video_id
stringlengths 11
11
| text
stringlengths 361
490
| start_second
int64 0
11.3k
| end_second
int64 18
11.3k
| url
stringlengths 48
52
| title
stringlengths 0
100
| thumbnail
stringlengths 0
52
|
|---|---|---|---|---|---|---|
733m6qBH-jI | Um, and if the vertical area of the vertical axis is depth, uh, a lot of all the strongest candidates for jobs are, um, T-shaped individuals. Meaning that you have a broad understanding of many different topics in the AI machine learning, and very deep understanding in, you know, maybe at least one area. Maybe more than one area. Um, and so I think by taking CS230 and doing the things that you're doing here, | 2,113 | 2,139 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2113s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | hopefully you're acquiring a deeper understanding of one of these areas of deep learning in particular. Um, but the other thing that even, you know, deepens your knowledge in one area will be the projects you work on. Um, the open source contributions you make, right. Uh, whether or not you've done research. Um, and maybe whether or not you've done an internship. | 2,139 | 2,164 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2139s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | Right? Okay. And I think these two elements, you know, broad area of skills, and then also going deeper to do a meaningful project in deep learning. Or, um, work with a Stanford professor, right? And do a meaningful research project, or make some contribution to open-source. Publish it on GitHub, and then let us use it. These are the things that let you deepen your knowledge and, | 2,164 | 2,185 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2164s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | and convince recruiters that you both have the broad technical skills, and when called on you're able to apply these in a, in a, in a meaningful way to an important problem, right? And in fact, um, the way we design CS230 is actually a microcosm of this. Where, um, you know, you learned about neural nets. Um, then about topics like Batch Norm, ConvNets, sequence models, right? | 2,185 | 2,210 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2185s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | I'm just gonna say RNNs. So, actually you've a breadth within the field of deep learning. And then what happens is, well, then, and the reason I want you to work on the project is so that you can pick one of these areas. And maybe go deep and build a meaningful project in one of these areas, which will, which will, and it's not just about making a resume look good, right? | 2,210 | 2,235 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2210s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | It's about giving you the practical experience to make sure you actually know how to make these things work, um, uh, and give you the learning. To make sure that you actually know how to make a CNN work, to make a RNN work. All right. And then of course it stands many students also list their projects on their resumes obviously. Um, so, I think the um, let's see. | 2,235 | 2,260 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2235s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | The- the- the- failure modes. The things, bad ways to navigate your career. Um, there are some students they just do this, right? There are some Stanford students that just take class, after class, after class, after class, and go equally in depth in a huge range of areas. And this is not terrible. You can actually still got a job uh, uh you still get. Sometimes you can even get into some Ph.D. programs like this with all the depth, | 2,260 | 2,285 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2260s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | but this is not the best way to navigate your career. All right? So, there are some Stanford students who's- that takes tons of classes. You can get a good GPA doing that, but do nothing else. And this is not terrible, but this is- this is not- this is not great. It's not as good as the alternative. Um, there's one other thing I've seen Stanford students do which is, | 2,285 | 2,304 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2285s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | uh, just try to do that, right? But you just try to jump in on day one, and go really really deep in one area. And again, um, this has its own challenges, I guess. You know, one, one, one failure mode, one mode is actually not great. As sometimes you actually get, um, undergrad freshmen at Stanford that have not yet learned a lot about coding, or software engineering, or machine learning, | 2,304 | 2,329 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2304s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | and try to jump into research projects right away. This turns out not to be very efficient because it turns out Stanford classes are, your online courses or Stanford classes are a very efficient way for you to learn about the broad range of areas. And after that going deeper and getting experience in one vertical area then deepens your knowledge. It makes so you know how to actually make those ideas work. | 2,329 | 2,347 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2329s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | Uh, so I do see sometimes unfortunately, you know, som- some Stanford freshmen join us already knowing how to code and have implemented, you know, some learning algorithms, but some students that do not yet have much experience try to jump into research projects right away. And that turns out not to be very productive for the student or for the research group because until you've taken | 2,347 | 2,367 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2347s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | the classes and mastered the basics it's difficult to understand what's really going on in the advanced projects, right? Um, so I would, I, I would say this is actually worse than that, right? This is, this is actually okay. This is actually pretty bad. It is I, I, I would not do this for your career, right? Yeah. Probably not. Yeah. Um, and then the other not-so-great | 2,367 | 2,390 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2367s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | mode