a lot of issues

25:39.800 --> 25:41.200
 that govern access to the data

25:41.200 --> 25:46.200
 because the hospital is legally responsible for the data.

25:46.280 --> 25:51.120
 And, you know, they have a lot to lose

25:51.120 --> 25:53.200
 if they give the data to the wrong person,

25:53.200 --> 25:55.360
 but they may not have a lot to gain

25:55.360 --> 25:58.680
 if they give it as a hospital, as a legal entity

25:59.920 --> 26:00.760
 as given it to you.

26:00.760 --> 26:02.800
 And the way, you know, what I would mention

26:02.800 --> 26:04.840
 happening in the future is the same thing

26:04.840 --> 26:06.840
 that happens when you're getting your driving license.

26:06.840 --> 26:09.880
 You can decide whether you want to donate your organs.

26:09.880 --> 26:11.600
 So you can imagine that whenever a person

26:11.600 --> 26:13.080
 goes to the hospital,

26:14.280 --> 26:18.080
 it should be easy for them to donate their data for research

26:18.080 --> 26:19.480
 and it can be different kind of,

26:19.480 --> 26:21.320
 do they only give you a test results

26:21.320 --> 26:25.880
 or only imaging data or the whole medical record?

26:27.080 --> 26:29.000
 Because at the end,

26:30.560 --> 26:33.880
 we all will benefit from all this insights.

26:33.880 --> 26:36.080
 And it's only gonna say, I want to keep my data private,

26:36.080 --> 26:38.800
 but I would really love to get it from other people

26:38.800 --> 26:40.760
 because other people are thinking the same way.

26:40.760 --> 26:45.760
 So if there is a mechanism to do this donation

26:45.760 --> 26:48.040
 and the patient has an ability to say

26:48.040 --> 26:50.840
 how they want to use their data for research,

26:50.840 --> 26:54.120
 it would be really a game changer.

26:54.120 --> 26:56.480
 People, when they think about this problem,

26:56.480 --> 26:58.480
 there's a depends on the population,

26:58.480 --> 27:00.160
 depends on the demographics,

27:00.160 --> 27:02.400
 but there's some privacy concerns.

27:02.400 --> 27:04.440
 Generally, not just medical data,

27:04.440 --> 27:05.880
 just any kind of data.

27:05.880 --> 27:07.840
 It's what you said, my data,

27:07.840 --> 27:09.600
 it should belong kinda to me,

27:09.600 --> 27:11.520
 I'm worried how it's gonna be misused.

27:12.520 --> 27:15.640
 How do we alleviate those concerns?

27:17.080 --> 27:19.440
 Because that seems like a problem that needs to be,

27:19.440 --> 27:23.000
 that problem of trust, of transparency needs to be solved

27:23.000 --> 27:27.240
 before we build large data sets that help detect cancer,

27:27.240 --> 27:30.160
 help save those very people in the future.

27:30.160 --> 27:31.920
 So seeing that two things that could be done,

27:31.920 --> 27:34.480
 there is a technical solutions

27:34.480 --> 27:38.240
 and there are societal solutions.

27:38.240 --> 27:40.200
 So on the technical end,

27:41.440 --> 27:46.440
 we today have ability to improve disambiguation,

27:48.120 --> 27:49.720
 like for instance, for imaging,

27:49.720 --> 27:54.720
 it's, you know, for imaging, you can do it pretty well.

27:55.600 --> 27:56.760
 What's disambiguation?

27:56.760 --> 27:58.520
 And disambiguation, sorry, disambiguation,

27:58.520 --> 27:59.840
 removing the identification,

27:59.840 --> 28:02.200
 removing the names of the people.

28:02.200 --> 28:04.840
 There are other data, like if it is a raw text,

28:04.840 --> 28:08.200
 you cannot really achieve 99.9%

28:08.200 --> 28:10.080
 but there are all these techniques

28:10.080 --> 28:12.480
 that actually some of them are developed at MIT,

28:12.480 --> 28:15.440
 how you can do learning on the encoded data

28:15.440 --> 28:17.400
 where you locally encode the image,

28:17.400 --> 28:19.040
 you train on network,

28:19.040 --> 28:22.440
 which only works on the encoded images

28:22.440 --> 28:24.960
 and then you send the outcome back to the hospital

28:24.960 --> 28:26.600
 and you can open it up.

28:26.600 --> 28:28.040
 So those are the technical solutions.

28:28.040 --> 28:30.720
 There are a lot of people who are working in this space

28:30.720 --> 28:33.320
 where the learning happens in the encoded form.

28:33.320 --> 28:36.160
 We are still early,

28:36.160 --> 28:39.240
 but this is an interesting research area

28:39.240 --> 28:41.880
 where I think we'll make more progress.

28:43.360 --> 28:45.640
 There is a lot of work in natural language processing

28:45.640 --> 28:48.600
 community, how to do the identification better.

28:50.400 --> 28:54.040
 But even today, there are already a lot of data

28:54.040 --> 28:55.920
 which can be identified perfectly,

28:55.920 --> 28:58.760
 like your test data, for instance, correct,

28:58.760 --> 29:00.040
 where you can just, you know,

29:00.040 --> 29:01.000
 the name of the patient,

29:01.000 --> 29:04.320
 you just want to extract the part with the numbers.

29:04.320 --> 29:07.480
 The big problem here is again,

29:08.440 --> 29:10.440
 hospitals don't see much incentive

29:10.440 --> 29:12.640
 to give this data away on one hand

29:12.640 --> 29:14.200
 and then there is general concern.

29:14.200 --> 29:17.720
 Now, when I'm talking about societal benefits

29:17.720 --> 29:19.640
 and about the education,

29:19.640 --> 29:23.640
 the public needs to understand

29:23.640 --> 29:27.800
 and I think that there are situations

29:27.800 --> 29:29.360
 that I still remember myself

29:29.360 --> 29:31.520
 when I really needed an answer.

29:31.520 --> 29:33.280
 I had to make a choice

29:33.280 --> 29:35.200
 and there was no information to make a choice.

29:35.200 --> 29:36.640
 You're just guessing.

29:36.640 --> 29:38.760
 And at that moment,

29:38.760 --> 29:41.040
 you feel that your life is at the stake,

29:41.040 --> 29:44.760
 but you just don't have information to make the choice.

29:44.760 --> 29:48.680
 And many times when I give talks,

29:48.680 --> 29:51.240
 I get emails from women who say,

29:51.240 --> 29:52.760
 you know, I'm in this situation,

29:52.760 --> 29:54.160
 can you please run statistic

29:54.160 --> 29:55.920
 and see what are the outcomes?

29:57.040 --> 30:00.000
 We get almost every week a mammogram

30:00.000 --> 30:02.520
 that comes by mail to my office at MIT.

30:02.520 --> 30:06.200
 I'm serious that people ask to run

30:06.200 --> 30:07.840
 because they need to make, you know,

30:07.840 --> 30:10.000
 life changing decisions.

30:10.000 --> 30:11.320
 And of course, you know,

30:11.320 --> 30:12.920
 I'm not planning to open a clinic here,

30:12.920 --> 30:16.600
 but we do run and give them the results for their doctors.

30:16.600 --> 30:20.040
 But the point that I'm trying to make

30:20.040 --> 30:23.760
 that we all at some point or our loved ones

30:23.760 --> 30:26.600
 will be in the situation where you need information

30:26.600 --> 30:28.840
 to make the best choice.

30:28.840 --> 30:31.840
 And if this information is not available,

30:31.840 --> 30:35.080
 you would feel vulnerable and unprotected.

30:35.080 --> 30:36.880
 And then the question is, you know,

30:36.880 --> 30:37.840
 what do I care more?

30:37.840 --> 30:40.320
 Because at the end everything is a trade off, correct?

30:40.320 --> 30:41.640
 Yeah, exactly.

30:41.640 --> 30:43.080
 Just out of curiosity,

30:43.080 --> 30:45.560
 what it seems like one possible solution,

30:45.560 --> 30:47.160
 I'd like to see what you think of it

30:47.160 --> 30:50.680
 based on what you just said,

30:50.680 --> 30:52.480
 based on wanting to know answers

30:52.480 --> 30:55.040
 for when you're yourself in that situation.

30:55.040 --> 30:58.400
 Is it possible for patients to own their data

30:58.400 --> 31:01.040
 as opposed to hospitals owning their data?

31:01.040 --> 31:02.280
 Of course, theoretically,

31:02.280 --> 31:04.120
 I guess patients own their data,

31:04.120 --> 31:06.640
 but can you walk out there with a USB stick

31:07.600 --> 31:10.600
 containing everything or upload it to the cloud

31:10.600 --> 31:13.400
 where a company, you know,

31:13.400 --> 31:15.680
 I remember Microsoft had a service,

31:15.680 --> 31:17.760
 like I try, I was really excited about

31:17.760 --> 31:19.240
 and Google Health was there.

31:19.240 --> 31:21.880
 I tried to give, I was excited about it.

31:21.880 --> 31:24.760
 Basically companies helping you upload your data

31:24.760 --> 31:27.920
 to the cloud so that you can move from hospital to hospital

31:27.920 --> 31:29.200
 from doctor to doctor.

31:29.200 --> 31:32.680
 Do you see a promise of that kind of possibility?

31:32.680 --> 31:34.640
 I absolutely think this is, you know,

31:34.640 --> 31:38.160
 the right way to exchange the data.