that you see some Stanford students do is get a lot of breadth, and then do a tiny project here. Do a tiny project there. Do a tiny project there. Do a tiny project there. And you end up with ten tiny projects, but no one or two really sec- significant projects. So again, this is not terrible, but, you know, beyond a certain point, by the way recruiters are not impressed by volume, right? | 2,390 | 2,416 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2390s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | So, having done 10 lame projects is actually not impressive. Not nearly as impressive as doing one great project or two great projects. And again, there's more to life than impressing recruiters, but recruiters are very rational. And the reason recruiters are less impressed by someone who's profile looks like this is because they're actually probably factually less skilled and less | 2,416 | 2,436 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2416s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | able at doing good work in machine learning compared to someone that, that has done a substantive project and knows what it takes to see, see the whole thing through. Does that make sense? So, when I say you'd have recruiters more or less impressed is because they're actually quite rational, in terms of, uh, trying to understand how good you are at um, uh, at, at, doing important work, | 2,436 | 2,456 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2436s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | at building meaningful AI systems. Makes sense? Um, and so in terms of building up both the horizontal piece and vertical piece, uh, this is what I recommend. Um, to build a horizontal piece, a lot of this is about building foundational skills. And, um, it turns out coursework is a very efficient way to do this. Uh, you know, in, in, in these courses, right, you know various instructors like us, | 2,456 | 2,486 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2456s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | but many other Stanford professors, um, put a lot of work into organizing the content to make it efficient for you to learn this material. Um, and then also reading research papers which we just talked about. Having a community will help you. Um, and then that is often, uh, building a more deep and, um, relevant project, right? And, and, and the pro- projects have to be relevant. | 2,486 | 2,519 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2486s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | So, you know, if you want to build a career machine learning, build a career in AI. Hopefully, the project is something that's relevant to CS, or machine learning, or AI deep learning. Um, I do see, I don't know, for some reason, I feel like, uh, a surprisingly large number of Stanford students I know are in the Stanford dance crew, and they spend a lot of time on that which is fine. | 2,519 | 2,538 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2519s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | If you enjoy dancing, go have fun. Don't, don't, you know, you, you don't need to work all the time. So, go join the dance crew, or go on the overseas exchange program. And go hang out in London and have fun, but those things do not as directly contribute to this, right? Yeah. I know, I think, I think, in an earlier version of this presentation, you know, students walked away, saying ha, you know, | 2,538 | 2,563 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2538s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | Andrew says we should not have fun we should work all the time and that's not the goal [LAUGHTER]. Um, All right. There is one. All right. Um, you know, there is the uh, Saturday morning problem which all of you will face. Right? Which is every week, uh, including this week on Saturday morning you have a choice. Um, you can, uh, read a paper [LAUGHTER] or work on research | 2,563 | 2,626 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2563s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | or work on open source or, I don't know what people do, or you can watch TV or something, [LAUGHTER] right? Um, and you will face this choice, like, maybe every Saturday, you know, for the rest of your life or for all Saturdays in the rest of your life. And, um, you know, you can build out that foundation skills, go deep or go have fun, and you should have fun, all right? Just for the record. | 2,626 | 2,651 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2626s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | But one of the problems that a lot of people face is that, um, even if you spend all Saturday and all Sunday reading a research paper, um, you know, the following Monday, or maybe spend all Saturday and Sunday working hard, it turns out that the following Monday, you're not that much better at deep learning. Is like, yeah, you work really hard. So you read five papers, you know, great. | 2,651 | 2,673 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2651s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | Uh, but if you work on a research group the professor or your manager if you're in a company, they have no idea how hard you work. So there's no one to come by and say ''Oh, good job working so hard all weekend.'' So no one knows these sacrifices you make all weekend to study or code open source, just no one knows. So there's almost no short-term reward to doing these things. | 2,673 | 2,694 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2673s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | Um, but they see- and, and, and, and whereas there might be short-term rewards for doing other things, right? Um, uh, but the secret to this is that it's not about reading papers really, really hard for one Saturday morning or for all Saturday once and it being done. The secret to this is to do this consistently, um, you know, for years, um, or at least a month. | 2,694 | 2,718 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2694s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | And it turns out that if you read, um, two papers a week, and you do that for a year then you have read 50 papers after a year and you will be much better at deep learning after that, right? I mean when you read, you have