31:38.160 --> 31:41.680
 I don't know now who's the biggest player in this field,

31:41.680 --> 31:46.280
 but I can clearly see that even for totally selfish health

31:46.280 --> 31:49.280
 reasons, when you are going to a new facility

31:49.280 --> 31:52.600
 and many of us are sent to some specialized treatment,

31:52.600 --> 31:55.720
 they don't easily have access to your data.

31:55.720 --> 31:58.960
 And today, you know, we would want to send us

31:58.960 --> 32:00.680
 Mammogram need to go to their hospital,

32:00.680 --> 32:01.760
 find some small office,

32:01.760 --> 32:04.760
 which gives them the CD and they ship as a CD.

32:04.760 --> 32:08.280
 So you can imagine we're looking at kind of decades old

32:08.280 --> 32:10.080
 mechanism of data exchange.

32:10.080 --> 32:15.080
 So I definitely think this is an area where hopefully

32:15.600 --> 32:20.360
 all the right regulatory and technical forces will align

32:20.360 --> 32:23.200
 and we will see it actually implemented.

32:23.200 --> 32:25.720
 It's sad because unfortunately,

32:25.720 --> 32:28.400
 and I have, I need to research why that happened,

32:28.400 --> 32:32.080
 but I'm pretty sure Google Health and Microsoft Health Vault

32:32.080 --> 32:34.640
 or whatever it's called, both closed down,

32:34.640 --> 32:37.560
 which means that there was either regulatory pressure

32:37.560 --> 32:39.080
 or there's not a business case

32:39.080 --> 32:41.760
 or there's challenges from hospitals,

32:41.760 --> 32:43.240
 which is very disappointing.

32:43.240 --> 32:46.480
 So when you say, you don't know what the biggest players are,

32:46.480 --> 32:50.520
 the two biggest that I was aware of closed their doors.

32:50.520 --> 32:53.120
 So I'm hoping I'd love to see why

32:53.120 --> 32:54.760
 and I'd love to see who else can come up.

32:54.760 --> 32:59.600
 It seems like one of those Elon Musk style problems

32:59.600 --> 33:01.280
 that are obvious needs to be solved

33:01.280 --> 33:02.360
 and somebody needs to step up

33:02.360 --> 33:07.360
 and actually do this large scale data collection.

33:07.360 --> 33:09.600
 So I know there is an initiative in Massachusetts,

33:09.600 --> 33:11.720
 a thing actually led by the governor

33:11.720 --> 33:15.440
 to try to create this kind of health exchange system

33:15.440 --> 33:17.840
 where at least to help people who are kind of when you show up

33:17.840 --> 33:20.160
 in emergency room and there is no information

33:20.160 --> 33:22.520
 about what are your allergies and other things.

33:23.480 --> 33:26.080
 So I don't know how far it will go,

33:26.080 --> 33:30.280
 but another thing that you said and I find it very interesting

33:30.280 --> 33:33.760
 is actually who are the successful players in this space

33:33.760 --> 33:36.080
 and the whole implementation.

33:36.080 --> 33:37.240
 How does it go?

33:37.240 --> 33:40.280
 To me, it is from the anthropological perspective,

33:40.280 --> 33:44.640
 it's more fascinating that AI that today goes in health care.

33:44.640 --> 33:49.640
 We've seen so many attempts and so very little successes

33:50.360 --> 33:54.200
 and it's interesting to understand that I by no means

33:54.200 --> 33:58.280
 have knowledge to assess why we are in the position

33:58.280 --> 33:59.600
 where we are.

33:59.600 --> 34:02.920
 Yeah, it's interesting because data is really fuel

34:02.920 --> 34:04.960
 for a lot of successful applications

34:04.960 --> 34:08.480
 and when that data requires regulatory approval

34:08.480 --> 34:12.400
 like the FDA or any kind of approval,

34:14.160 --> 34:16.920
 it seems that the computer scientists are not quite there yet

34:16.920 --> 34:18.840
 in being able to play the regulatory game,

34:18.840 --> 34:21.200
 understanding the fundamentals of it.

34:21.200 --> 34:26.200
 I think that in many cases when even people do have data,

34:26.480 --> 34:31.480
 we still don't know what exactly do you need to demonstrate

34:31.480 --> 34:35.040
 to change the standard of care.

34:35.040 --> 34:40.040
 Let me give you an example related to my breast cancer research.

34:41.000 --> 34:45.400
 So in traditional breast cancer risk assessment,

34:45.400 --> 34:47.040
 there is something called density

34:47.040 --> 34:50.400
 which determines the likelihood of a woman to get cancer

34:50.400 --> 34:52.680
 and this is pretty much says how much white

34:52.680 --> 34:54.120
 do you see on the mammogram?

34:54.120 --> 34:58.840
 The whiter it is, the more likely the tissue is dense.

34:58.840 --> 35:02.640
 And the idea behind density,

35:02.640 --> 35:03.560
 it's not a bad idea,

35:03.560 --> 35:07.960
 in 1967 a radiologist called Wolf decided to look back

35:07.960 --> 35:09.680
 at women who were diagnosed

35:09.680 --> 35:12.320
 and see what is special in their images.

35:12.320 --> 35:14.600
 Can we look back and say that they're likely to develop?

35:14.600 --> 35:16.080
 So he come up with some patterns

35:16.080 --> 35:20.520
 and it was the best that his human eye can identify

35:20.520 --> 35:24.160
 then it was kind of formalized and coded into four categories

35:24.160 --> 35:26.840
 and that's what we are using today.

35:26.840 --> 35:31.840
 And today this density assessment is actually a federal law

35:32.240 --> 35:36.120
 from 2019 approved by President Trump

35:36.120 --> 35:38.640
 and for the previous FDA commissioner

35:40.040 --> 35:43.560
 where women are supposed to be advised by their providers

35:43.560 --> 35:45.040
 if they have high density,

35:45.040 --> 35:47.200
 putting them into higher risk category

35:47.200 --> 35:51.240
 and in some states you can actually get supplementary screening

35:51.240 --> 35:53.640
 paid by your insurance because you are in this category.

35:53.640 --> 35:56.720
 Now you can say how much science do we have behind it?

35:56.720 --> 36:00.800
 Whatever biological science or epidemiological evidence.

36:00.800 --> 36:05.080
 So it turns out that between 40 and 50% of women

36:05.080 --> 36:06.600
 have dense breast.

36:06.600 --> 36:11.040
 So above 40% of patients are coming out of their screening

36:11.040 --> 36:14.960
 and somebody tells them you are in high risk.

36:14.960 --> 36:16.800
 Now what exactly does it mean

36:16.800 --> 36:19.520
 if you as half of the population in high risk

36:19.520 --> 36:21.960
 gets from saying maybe I'm not,

36:21.960 --> 36:23.600
 or what do I really need to do with it?

36:23.600 --> 36:28.280
 Because the system doesn't provide me a lot of the solutions

36:28.280 --> 36:30.080
 because there are so many people like me,

36:30.080 --> 36:34.560
 we cannot really provide very expensive solutions for them.

36:34.560 --> 36:38.680
 And the reason this whole density became this big deal

36:38.680 --> 36:40.720
 it's actually advocated by the patients

36:40.720 --> 36:43.560
 who felt very unprotected because many women

36:43.560 --> 36:46.200
 when did the mammograms which were normal

36:46.200 --> 36:49.400
 and then it turns out that they already had cancer,

36:49.400 --> 36:50.520
 quite developed cancer.

36:50.520 --> 36:54.320
 So they didn't have a way to know who is really at risk

36:54.320 --> 36:56.240
 and what is the likelihood that when the doctor tells you

36:56.240 --> 36:58.000
 you're okay, you are not okay.

36:58.000 --> 37:02.080
 So at the time and it was 15 years ago,

37:02.080 --> 37:06.760
 this maybe was the best piece of science that we had

37:06.760 --> 37:12.120
 and it took quite 15, 16 years to make it federal law.

37:12.120 --> 37:15.600
 But now this is a standard.

37:15.600 --> 37:17.560
 Now with a deep learning model

37:17.560 --> 37:19.600
 we can so much more accurately predict

37:19.600 --> 37:21.560
 who is gonna develop breast cancer

37:21.560 --> 37:23.680
 just because you're trained on a logical thing.

37:23.680 --> 37:26.040
 And instead of describing how much white

37:26.040 --> 37:27.360
 and what kind of white machine

37:27.360 --> 37:30.120
 can systematically identify the patterns

37:30.120 --> 37:32.760
 which was the original idea behind the sort

37:32.760 --> 37:33.680
 of the tradiologist,

37:33.680 --> 37:35.680
 machines can do it much more systematically

37:35.680 --> 37:38.240
 and predict the risk when you're training the machine

37:38.240 --> 37:42.080
 to look at the image and to say the risk in one to five years.

37:42.080 --> 37:45.000
 Now you can ask me how long it will take

37:45.000 --> 37:46.400
 to substitute this density

37:46.400 --> 37:48.560
 which is broadly used across the country

37:48.560 --> 37:53.560
 and really it's not helping to bring this new models.

37:54.320 --> 37:56.640
 And I would say it's not a matter of the algorithm.

37:56.640 --> 37:58.720
 Algorithm is already orders of magnitude better

37:58.720 --> 38:00.360
 than what is currently in practice.

38:00.360 --> 38:02.440
 I think it's really the question,

38:02.440 --> 38:04.320
 who do you need to convince?

38:04.320 --> 38:07.400
 How many hospitals do you need to run the experiment?

38:07.400 --> 38:11.560
 What, you know, all this mechanism of adoption

38:11.560 --> 38:15.120
 and how do you explain to patients

38:15.120 --> 38:17.520
 and to women across the country

38:17.520 --> 38:20.400
 that this is really a better measure?