read 100 papers in the year if you read two papers a week. And so for you to be successful is much less about the intense burst of effort you put in over one weekend. | 2,718 | 2,742 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2718s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | It's much more about whether you can find a little bit of time every week to read a few papers or contribute to open source or take some online courses, uh, but- and if you do that you know every week for six months or do that every week for a year, you will actually learn a lot about these fields and be much better off, and be much more capable at deep learning and machine learning or whatever, right? | 2,742 | 2,763 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2742s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | Um, yeah. So, um, yeah, and yeah she- my wife and I actually do not own a TV. [LAUGHTER] For what it's worth. Okay, but again, if you own one go ahead. Make sure- don't, don't drive yourself crazy. There's a healthy work-life integration as well. All right. So, um, so I hope that doing these things more is not about finding a job, it's about doing these things and make you more capable as a machine learning person, | 2,763 | 2,800 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2763s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | so that you have the power to go out and implement stuff that matters, right? To do stuff that actually, do, do work that matters. Well the second thing we'll chat about is selecting a job, right? And it's actually really interesting. Um, I, uh, gave this part of presentation, um, last year, uh, actually sorry earlier this year and shortly after that presentation, um, | 2,800 | 2,824 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2800s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | there was a student in the class that was already in a company who emailed me saying, "Boy Andrew, I wish you'd told me this before I accepted my current job." Um, and so [LAUGHTER] let's see. Hopefully this will be useful to you. Um, so it turns out that ,um, uh, you know, I, so when you're- at some point you're deciding, you know, what Ph.D program do you want to apply for, | 2,824 | 2,849 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2824s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | what companies you want to apply for a job at and, um, I can tell you what, uh, so if you want to keep learning new things, um, I think one of the biggest predictors of your success will be whether or not you're working with great people and projects, right? And in particular, um, you know, there are these fascinating results from, uh, what are these, I wanna say from the social sciences showing that, | 2,849 | 2,884 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2849s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | um, if your closest friends, if your five closest friends or ten closest friends are all smokers, there's a much higher chance you become a smoker as well, right? And if your five or 10 close friends are, uh, um, you know, overweight, there's a much high chance you'd do the same or- and conversely if there's a, you know, so I think that if your five closest friends work really hard, | 2,884 | 2,905 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2884s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | read a lot of research papers, care about their work, learning and making themselves better, then there's actually a very good chance that you will be, that they'll influence you that way as well. So we're all human. We're all influenced by the people around us, right? And so, um, I think that- and I've been fortunate, I've taught at Stanford for a long time now, | 2,905 | 2,924 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2905s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | so I've been fortunate to have seen a lot of students go from Stanford to various careers and because I've seen how many hundreds or maybe thousands of Stanford students, that I knew right, when they were still Stanford students, go on to separate jobs. I saw many of them have amazing careers. Um, I saw, you know, a few have, like, okay careers, um, that I think over time I've learned to pattern match what is | 2,924 | 2,948 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2924s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | predictive of your future success after you leave Stanford and I'll share with you some of those patterns, share with you some of those patterns as you navigate your career. And it's just there's so many options in machine learning today that its's kind of tragic if you don't, you know, navigate to hopefully maximize your chance of being one of the people that gets to do fun and important work that helps others. | 2,948 | 2,967 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2948s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | Um, so when selecting a position, um, I would advise you to focus on the team, um, [NOISE] you interact with and by team I mean, you know, somewhere between 10 to 30 persons, right, maybe up to 50, right? Um, because it turns out that if yo- there will be some group of people. Maybe 10 to 30 people, maybe 50 people that you interact with quite closely and these will be | 2,967 | 3,003 | https://www.youtube.com/watch?v=733m6qBH-jI&t=2967s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | appears in the people that will influence you the most, right? Um, because if you join a company with 10,000 people, you will not interact with all 10,000 people. There will be a core of 10 or 30 or 50 people that you interact with the most, and it's those people how much they know, how much they teach you, how hard working they are, whether they're learning themselves that will influence you the most, | 3,003 | 3,025 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3003s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | rather than all these other hypothetical 10,000 people in a giant company. Um, and of these people, one of the ones that will influence you the most is your manager, all right? So make sure you meet your manager and get to know them and make sure they're someone you want to work with. Um, and in particular, I would recommend focusing on these things and not on the brand, | 