38:20.400 --> 38:22.680
 And again, I don't think it's an AI question.

38:22.680 --> 38:25.880
 We can walk more and make the algorithm even better

38:25.880 --> 38:29.240
 but I don't think that this is the current, you know,

38:29.240 --> 38:32.000
 the barrier, the barrier is really this other piece

38:32.000 --> 38:35.200
 that for some reason is not really explored.

38:35.200 --> 38:36.800
 It's like anthropological piece.

38:36.800 --> 38:39.760
 And coming back to your question about books,

38:39.760 --> 38:42.920
 there is a book that I'm reading.

38:42.920 --> 38:47.920
 It's called American Sickness by Elizabeth Rosenthal

38:48.240 --> 38:51.560
 and I got this book from my clinical collaborator,

38:51.560 --> 38:53.080
 Dr. Kony Lehman.

38:53.080 --> 38:54.800
 And I said, I know everything that I need to know

38:54.800 --> 38:56.000
 about American health system,

38:56.000 --> 38:59.200
 but you know, every page doesn't fail to surprise me.

38:59.200 --> 39:03.080
 And I think that there is a lot of interesting

39:03.080 --> 39:06.840
 and really deep lessons for people like us

39:06.840 --> 39:09.600
 from computer science who are coming into this field

39:09.600 --> 39:13.640
 to really understand how complex is the system of incentives

39:13.640 --> 39:17.160
 in the system to understand how you really need

39:17.160 --> 39:18.760
 to play to drive adoption.

39:19.720 --> 39:21.120
 You just said it's complex,

39:21.120 --> 39:23.960
 but if we're trying to simplify it,

39:23.960 --> 39:27.360
 who do you think most likely would be successful

39:27.360 --> 39:29.480
 if we push on this group of people?

39:29.480 --> 39:30.720
 Is it the doctors?

39:30.720 --> 39:31.760
 Is it the hospitals?

39:31.760 --> 39:34.240
 Is it the governments or policy makers?

39:34.240 --> 39:37.240
 Is it the individual patients, consumers

39:37.240 --> 39:42.240
 who needs to be inspired to most likely lead to adoption?

39:45.200 --> 39:47.120
 Or is there no simple answer?

39:47.120 --> 39:48.280
 There's no simple answer,

39:48.280 --> 39:52.000
 but I think there is a lot of good people in medical system

39:52.000 --> 39:55.240
 who do want to make a change.

39:56.520 --> 40:01.520
 And I think a lot of power will come from us as a consumers

40:01.600 --> 40:04.320
 because we all are consumers or future consumers

40:04.320 --> 40:06.560
 of healthcare services.

40:06.560 --> 40:11.560
 And I think we can do so much more

40:12.080 --> 40:15.560
 in explaining the potential and not in the hype terms

40:15.560 --> 40:17.920
 and not saying that we're now cured or Alzheimer

40:17.920 --> 40:20.560
 and I'm really sick of reading this kind of articles

40:20.560 --> 40:22.120
 which make these claims.

40:22.120 --> 40:24.800
 But really to show with some examples

40:24.800 --> 40:26.520
 what this implementation does

40:26.520 --> 40:29.080
 and how it changes the care.

40:29.080 --> 40:30.040
 Because I can't imagine,

40:30.040 --> 40:33.240
 it doesn't matter what kind of politician it is,

40:33.240 --> 40:35.240
 we all are susceptible to these diseases.

40:35.240 --> 40:37.800
 There is no one who is free.

40:37.800 --> 40:41.080
 And eventually, we all are humans

40:41.080 --> 40:44.880
 and we are looking for a way to alleviate the suffering.

40:44.880 --> 40:47.320
 And this is one possible way

40:47.320 --> 40:49.360
 where we currently are underutilizing,

40:49.360 --> 40:50.960
 which I think can help.

40:51.880 --> 40:55.120
 So it sounds like the biggest problems are outside of AI

40:55.120 --> 40:58.000
 in terms of the biggest impact at this point.

40:58.000 --> 41:00.440
 But are there any open problems

41:00.440 --> 41:03.800
 in the application of ML to oncology in general?

41:03.800 --> 41:05.400
 So improving the detection

41:05.400 --> 41:07.600
 or any other creative methods,

41:07.600 --> 41:09.640
 whether it's on the detection segmentations

41:09.640 --> 41:11.800
 or the vision perception side

41:11.800 --> 41:16.320
 or some other clever of inference.

41:16.320 --> 41:18.560
 Yeah, what in general in your view

41:18.560 --> 41:20.320
 are the open problems in this space?

41:20.320 --> 41:22.480
 So I just want to mention that beside detection,

41:22.480 --> 41:24.880
 another area where I am kind of quite active

41:24.880 --> 41:28.640
 and I think it's really an increasingly important area

41:28.640 --> 41:30.980
 in healthcare is drug design.

41:30.980 --> 41:32.820
 Absolutely.

41:32.820 --> 41:36.940
 Because it's fine if you detect something early,

41:36.940 --> 41:41.140
 but you still need to get drugs

41:41.140 --> 41:43.900
 and new drugs for these conditions.

41:43.900 --> 41:48.300
 And today, all of the drug design, ML is non existent there.

41:48.300 --> 41:53.020
 We don't have any drug that was developed by the ML model

41:53.020 --> 41:54.940
 or even not developed,

41:54.940 --> 41:56.220
 but at least even you,

41:56.220 --> 41:59.300
 that ML model plays some significant role.

41:59.300 --> 42:03.300
 I think this area with all the new ability

42:03.300 --> 42:05.820
 to generate molecules with desired properties

42:05.820 --> 42:10.820
 to do in silica screening is really a big open area.

42:11.300 --> 42:12.660
 It to be totally honest with you,

42:12.660 --> 42:14.940
 when we are doing diagnostics and imaging,

42:14.940 --> 42:17.300
 primarily taking the ideas that were developed

42:17.300 --> 42:20.500
 for other areas and you're applying them with some adaptation.

42:20.500 --> 42:24.700
 The area of drug design

42:24.700 --> 42:27.980
 is really technically interesting and exciting area.

42:27.980 --> 42:30.380
 You need to work a lot with graphs

42:30.380 --> 42:34.620
 and capture various 3D properties.

42:34.620 --> 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.

42:46.780 --> 42:48.820
 We're already getting a lot of successes

42:48.820 --> 42:52.700
 even with the kind of the first generation of this models,

42:52.700 --> 42:56.540
 but there is much more new creative things that you can do.

42:56.540 --> 43:01.540
 And what's very nice to see is actually the more powerful,

43:03.060 --> 43:05.420
 the more interesting models actually do better.

43:05.420 --> 43:10.420
 So there is a place to innovate in machine learning

43:11.300 --> 43:12.500
 in this area.

43:13.900 --> 43:16.820
 And some of these techniques are really unique too,

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.

43:30.660 --> 43:31.980
 How do you discover a drug?

43:31.980 --> 43:33.340
 Is this chemistry?

43:33.340 --> 43:37.500
 Is this trying to predict different chemical reactions?

43:37.500 --> 43:39.620
 Or is it some kind of...

43:39.620 --> 43:42.100
 What do graphs even represent in this space?

43:42.100 --> 43:43.940
 Oh, sorry.

43:43.940 --> 43:45.300
 And what's a drug?

43:45.300 --> 43:47.820
 Okay, so let's say you think that there are many different

43:47.820 --> 43:49.580
 types of drugs, but let's say you're going to talk

43:49.580 --> 43:51.900
 about small molecules because I think today,

43:51.900 --> 43:53.580
 the majority of drugs are small molecules.

43:53.580 --> 43:55.020
 So small molecule is a graph.

43:55.020 --> 44:00.020
 The molecule is just where the node in the graph is an atom

44:00.060 --> 44:01.460
 and then you have the bond.

44:01.460 --> 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, well,

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 --> 44:17.740
 how it is done at scale in the companies,

44:17.740 --> 44:20.220
 you're without machine learning,

44:20.220 --> 44:22.100
 you have high throughput screening.

44:22.100 --> 44:24.540
 So you know that you are interested to get certain

44:24.540 --> 44:26.580
 biological activity of the compounds.

44:26.580 --> 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 --> 44:36.100
 You identify some compounds which have the right activity

44:36.100 --> 44:39.260
 and then at this point, the chemists come

44:39.260 --> 44:44.260
 and they're trying to now to optimize this original heat

44:44.340 --> 44:46.340
 to different properties that you want it to be,

44:46.340 --> 44:49.100
 maybe soluble, you want to decrease toxicity,

44:49.100 --> 44:51.660
 you want to decrease the side effects.

44:51.660 --> 44:54.060
 Are those, sorry, again to the drop,

44:54.060 --> 44:55.540
 can that be done in simulation

44:55.540 --> 44:57.700
 or just by looking at the molecules

44:57.700 --> 44:59.860
 or do you need to actually run reactions

44:59.860 --> 45:02.180
 in real labs with lab posts and stuff?

45:02.180 --> 45:04.060
 So when you do high throughput screening,

45:04.060 --> 45:07.060
 you really do screening, it's in the lab.

45:07.060 --> 45:09.180
 It's really the lab screening,

45:09.180 --> 45:10.980
 you screen the molecules, correct?

45:10.980 --> 45:12.540
 I don't know what screening is.

45:12.540 --> 45:15.100
 The screening, you just check them for certain property.

45:15.100 --> 45:17.340
 Like in the physical space, in the physical world,

45:17.340 --> 45:18.780
 like actually there's a machine probably

45:18.780 --> 45:21.460
 that's actually running the reaction.

45:21.460 --> 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 --> 45:26.860
 and that's why it's called high throughput,

45:26.860 --> 45:29.580
 you know, it becomes cheaper and faster

45:29.580 --> 45:33.820
 to do it on very big number of molecules.

45:33.820 --> 45:38.340
 You run the screening, you identify potential,

45:38.340 --> 45:40.300
 you know, potential good starts

45:40.300 --> 45:42.340
 and then where the chemists come in

45:42.340 --> 45:44.060
 who, you know, have done it many times

45:44.060 --> 45:45.900
 and then they can try to look at it

45:45.900 --> 45:48.260
 and say, how can you change the molecule

45:48.260 --> 45:53.260
 to get the desired profile 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, you know, 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 --> 46:10.180
 they have a lot of domain knowledge of, you know,

46:10.180 --> 46:12.900
 what works, how do you decrease the CCT 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, you know,

46:17.860 --> 46:20.820
 in the FDA approved drugs or even drugs

46:20.820 --> 46:22.140
 that are in clinical trials,

46:22.140 --> 46:27.140
 they are designed using these domain experts

46:27.140 --> 46:30.060
 which goes through this combinatorial space

46:30.060 --> 46:31.980
 of molecules or graphs or whatever

46:31.980 --> 46:35.180
 and find the right one or adjust it to be the right ones.

46:35.180 --> 46:39.260
 Sounds like the breast density heuristic from 67,

46:39.260 --> 46:40.500
 the same echoes.

46:40.500 --> 46:41.820
 It's not necessarily that.

46:41.820 --> 46:45.380
 It's really, you know, 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 change 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's in terms of cost, it's prohibitive.

47:13.940 --> 47:15.820
 If you measure it in terms of times, you know,

47:15.820 --> 47:18.420
 we have lots of diseases for which we don't have any drugs

47:18.420 --> 47:20.100
 and we don't even know how to approach.

47:20.100 --> 47:23.460
 I don't need to mention few drugs

47:23.460 --> 47:26.980
 or degenerative disease drugs that fail, you know.

47:26.980 --> 47:30.900
 So there are lots of, you know, trials that fail,

47:30.900 --> 47:32.180
 you know, in later stages,

47:32.180 --> 47:35.180
 which is really catastrophic from the financial perspective.

47:35.180 --> 47:38.260
 So, you know, is it the effective,

47:38.260 --> 47:39.540
 the most effective mechanism?

47:39.540 --> 47:42.740
 Absolutely no, but this is the only one that currently works.

47:42.740 --> 47:46.660
 And I would, you know, I was closely interacting

47:46.660 --> 47:48.020
 with people in pharmaceutical industry.

47:48.020 --> 47:50.020
 I was really fascinating on how sharp

47:50.020 --> 47:53.860
 and what a deep understanding of the domain do they have.

47:53.860 --> 47:55.500
 It's not observation driven.

47:55.500 --> 47:58.660
 There is really a lot of science behind what they do.

47:58.660 --> 48:00.940
 But if you ask me, can machine learning change it?

48:00.940 --> 48:03.460
 I firmly believe yes,

48:03.460 --> 48:07.020
 because even the most experienced chemists cannot, you know,

48:07.020 --> 48:09.460
 hold in their memory and understanding

48:09.460 --> 48:11.020
 everything that you can learn, you know,

48:11.020 --> 48:14.140
 from millions of molecules and reactions.

48:14.140 --> 48:18.380
 And the space of graphs is a totally new space.

48:18.380 --> 48:20.460
 I mean, it's a really interesting space

48:20.460 --> 48:22.540
 for machine learning to explore, graph generation.

48:22.540 --> 48:24.740
 Yeah, so there are a lot of things that you can do here.

48:24.740 --> 48:27.140
 So we do a lot of work.

48:27.140 --> 48:29.940
 So the first tool that we started with

48:29.940 --> 48:34.940
 was the tool that can predict properties of the molecules.

48:34.940 --> 48:37.820
 So you can just give the molecule and the property.

48:37.820 --> 48:39.900
 It can be bioactivity properties.

48:39.900 --> 48:42.700
 Or it can be some other property.

48:42.700 --> 48:48.580
 And you train the molecules and you can now take a new molecule

48:48.580 --> 48:50.620
 and predict this property.

48:50.620 --> 48:53.420
 Now, when people started working in this area,

48:53.420 --> 48:54.620
 it is something very simple.

48:54.620 --> 48:57.220
 They do kind of existing, you know, fingerprints,

48:57.220 --> 48:59.380
 which is kind of handcrafted features of the molecule

48:59.380 --> 49:01.420
 when you break the graph to substructures

49:01.420 --> 49:04.420
 and then you run, I don't know, a feedforward neural network.

49:04.420 --> 49:07.100
 And what was interesting to see that clearly, you know,

49:07.100 --> 49:09.980
 this was not the most effective way to proceed.

49:09.980 --> 49:12.900
 And you need to have much more complex models

49:12.900 --> 49:15.140
 that can induce a representation

49:15.140 --> 49:18.220
 which can translate this graph into the embeddings

49:18.220 --> 49:20.100
 and do these predictions.

49:20.100 --> 49:22.020
 So this is one direction.

49:22.020 --> 49:24.180
 And another direction, which is kind of related,

49:24.180 --> 49:27.940
 is not only to stop by looking at the embedding itself,

49:27.940 --> 49:31.580
 but actually modify it to produce better molecules.

49:31.580 --> 49:34.780
 So you can think about it as the machine translation

49:34.780 --> 49:37.220
 that you can start with a molecule

49:37.220 --> 49:39.500
 and then there is an improved version of molecule.

49:39.500 --> 49:41.380
 And you can again, with encoder,

49:41.380 --> 49:42.820
 translate it into the hidden space

49:42.820 --> 49:45.020
 and then learn how to modify it to improve

49:45.020 --> 49:48.220
 the in some ways version of the molecules.

49:48.220 --> 49:51.540
 So that's, it's kind of really exciting.

49:51.540 --> 49:54.220
 We already have seen that the property prediction works

49:54.220 --> 49:58.740
 pretty well and now we are generating molecules

49:58.740 --> 50:00.780
 and there is actually labs

50:00.780 --> 50:03.140
 which are manufacturing this molecule.

50:03.140 --> 50:05.180
 So we'll see why it will get to us.

50:05.180 --> 50:06.540
 Okay, that's really exciting.

50:06.540 --> 50:07.980
 There's a lot of problems.

50:07.980 --> 50:10.740
 Speaking of machine translation and embeddings,

50:10.740 --> 50:15.180
 you have done a lot of really great research in NLP,

50:15.180 --> 50:16.740
 natural language processing.

50:18.020 --> 50:20.380
 Can you tell me your journey through NLP,

50:20.380 --> 50:23.860
 what ideas, problems, approaches were you working on

50:23.860 --> 50:26.820
 were you fascinated with, did you explore

50:26.820 --> 50:31.820
 before this magic of deep learning reemerged and after?

50:31.820 --> 50:35.740
 So when I started my work in NLP, it was in 97.

50:35.740 --> 50:37.260
 This was a very interesting time.

50:37.260 --> 50:40.780
 It was exactly the time that I came to ACL

50:40.780 --> 50:43.540
 and the dynamic would barely understand English.

50:43.540 --> 50:46.140
 But it was exactly like the transition point

50:46.140 --> 50:51.140
 because half of the papers were really rule based approaches

50:51.140 --> 50:53.820
 where people took more kind of heavy linguistic approaches

50:53.820 --> 50:57.820
 for small domains and try to build up from there.

50:57.820 --> 50:59.980
 And then there were the first generation of papers

50:59.980 --> 51:01.980
 which were corpus based papers.

51:01.980 --> 51:03.900
 And they were very simple in our terms

51:03.900 --> 51:05.420
 when you collect some statistics

51:05.420 --> 51:07.300
 and do prediction based on them.

51:07.300 --> 51:10.700
 But I found it really fascinating that one community

51:10.700 --> 51:16.700
 can think so very differently about the problem.

51:16.700 --> 51:20.260
 And I remember my first papers that I wrote,

51:20.260 --> 51:21.940
 it didn't have a single formula,

51:21.940 --> 51:25.700
 it didn't have evaluation, it just had examples of outputs.

51:25.700 --> 51:29.500
 And this was a standard of the first generation

51:29.500 --> 51:32.020
 of the field at a time.

51:32.020 --> 51:35.820
 In some ways, I mean, people maybe just started emphasizing

51:35.820 --> 51:37.820
 the empirical evaluation,

51:37.820 --> 51:39.780
 but for many applications like summarization,

51:39.780 --> 51:42.780
 you just wrote some examples of outputs.

51:42.780 --> 51:44.460
 And then increasingly you can see

51:44.460 --> 51:48.300
 that how the statistical approach has dominated the field.

51:48.300 --> 51:52.060
 And we've seen increased performance

51:52.060 --> 51:56.020
 across many basic tasks.

51:56.020 --> 52:00.020
 The sad part of the story may be that if you look

52:00.020 --> 52:01.580
 again through this journey,

52:01.580 --> 52:05.060
 we see that the role of linguistics

52:05.060 --> 52:07.420
 in some ways greatly diminishes.

52:07.420 --> 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 --> 52:18.940
 You mean today?

52:18.940 --> 52:19.780
 Today.

52:19.780 --> 52:20.620
 Today.

52:20.620 --> 52:21.460
 This was definitely...