3,025 | 3,050 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3025s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | um, of the company. Because it turns out that the brand of the company you work with is actually not that correlated. Yeah maybe there's a very weak correlation, but it's actually not that correlated with what your personal experience would be like if that makes sense, right? Um, and so, um, [NOISE] and by the way, again, just full disclosure. I'm one of the- I have a research group here at Stanford, right? | 3,050 | 3,081 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3050s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | My research group at Stanford is one of the better known research groups in the world but just don't join us because you think we are well-known, right? It's just not a good reason to join us for the brand. Instead, if you want to work with someone, meet the people and evaluate the individuals, or look at the people and see if you think these are people you can learn from and work with, | 3,081 | 3,098 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3081s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | and are good people, makes sense? [NOISE] So, um, in today's world there are a lot of companies, um, recruiting Stanford students. So let me give you some advice. This piece I only give because many years- well I'll just give you advice. So sometimes, there are giant companies with let's say, uh, 50,000 people, right? And I'm not thinking of any one specific company. | 3,098 | 3,139 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3098s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | If you're trying to guess what company I'm thinking of, there is no one specific company I'm thinking of but this pattern matches, uh, to many large companies. But maybe there's a giant company with, you know, 50,000 people, right? And, um, let's say that they have a 300 person, right, AI team, um, it turns out that if you look at the work of the 300 persons in | 3,139 | 3,167 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3139s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | the AI team and if they send you a job offer to join the 300 person AI team, that might be pretty good, right? Since this may be the group, you know, whose work you hear about, they publish papers, you read them on the news. Um and so if you've got a job offer to work with this group, that might be pretty good or even better would be sometimes even within the 30 person AI team it's actually difficult to tell what's good and what's not. | 3,167 | 3,189 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3167s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | There is often a lot of variance even with this, what's even better would be if you get a job offer to join a 30 person team. So you actually know who's your manager, who are your peers, who you're working with. And if you think these are 30 great people you can learn from, that could be a great job offer. The failure mode that unfortunately I've seen, um, several Stanford students go down and it's actually this is a true story. | 3,189 | 3,214 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3189s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | There was once, uh, several years ago there's a Stanford student I knew that I thought was a great guy, right? You know, I knew his work, he was coding machine learning algorithms. I thought he was very sharp and did very good work, uh, working with some of my Ph.D students. He got a job offer from one of these giant companies with- that has a great AI group. | 3,214 | 3,232 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3214s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | Um, and his offer wasn't to go to the AI group, his offer was to, um, join us and then we'll assign you to a team. So this particular student, that was a Stanford student that I know about and care about, um, he wound up being assigned to a really boring Java back end payments team and, uh, so after he accepted the job offer, he wound up being assigned to a, | 3,232 | 3,255 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3232s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | you know, back-end- and I apologize. I know you work on Java back-end payment process systems I think they're great [LAUGHTER] but the student was assigned to that team and he was really bored and so, um, I think that this was a student whose career- I personally saw his career rising, while he was at Stanford and after he went to this, you know, frankly not very interesting team, | 3,255 | 3,275 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3255s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | I saw his career plateau, um, and after about a year and a half he resigned from this company after wasting a year and a half of his life and missing out really on a year and a half of this very exciting growth of AI machine learning, right? So it was very unfortunate. Um, uh, and it was actually after I told this story, um, last time I taught this class earlier this year that actually someone from, um, | 3,275 | 3,298 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3275s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | actually it was from the same big company [LAUGHTER] he found me and said, "Boy, Andrew I wish you'd told me the story earlier, because this is exactly what happened to me, at the same big company [LAUGHTER]. Now, I wanna share with you, uh, a different, um, so- so I would just be careful about rotation programs as well. You know, when the company is trying to recruit you, | 3,298 | 3,324 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3298s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | if a company refuses to tell you what project you work on, who's your manager, exactly what team you're joining, I personally do not find those job offers that attractive because if they can't, you know, if they refuse to tell you what team you're gonna work with, well chances are, right, telling you the answer will not make the job attractive to you. That's why they're not telling you. | 3,324 | 3,345 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3324s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | So I'd just be very careful. And sometimes rotation programs sound