52:21.460 --> 52:22.300
 Things like syntactic trees,

52:22.300 --> 52:24.380
 just even basically against our conversation

52:24.380 --> 52:27.500
 about human understanding of language,

52:27.500 --> 52:30.260
 which I guess what linguistics would be,

52:30.260 --> 52:34.260
 structured hierarchical representing language

52:34.260 --> 52:35.700
 in a way that's human explainable,

52:35.700 --> 52:39.420
 understandable is missing today.

52:39.420 --> 52:41.100
 I don't know if it is,

52:41.100 --> 52:43.580
 what is explainable and understandable.

52:43.580 --> 52:45.900
 At the end, we perform functions

52:45.900 --> 52:50.100
 and it's okay to have machine which performs a function.

52:50.100 --> 52:53.180
 Like when you're thinking about your calculator, correct?

52:53.180 --> 52:55.420
 Your calculator can do calculation

52:55.420 --> 52:57.580
 very different from you would do the calculation,

52:57.580 --> 52:58.820
 but it's very effective in it.

52:58.820 --> 52:59.700
 And this is fine.

52:59.700 --> 53:04.420
 If we can achieve certain tasks with high accuracy,

53:04.420 --> 53:07.100
 it doesn't necessarily mean that it has to understand

53:07.100 --> 53:09.260
 in the same way as we understand.

53:09.260 --> 53:11.220
 In some ways, it's even naive to request

53:11.220 --> 53:14.900
 because you have so many other sources of information

53:14.900 --> 53:17.860
 that are absent when you are training your system.

53:17.860 --> 53:19.180
 So it's okay.

53:19.180 --> 53:20.020
 Is it delivered?

53:20.020 --> 53:21.460
 And I would tell you one application

53:21.460 --> 53:22.780
 that's just really fascinating.

53:22.780 --> 53:24.300
 In 97, when it came to ACL,

53:24.300 --> 53:25.900
 there were some papers on machine translation.

53:25.900 --> 53:27.460
 They were like primitive,

53:27.460 --> 53:31.100
 like people were trying really, really simple.

53:31.100 --> 53:34.300
 And the feeling, my feeling was that,

53:34.300 --> 53:36.300
 to make real machine translation system,

53:36.300 --> 53:39.580
 it's like to fly at the moon and build a house there

53:39.580 --> 53:41.620
 and the garden and live happily ever after.

53:41.620 --> 53:42.620
 I mean, it's like impossible.

53:42.620 --> 53:46.740
 I never could imagine that within 10 years,

53:46.740 --> 53:48.580
 we would already see the system working.

53:48.580 --> 53:51.620
 And now nobody is even surprised

53:51.620 --> 53:54.460
 to utilize the system on daily basis.

53:54.460 --> 53:56.260
 So this was like a huge, huge progress,

53:56.260 --> 53:57.900
 saying that people for very long time

53:57.900 --> 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 --> 54:06.180
 That's why I'm coming back to a question about biology,

54:06.180 --> 54:10.820
 that in linguistics, people try to go this way

54:10.820 --> 54:13.540
 and try to write the syntactic trees

54:13.540 --> 54:14.860
 and try to obstruct it

54:14.860 --> 54:17.060
 and to find the right representation.

54:17.060 --> 54:22.060
 And, you know, they couldn't get very far

54:22.220 --> 54:25.940
 with this understanding while these models,

54:25.940 --> 54:29.620
 using, you know, other sources actually capable

54:29.620 --> 54:31.660
 to make a lot of progress.

54:31.660 --> 54:33.940
 Now, I'm not naive to think

54:33.940 --> 54:36.860
 that we are in this paradise space in NLP

54:36.860 --> 54:38.540
 and I'm sure as you know,

54:38.540 --> 54:40.860
 that when we slightly change the domain

54:40.860 --> 54:42.580
 and when we decrease the amount of training,

54:42.580 --> 54:44.740
 it can do like really bizarre and funny thing.

54:44.740 --> 54:47.140
 But I think it's just a matter of improving

54:47.140 --> 54:48.540
 generalization capacity,

54:48.540 --> 54:51.500
 which is just a technical question.

54:51.500 --> 54:54.300
 Well, so that's the question.

54:54.300 --> 54:57.020
 How much of language understanding

54:57.020 --> 54:59.180
 can be solved 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 --> 55:18.300
 and something that may be to me or to us,

55:18.300 --> 55:21.620
 so the humans feels like it needs real understanding.

55:21.620 --> 55:23.500
 How much can I be achieved

55:23.500 --> 55:27.140
 with these neural networks or statistical methods?

55:28.060 --> 55:33.060
 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 --> 55:46.060
 which are trained on large amounts of data,

55:46.060 --> 55:48.820
 they really can do a remarkable job

55:48.820 --> 55:51.380
 relatively to where they've been a few years ago.

55:51.380 --> 55:54.660
 And if you project into the future,

55:54.660 --> 55:57.020
 if it will be the same speed of improvement,

55:59.380 --> 56:00.220
 this is great.

56:00.220 --> 56:01.060
 Now, does it bother me

56:01.060 --> 56:04.900
 that it's not doing the same translation as we are doing?

56:04.900 --> 56:06.660
 Now, if you go to cognitive science,

56:06.660 --> 56:09.460
 we still don't really understand what we are doing.

56:10.460 --> 56:11.900
 I mean, there are a lot of theories

56:11.900 --> 56:13.860
 and there is obviously a lot of progress and studying,

56:13.860 --> 56:17.580
 but our understanding what exactly goes on in our brains

56:17.580 --> 56:21.060
 when we process language is still not crystal clear

56:21.060 --> 56:25.500
 and precise that we can translate it into machines.

56:25.500 --> 56:29.820
 What does bother me is that, again,

56:29.820 --> 56:31.740
 that machines can be extremely brittle

56:31.740 --> 56:34.540
 when you go out of your comfort zone of there,

56:34.540 --> 56:36.100
 when there is a distributional shift

56:36.100 --> 56:37.340
 between training and testing.

56:37.340 --> 56:39.060
 And it have been years and years,

56:39.060 --> 56:41.540
 every year when they teach a NLP class,

56:41.540 --> 56:43.580
 show them some examples of translation

56:43.580 --> 56:45.740
 from some newspaper in Hebrew,

56:45.740 --> 56:47.340
 whatever, it was perfect.

56:47.340 --> 56:48.860
 And then they have a recipe

56:48.860 --> 56:52.620
 that Tomi Akala's system sent me a while ago

56:52.620 --> 56:55.740
 and it was written in Finnish of Carillian pies.

56:55.740 --> 56:59.340
 And it's just a terrible translation.

56:59.340 --> 57:01.500
 You cannot understand anything what it does.

57:01.500 --> 57:03.180
 It's not like some syntactic mistakes.

57:03.180 --> 57:04.340
 It's just terrible.

57:04.340 --> 57:07.140
 And year after year, I tried it and will translate it.

57:07.140 --> 57:09.020
 And year after year, it does this terrible work

57:09.020 --> 57:12.060
 because I guess the recipes are not big part

57:12.060 --> 57:14.620
 of their training repertoire.

57:15.500 --> 57:17.780
 So, but in terms of outcomes,

57:17.780 --> 57:20.260
 that's a really clean, good way to look at it.

57:21.140 --> 57:23.180
 I guess the question I was asking is,

57:24.100 --> 57:27.740
 do you think, imagine a future,

57:27.740 --> 57:29.820
 do you think the current approaches

57:29.820 --> 57:32.540
 can pass the Turing test in the way,

57:34.740 --> 57:37.060
 in the best possible formulation of the Turing test?

57:37.060 --> 57:39.500
 Which is, would you want to have a conversation

57:39.500 --> 57:42.380
 with a neural network for an hour?

57:42.380 --> 57:45.860
 Oh God, no, no, there are not that many people

57:45.860 --> 57:48.140
 that I would want to talk for an hour.

57:48.140 --> 57:51.540
 But there are some people in this world, alive or not,

57:51.540 --> 57:53.300
 that you would like to talk to for an hour,

57:53.300 --> 57:56.740
 could a neural network achieve that outcome?

57:56.740 --> 57:58.220
 So I think it would be really hard

57:58.220 --> 58:01.180
 to create a successful training set,

58:01.180 --> 58:03.380
 which would enable it to have a conversation

58:03.380 --> 58:07.140
 for an intercontextual conversation for an hour.

58:07.140 --> 58:08.180
 So you think it's a problem of data, perhaps?

58:08.180 --> 58:09.980
 I think in some ways it's an important data.

58:09.980 --> 58:12.500
 It's a problem both of data and the problem

58:12.500 --> 58:15.780
 of the way we are training our systems,

58:15.780 --> 58:18.100
 their ability to truly to generalize,

58:18.100 --> 58:21.300
 to be very compositional, in some ways, it's limited,

58:21.300 --> 58:24.120
 in the current capacity, at least.

58:25.580 --> 58:28.020
 You know, we can translate well,

58:28.020 --> 58:32.540
 we can find information well, we can extract information.

58:32.540 --> 58:35.220
 So there are many capacities in which it's doing very well.

58:35.220 --> 58:38.020
 And you can ask me, would you trust the machine

58:38.020 --> 58:39.900
 to translate for you and use it as a source?

58:39.900 --> 58:41.900
 I would say absolutely, especially if we're talking

58:41.900 --> 58:44.180
 about newspaper data or other data,

58:44.180 --> 58:46.780
 which is in the realm of its own training set,

58:46.780 --> 58:47.940
 I would say yes.

58:48.940 --> 58:52.940
 But, you know, having conversations with the machine,

58:52.940 --> 58:56.500
 it's not something that I would choose to do.

58:56.500 --> 58:58.180
 But you know, I would tell you something,

58:58.180 --> 58:59.460
 talking about Turing tests

58:59.460 --> 59:02.980
 and about all this kind of ELISA conversations.

59:02.980 --> 59:05.580
 I remember visiting Tencent in China

59:05.580 --> 59:06.980
 and they have this chat board.

59:06.980 --> 59:09.540
 And they claim that it is like really humongous amount

59:09.540 --> 59:10.820
 of the local population,

59:10.820 --> 59:12.980
 which like for hours talks to the chat board,

59:12.980 --> 59:15.380
 to me it was, I cannot believe it,

59:15.380 --> 59:17.140
 but apparently it's like documented

59:17.140 --> 59:20.820
 that there are some people who enjoy this conversation.

59:20.820 --> 59:24.580
 And you know, it brought to me another MIT story

59:24.580 --> 59:26.940
 about ELISA and Weizimbau.

59:26.940 --> 59:29.380
 I don't know if you're familiar with the story.

59:29.380 --> 59:31.060
 So Weizimbau was a professor at MIT

59:31.060 --> 59:32.620
 and when he developed this ELISA,

59:32.620 --> 59:36.740
 which was just doing string matching, very trivial,

59:36.740 --> 59:38.580
 like restating of what you said,

59:38.580 --> 59:41.300
 with very few rules, no syntax.

59:41.300 --> 59:43.780
 Apparently there were secretaries at MIT

59:43.780 --> 59:48.220
 that would sit for hours and converse with this trivial thing.

59:48.220 --> 59:50.220
 And at the time there was no beautiful interfaces.

59:50.220 --> 59:53.580
 So you actually need to go through the pain of communicating.

59:53.580 --> 59:56.980
 And Weizimbau himself was so horrified by this phenomenon

59:56.980 --> 59:59.300
 that people can believe enough to the machine

59:59.300 --> 1:00:00.860
 that you just need to give them the hint

1:00:00.860 --> 1:00:02.060
 that machine understands you

1:00:02.060 --> 1:00:03.980
 and you can complete the rest.

1:00:03.980 --> 1:00:05.460
 So he kind of stopped this research

1:00:05.460 --> 1:00:08.700
 and went into kind of trying to understand

1:00:08.700 --> 1:00:11.500
 what this artificial intelligence can do to our brains.

1:00:12.780 --> 1:00:15.580
 So my point is, you know, how much,

1:00:15.580 --> 1:00:19.380
 it's not how good is the technology,

1:00:19.380 --> 1:00:22.660
 it's how ready we are to believe

1:00:22.660 --> 1:00:25.620
 that it delivers the good that we are trying to get.

1:00:25.620 --> 1:00:27.220
 That's a really beautiful way to put it.

1:00:27.220 --> 1:00:29.780
 I, by the way, I'm not horrified by that possibility

1:00:29.780 --> 1:00:34.780
 but inspired by it because, I mean, human connection,

1:00:35.940 --> 1:00:38.100
 whether it's through language or through love,

1:00:39.180 --> 1:00:44.180
 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 --> 1:00:58.060
 as enjoying spending hours with those we love

1:00:58.060 --> 1:01:00.860
 for very silly reasons.

1:01:00.860 --> 1:01:02.820
 All we're doing is keyword matching as well.

1:01:02.820 --> 1:01:05.140
 So I'm not sure how much intelligence

1:01:05.140 --> 1:01:08.180
 we exhibit to each other with the people we love

1:01:08.180 --> 1:01:09.860
 that we're close with.

1:01:09.860 --> 1:01:12.700
 So it's a very interesting point

1:01:12.700 --> 1:01:16.060
 of what it means to pass the Turing test with language.

1:01:16.060 --> 1:01:16.900
 I think you're right.

1:01:16.900 --> 1:01:18.260
 In terms of conversation,

1:01:18.260 --> 1:01:23.140
 I think machine translation has very clear performance

1:01:23.140 --> 1:01:24.460
 and improvement, right?

1:01:24.460 --> 1:01:28.060
 What it means to have a fulfilling conversation

1:01:28.060 --> 1:01:31.060
 is very, very person dependent

1:01:31.060 --> 1:01:33.580
 and context dependent and so on.

1:01:33.580 --> 1:01:36.060
 That's, yeah, it's very well put.

1:01:36.060 --> 1:01:38.340
 So, but in your view,

1:01:38.340 --> 1:01:41.940
 what's a benchmark in natural language, a test,

1:01:41.940 --> 1:01:43.700
 that's just out of reach right now,

1:01:43.700 --> 1:01:46.060
 but we might be able to, that's exciting.

1:01:46.060 --> 1:01:49.140
 Is it in machine, isn't perfecting machine translation

1:01:49.140 --> 1:01:51.940
 or is there other, is it summarization?

1:01:51.940 --> 1:01:53.340
 What's out there just out of reach?

1:01:53.340 --> 1:01:55.860
 It goes across specific application.

1:01:55.860 --> 1:01:58.300
 It's more about the ability to learn

1:01:58.300 --> 1:02:00.100
 from few examples for real,

1:02:00.100 --> 1:02:03.340
 what we call future planning and all these cases.

1:02:03.340 --> 1:02:05.940
 Because, you know, the way we publish these papers today,

1:02:05.940 --> 1:02:09.940
 we say, if we have like naively, we get 55,

1:02:09.940 --> 1:02:12.500
 but now we had a few example and we can move to 65.

1:02:12.500 --> 1:02:14.020
 None of these methods actually

1:02:14.020 --> 1:02:15.980
 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 --> 1:02:28.980
 or to be autonomous in finding the data

1:02:28.980 --> 1:02:30.300
 that you need to learn,

1:02:31.340 --> 1:02:34.260
 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

1:02:39.220 --> 1:02:43.060
 to move forward to and we are not yet there.

1:02:43.060 --> 1:02:45.060
 Are you at all excited,

1:02:45.060 --> 1:02:48.540
 curious by the possibility of creating human level intelligence?

1:02:49.900 --> 1:02:52.540
 Is this, because 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 --> 1:03:09.860
 But, you know, human level intelligence is a thing

1:03:09.860 --> 1:03:13.100
 that our civilizations dream about creating

1:03:13.100 --> 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 --> 1:03:19.140
 Do you think it's at all within our reach?

1:03:20.420 --> 1:03:22.660
 So as you said yourself earlier,

1:03:22.660 --> 1:03:26.660
 talking about, you know, how do you perceive,

1:03:26.660 --> 1:03:28.980
 you know, our communications with each other

1:03:28.980 --> 1:03:30.700
 that, you know, we're matching keywords

1:03:30.700 --> 1:03:33.020
 and certain behaviors and so on.

1:03:33.020 --> 1:03:36.860
 So at the end, whenever one assesses,

1:03:36.860 --> 1:03:38.660
 let's say relations with another person,

1:03:38.660 --> 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, you know,

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 --> 1:03:49.580
 What is the intelligence?

1:03:49.580 --> 1:03:51.260
 Is it the fact that now we are going to do

1:03:51.260 --> 1:03:52.940
 the same way 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 an LPL,

1:04:01.300 --> 1:04:03.940
 not just to translate or just to answer questions,

1:04:03.940 --> 1:04:06.900
 but across many, many areas that we can achieve

1:04:06.900 --> 1:04:09.740
 the functionalities that humans can achieve

1:04:09.740 --> 1:04:12.380
 with their ability to learn 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 --> 1:04:20.340
 how far we are, and that's what makes me excited

1:04:20.340 --> 1:04:22.420
 that we, you know, in my lifetime,

1:04:22.420 --> 1:04:23.780
 at least so far what we've seen,

1:04:23.780 --> 1:04:26.260
 it's like tremendous progress across

1:04:26.260 --> 1:04:28.700
 with 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 --> 1:04:40.020
 And again, one way to think about is there are machines

1:04:40.020 --> 1:04:41.820
 which are improving their functionality.

1:04:41.820 --> 1:04:44.900
 Another one is to think about us with our brains,

1:04:44.900 --> 1:04:49.020
 which are imperfect, how they can be accelerated

1:04:49.020 --> 1:04:54.020
 by this technology as it becomes stronger and stronger.

1:04:55.860 --> 1:04:58.580
 Coming back to another book that I love,

1:04:58.580 --> 1:05:02.060
 Flowers for Algernon, have you read this book?

1:05:02.060 --> 1:05:02.900
 Yes.

1:05:02.900 --> 1:05:05.740
 You know, there is this point that the patient gets

1:05:05.740 --> 1:05:08.020
 this miracle cure which changes his brain

1:05:08.020 --> 1:05:11.060
 and all of a sudden they see life in a different way

1:05:11.060 --> 1:05:13.340
 and can do certain things better,

1:05:13.340 --> 1:05:14.900
 but certain things much worse.

1:05:16.540 --> 1:05:21.540
 So you can imagine this kind of computer augmented cognition

1:05:22.420 --> 1:05:24.820
 where it can bring you that now in the same way

1:05:24.820 --> 1:05:28.140
 as, you know, the cars enable us to get to places

1:05:28.140 --> 1:05:30.100
 where we've never been before.

1:05:30.100 --> 1:05:31.620
 Can we think differently?

1:05:31.620 --> 1:05:32.820
 Can we think faster?

1:05:32.820 --> 1:05:36.700
 So, and we already see a lot of it happening

1:05:36.700 --> 1:05:38.260
 in how it impacts us,

1:05:38.260 --> 1:05:42.180
 but I think we have a long way to go there.

1:05:42.180 --> 1:05:45.020
 So that's sort of artificial intelligence

1:05:45.020 --> 1:05:47.260
 and technology affecting our,

1:05:47.260 --> 1:05:50.500
 augmenting our intelligence as humans.

1:05:50.500 --> 1:05:55.500
 Yesterday, a company called Neuralink announced

1:05:55.540 --> 1:05:56.820
 they did this whole demonstration.

1:05:56.820 --> 1:05:57.980
 I don't know if you saw it.

1:05:57.980 --> 1:06:00.980
 It's, they demonstrated brain, computer,

1:06:00.980 --> 1:06:05.260
 brain machine interface where there's like a sewing machine

1:06:05.260 --> 1:06:06.340
 for the brain.

1:06:06.340 --> 1:06:11.140
 Do you, you know, a lot of that is quite out there

1:06:11.140 --> 1:06:15.300
 in terms of things that some people would say are impossible,

1:06:15.300 --> 1:06:18.100
 but they're dreamers and want to engineer systems like that.

1:06:18.100 --> 1:06:20.380
 Do you see, based on what you just said,

1:06:20.380 --> 1:06:23.820
 a hope for that more direct interaction with the brain?

1:06:25.140 --> 1:06:27.020
 I think there are different ways.

1:06:27.020 --> 1:06:28.980
 One is a direct interaction with the brain.

1:06:28.980 --> 1:06:30.900
 And again, there are lots of companies

1:06:30.900 --> 1:06:32.220
 that work in this space.

1:06:32.220 --> 1:06:35.060
 And I think there will be a lot of developments.

1:06:35.060 --> 1:06:36.540
 When I'm just thinking that many times

1:06:36.540 --> 1:06:39.020
 we are not aware of our feelings

1:06:39.020 --> 1:06:41.420
 of motivation, what drives us.

1:06:41.420 --> 1:06:44.100
 Like let me give you a trivial example, our attention.

1:06:45.500 --> 1:06:47.260
 There are a lot of studies that demonstrate

1:06:47.260 --> 1:06:49.220
 that it takes a while to a person to understand

1:06:49.220 --> 1:06:51.100
 that they are not attentive anymore.

1:06:51.100 --> 1:06:52.180
 And we know that there are people

1:06:52.180 --> 1:06:54.540
 who really have strong capacity to hold attention.

1:06:54.540 --> 1:06:55.980
 There are another end of the spectrum,

1:06:55.980 --> 1:06:57.980
 people with ADD and other issues

1:06:57.980 --> 1:07:00.740
 that they have problem to regulate their attention.

1:07:00.740 --> 1:07:03.540
 Imagine to yourself that you have like a cognitive aid

1:07:03.540 --> 1:07:06.260
 that just alerts you based on your gaze.

1:07:06.260 --> 1:07:09.300
 That your attention is now not on what you are doing.

1:07:09.300 --> 1:07:11.460
 And instead of writing a paper, you're now dreaming

1:07:11.460 --> 1:07:12.740
 of what you're gonna do in the evening.

1:07:12.740 --> 1:07:16.340
 So even this kind of simple measurement things,

1:07:16.340 --> 1:07:18.020
 how they can change us.

1:07:18.020 --> 1:07:22.380
 And I see it even in the simple ways with myself.

1:07:22.380 --> 1:07:26.460
 I have my zone up from that I got in MIT gym.

1:07:26.460 --> 1:07:28.780
 It kind of records how much did you run

1:07:28.780 --> 1:07:31.940
 and you have some points and you can get some status,

1:07:31.940 --> 1:07:32.900
 whatever.

1:07:32.900 --> 1:07:35.860
 Like I said, what is this ridiculous thing?

1:07:35.860 --> 1:07:38.820
 Who would ever care about some status in some arm?

1:07:38.820 --> 1:07:39.660
 Guess what?

1:07:39.660 --> 1:07:41.580
 So to maintain the status,

1:07:41.580 --> 1:07:44.660
 you have to set a number of points every month.

1:07:44.660 --> 1:07:48.060
 And not only is that they do it every single month

1:07:48.060 --> 1:07:50.580
 for the last 18 months,

1:07:50.580 --> 1:07:54.180
 it went to the point that I was injured.

1:07:54.180 --> 1:07:56.180
 And when I could run again,

1:07:56.180 --> 1:08:01.180
 I in two days, I did like some humongous amount

1:08:01.860 --> 1:08:04.020
 of writing just to complete the points.

1:08:04.020 --> 1:08:05.820
 It was like really not safe.

1:08:05.820 --> 1:08:08.340
 It's like, I'm not gonna lose my status

1:08:08.340 --> 1:08:10.100
 because I want to get there.

1:08:10.100 --> 1:08:13.180
 So you can already see that this direct measurement

1:08:13.180 --> 1:08:16.180
 and the feedback, we're looking at video games

1:08:16.180 --> 1:08:18.540
 and see why the addiction aspect of it,

1:08:18.540 --> 1:08:20.340
 but you can imagine that the same idea

1:08:20.340 --> 1:08:23.500
 can be expanded to many other areas of our life

1:08:23.500 --> 1:08:25.820
 when we really can get feedback

1:08:25.820 --> 1:08:28.380
 and imagine in your case in relations

1:08:29.740 --> 1:08:31.220
 when we are doing keyword matching,

1:08:31.220 --> 1:08:36.100
 imagine that the person who is generating the key ones,

1:08:36.100 --> 1:08:37.700
 that person gets direct feedback

1:08:37.700 --> 1:08:39.540
 before the whole thing explodes.

1:08:39.540 --> 1:08:41.940
 Is it maybe at this happy point,

1:08:41.940 --> 1:08:43.980
 we are going in the wrong direction?

1:08:43.980 --> 1:08:47.980
 Maybe it will be really a behavior modifying moment.

1:08:47.980 --> 1:08:51.300
 So yeah, it's a relationship management too.

1:08:51.300 --> 1:08:54.180
 So yeah, that's a fascinating whole area

1:08:54.180 --> 1:08:56.100
 of psychology actually as well,

1:08:56.100 --> 1:08:58.220
 of seeing how our behavior has changed

1:08:58.220 --> 1:09:00.820
 with basically all human relations

1:09:00.820 --> 1:09:05.820
 now have other non human entities helping us out.

1:09:06.180 --> 1:09:09.420
 So you've, you teach a large,

1:09:09.420 --> 1:09:12.600
 a huge machine learning course here at MIT.

1:09:13.620 --> 1:09:15.340
 I can ask you a million questions,

1:09:15.340 --> 1:09:17.580
 but you've seen a lot of students.

1:09:17.580 --> 1:09:20.900
 What ideas do students struggle with the most

1:09:20.900 --> 1:09:23.940
 as they first enter this world of machine learning?

1:09:25.700 --> 1:09:28.020
 Actually, this year was the first time

1:09:28.020 --> 1:09:30.060
 I started teaching a small machine learning class

1:09:30.060 --> 1:09:32.860
 and it came as a result of what I saw

1:09:32.860 --> 1:09:35.620
 in my big machine learning class that Tommy Ackle

1:09:35.620 --> 1:09:38.280
 and I built maybe six years ago.

1:09:39.660 --> 1:09:42.820
 What we've seen that as this area become more and more popular,

1:09:42.820 --> 1:09:46.940
 more and more people at MIT want to take this class.

1:09:46.940 --> 1:09:49.940
 And while we designed it for computer science majors,

1:09:49.940 --> 1:09:52.340
 there were a lot of people who really are interested

1:09:52.340 --> 1:09:54.220
 to learn it, but unfortunately,

1:09:54.220 --> 1:09:57.380
 their background was not enabling them

1:09:57.380 --> 1:09:58.780
 to do well in the class.

1:09:58.780 --> 1:10:01.000
 And many of them associated machine learning

1:10:01.000 --> 1:10:02.980
 with a world struggle and failure,

1:10:04.380 --> 1:10:06.380
 primarily for non majors.

1:10:06.380 --> 1:10:08.660
 And that's why we actually started a new class

1:10:08.660 --> 1:10:12.620
 which we call machine learning from algorithms to modeling,

1:10:12.620 --> 1:10:16.740
 which emphasizes more the modeling aspects of it

1:10:16.740 --> 1:10:21.700
 and focuses on, it has majors and non majors.

1:10:21.700 --> 1:10:25.300
 So we kind of try to extract the relevant parts

1:10:25.300 --> 1:10:27.340
 and make it more accessible

1:10:27.340 --> 1:10:29.620
 because the fact that we're teaching 20 classifiers

1:10:29.620 --> 1:10:31.020
 in standard machine learning class

1:10:31.020 --> 1:10:34.100
 is really a big question we really needed.

1:10:34.100 --> 1:10:36.380
 But it was interesting to see this

1:10:36.380 --> 1:10:38.300
 from first generation of students,

1:10:38.300 --> 1:10:40.900
 when they came back from their internships

1:10:40.900 --> 1:10:43.940
 and from their jobs,

1:10:43.940 --> 1:10:47.460
 what different and exciting things they can do

1:10:47.460 --> 1:10:48.380
 is that they would never think

1:10:48.380 --> 1:10:51.100
 that you can even apply machine learning to.

1:10:51.100 --> 1:10:53.740
 Some of them are like matching their relations

1:10:53.740 --> 1:10:56.020
 and other things like variety of different applications.

1:10:56.020 --> 1:10:58.060
 Everything is amenable to machine learning.

1:10:58.060 --> 1:11:00.260
 That actually brings up an interesting point

1:11:00.260 --> 1:11:02.580
 of computer science in general.

1:11:02.580 --> 1:11:05.420
 It almost seems, maybe I'm crazy,

1:11:05.420 --> 1:11:08.420
 but it almost seems like everybody needs to learn

1:11:08.420 --> 1:11:10.060
 how to program these days.

1:11:10.060 --> 1:11:13.340
 If you're 20 years old or if you're starting school,

1:11:13.340 --> 1:11:15.900
 even if you're an English major,

1:11:15.900 --> 1:11:18.100
 it seems like programming

1:11:18.980 --> 1:11:21.860
 unlocks so much possibility in this world.

1:11:21.860 --> 1:11:24.980
 So when you interact with those non majors,

1:11:24.980 --> 1:11:29.980
 is there skills that they were simply lacking at the time

1:11:30.220 --> 1:11:31.980
 that you wish they had

1:11:31.980 --> 1:11:34.620
 and that they learned in high school and so on?

1:11:34.620 --> 1:11:37.460
 Like how should education change

1:11:37.460 --> 1:11:41.260
 in this computerized world that we live in?

1:11:41.260 --> 1:11:43.500
 So seeing because they knew that there is a Python component

1:11:43.500 --> 1:11:44.820
 in the class,

1:11:44.820 --> 1:11:47.020
 their Python skills were okay

1:11:47.020 --> 1:11:49.140
 and the class is not really heavy on programming.

1:11:49.140 --> 1:11:52.420
 They primarily kind of add parts to the programs.

1:11:52.420 --> 1:11:55.420
 I think it was more of their mathematical barriers

1:11:55.420 --> 1:11:58.220
 and the class, again, with the design on the majors

1:11:58.220 --> 1:12:01.220
 was using the notation like big O for complexity

1:12:01.220 --> 1:12:04.540
 and others, people who come from different backgrounds

1:12:04.540 --> 1:12:05.820
 just don't have it in the lexical.

1:12:05.820 --> 1:12:09.140
 So necessarily very challenging notion,

1:12:09.140 --> 1:12:11.500
 but they were just not aware.

1:12:12.380 --> 1:12:15.340
 So I think that, you know, kind of linear algebra

1:12:15.340 --> 1:12:17.660
 and probability, the basics, the calculus,

1:12:17.660 --> 1:12:20.860
 want to vary the calculus, things that can help.

1:12:20.860 --> 1:12:23.580
 What advice would you give to students

1:12:23.580 --> 1:12:26.620
 interested in machine learning, interested,

1:12:26.620 --> 1:12:30.100
 if you've talked about detecting curing cancer,

1:12:30.100 --> 1:12:33.140
 drug design, if they want to get into that field,

1:12:33.140 --> 1:12:34.540
 what should they do?

1:12:36.380 --> 1:12:39.100
 Get into it and succeed as researchers

1:12:39.100 --> 1:12:42.100
 and entrepreneurs.

1:12:43.300 --> 1:12:45.260
 The first good piece of news is that right now

1:12:45.260 --> 1:12:47.420
 there are lots of resources

1:12:47.420 --> 1:12:50.180
 that are created at different levels

1:12:50.180 --> 1:12:51.860
 and you can find online

1:12:51.860 --> 1:12:54.820
 on your school classes,

1:12:54.820 --> 1:12:57.580
 which are more mathematical or more applied and so on.

1:12:57.580 --> 1:13:01.340
 So you can find a kind of a preacher

1:13:01.340 --> 1:13:02.780
 which preaches your own language

1:13:02.780 --> 1:13:04.580
 where you can enter the field

1:13:04.580 --> 1:13:06.740
 and you can make many different types of contribution

1:13:06.740 --> 1:13:09.620
 depending of what is your strengths.

1:13:10.740 --> 1:13:12.020
 And the second point,

1:13:12.020 --> 1:13:15.300
 I think it's really important to find some area

1:13:15.300 --> 1:13:18.140
 which you really care about

1:13:18.140 --> 1:13:20.220
 and it can motivate your learning

1:13:20.220 --> 1:13:22.620
 and it can be for somebody curing cancer

1:13:22.620 --> 1:13:25.380
 or doing cell driving cars or whatever,

1:13:25.380 --> 1:13:29.660
 but to find an area where there is data

1:13:29.660 --> 1:13:31.300
 where you believe there are strong patterns

1:13:31.300 --> 1:13:32.340
 and we should be doing it

1:13:32.340 --> 1:13:33.580
 and we're still not doing it

1:13:33.580 --> 1:13:35.260
 or you can do it better

1:13:35.260 --> 1:13:37.860
 and just start there

1:13:37.860 --> 1:13:39.700
 and see a way it can bring you.

1:13:40.780 --> 1:13:44.060
 So you've been very successful

1:13:44.060 --> 1:13:45.580
 in many directions in life,

1:13:46.460 --> 1:13:48.860
 but you also mentioned Flowers of Argonaut.

1:13:51.020 --> 1:13:53.820
 And I think I've read or listened to you mention somewhere

1:13:53.820 --> 1:13:55.340
 that researchers often get lost

1:13:55.340 --> 1:13:56.740
 in the details of their work.

1:13:56.740 --> 1:14:00.220
 This is per our original discussion with cancer and so on

1:14:00.220 --> 1:14:02.180
 and don't look at the bigger picture,

1:14:02.180 --> 1:14:05.340
 the bigger questions of meaning and so on.

1:14:05.340 --> 1:14:07.460
 So let me ask you the impossible question

1:14:09.900 --> 1:14:11.620
 of what's the meaning of this thing,

1:14:11.620 --> 1:14:16.740
 of life, of your life, of research.

1:14:16.740 --> 1:14:21.460
 Why do you think we descendant of great apes

1:14:21.460 --> 1:14:24.500
 are here on this spinning ball?

1:14:26.820 --> 1:14:30.300
 You know, I don't think that I have really a global answer

1:14:30.300 --> 1:14:32.900
 you know, maybe that's why I didn't go to humanities

1:14:33.780 --> 1:14:36.500
 and I didn't take humanities classes in my undergrad.

1:14:39.500 --> 1:14:43.580
 But the way I am thinking about it,

1:14:43.580 --> 1:14:48.220
 each one of us inside of them have their own set of,

1:14:48.220 --> 1:14:51.140
 you know, things that we believe are important.

1:14:51.140 --> 1:14:53.380
 And it just happens that we are busy

1:14:53.380 --> 1:14:54.820
 with achieving various goals,

1:14:54.820 --> 1:14:56.260
 busy listening to others

1:14:56.260 --> 1:14:58.100
 and to kind of try to conform

1:14:58.100 --> 1:15:03.100
 to be part of the crowd that we don't listen to that part.

1:15:04.580 --> 1:15:09.580
 And, you know, we all should find some time to understand

1:15:09.580 --> 1:15:11.820
 what is our own individual missions

1:15:11.820 --> 1:15:14.060
 and we may have very different missions

1:15:14.060 --> 1:15:18.180
 and to make sure that while we are running 10,000 things,

1:15:18.180 --> 1:15:21.900
 we are not, you know, missing out

1:15:21.900 --> 1:15:24.420
 and we're putting all the resources

1:15:24.420 --> 1:15:28.500
 to satisfy our own mission.

1:15:28.500 --> 1:15:31.500
 And if I look over my time,

1:15:31.500 --> 1:15:34.820
 when I was younger, most of these missions,

1:15:34.820 --> 1:15:38.620
 you know, I was primarily driven by the external stimulus,

1:15:38.620 --> 1:15:41.540
 you know, to achieve this or to be that.

1:15:41.540 --> 1:15:46.540
 And now a lot of what I do is driven by really thinking

1:15:47.660 --> 1:15:51.380
 what is important for me to achieve independently

1:15:51.380 --> 1:15:55.140
 of the external recognition.

1:15:55.140 --> 1:16:00.140
 And, you know, I don't mind to be viewed in certain ways.

1:16:01.380 --> 1:16:05.740
 The most important thing for me is to be true to myself,

1:16:05.740 --> 1:16:07.500
 to what I think is right.

1:16:07.500 --> 1:16:08.700
 How long did it take?

1:16:08.700 --> 1:16:13.220
 How hard was it to find the you that you have to be true to?

1:16:14.180 --> 1:16:17.740
 So it takes time and even now sometimes, you know,

1:16:17.740 --> 1:16:20.860
 the vanity and the triviality can take, you know.

1:16:20.860 --> 1:16:26.060
 Yeah, it can everywhere, you know, it's just the vanity.

1:16:26.060 --> 1:16:28.140
 The vanity is different, the vanity in different places,

1:16:28.140 --> 1:16:30.940
 but we all have our piece of vanity.

1:16:30.940 --> 1:16:34.700
 But I think actually for me,

1:16:34.700 --> 1:16:39.700
 the many times the place to get back to it is, you know,

1:16:41.700 --> 1:16:45.820
 when I'm alone and also when I read.

1:16:45.820 --> 1:16:47.740
 And I think by selecting the right books,

1:16:47.740 --> 1:16:52.740
 you can get the right questions and learn from what you read.

1:16:54.900 --> 1:16:58.060
 So, but again, it's not perfect,

1:16:58.060 --> 1:17:02.020
 like vanity sometimes dominates.

1:17:02.020 --> 1:17:04.780
 Well, that's a beautiful way to end.

1:17:04.780 --> 1:17:06.380
 Thank you so much for talking today.

1:17:06.380 --> 1:17:07.820
 Thank you. That was fun.

1:17:07.820 --> 1:17:17.820
 It was fun.