good on paper but it is really, you know, well we'll figure out where to send you later. So, I feel like I've seen some students go into rotation programs that sound good on paper, that sound like a good idea but just as you wouldn't- after you graduate from Stanford, would you wanna do four internships and then apply for a job? | 3,345 | 3,364 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3345s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | That would be a weird thing to do. So, sometimes rotation programs are yeah, come and do four internships and then we'll let you apply for a job and see where we wanna send you. It could be a job at back end payment processing system, right? So, um, so so just just be cautious about the marketing of rotation programs. Um uh, and again, if you do if but if- but if what they say is do rotation and then you join this team, | 3,364 | 3,387 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3364s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | then you can look at this team and say yep, that's a great team. I wanna do rotation but then I would go and work with this team and and these are the 30 people I'll work with. So that could be great. But do a rotation and then we can send you anywhere in this giant company, that I would just be very careful about. Um, now on the flip side, there are some companies, | 3,387 | 3,405 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3387s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | I'm not gonna mention any companies, but there are some companies with you know, not as glamorous, not as- not as like cool brands, and maybe this is a, I don't know, 10,000 person company or 1,000 or 50,000 person or whatever. Let's say 10,000 person company. I have seen many companies that are not super well-known in the AI world, they are not in the news all the time, | 3,405 | 3,427 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3405s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | but they may have a very elite team of 100 people doing great work in machine learning, right? And there are definitely companies whose brands are not you know, the first companies you think of when you think of big AI companies that sometimes have a really really great 10 person or 50 person or 100 person team that works on learning algorithms. And even if the overall brand or the overall company, | 3,427 | 3,453 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3427s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | you know, isn't as like, is a little bit sucky. If you manage to track down this team and if you have a job offer to join this elite team in a much bigger company, you could actually learn a lot from these people and do important work. You know, one of the things about Silicon Valley is that uh, the brand of your resume matters less and less, right? Less than never before. | 3,453 | 3,474 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3453s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | I mean, I guess, I think the exception of the Stanford brand, you totally want the Stanford brand in your resume but with that exception, but really you know, Silicon Valley is becoming really good. Sili- the world, right? Has become really good at evaluating people for your genuine technical skills and your genuine capability and less for your brand and so, | 3,474 | 3,493 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3474s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | I would recommend that instead of trying to get the best stamps of approval on your resume to go and take the positions that let you have the best learning experiences and also allows you to do the most important work and that is really shaped by the you know, 30 or 50 people you work with and not by the overall brand of the company you work with, right? So the variance across um uh-so there's a huge variance across teams within | 3,493 | 3,518 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3493s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | one company and that variance is actually pretty bigger or might be bigger than the variance across different companies, does that make sense? Since I would, and if a company refuses to tell you what team you would join, I would seriously consider just, you know, doing something- if you have a better option, I would, I would do something else. Um, and then finally, um, | 3,518 | 3,539 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3518s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | yeah and- and so really I- again I guess I don't wanna name these companies but you know think of some of the large retailers or some of the large healthcare systems or there are a lot of companies that are not well known in the AI world but that I've met their AI teams and I think they're great. And so if you're able to find those jobs and meet their people you can actually get very exciting jobs in there. | 3,539 | 3,560 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3539s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | All right but of course, for the giant companies with elite AI teams, you can join that elite AI team, right? That's also- that's also great. I'm a bit biased since I use to lead some of these elite AI teams. So- so I think those teams are great but the loss of some teams in a, um, ah, yeah. All right. Um, lastly, you know, just general advice, this is how I really live my life. | 3,560 | 3,585 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3560s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | I tend to choose the things to work on that will allow you to learn the most and you know, try to do important work, right? So, you know especially if you're relatively early in your career, whatever you learn in your career will pay off for a long time and so um, uh and so joining the teams that are working with a great set of 10 or 30 or 50 teammates will let you learn a lot, and then also, | 3,585 | 3,625 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3585s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | you know, hopefully, I mean, yeah and- and just don't- don't don't join a like a cigarette company and hope you know, give more people cancer or stuff like that. Just don't- don't do this. Don't- don't do stuff like that. But if you can do meaningful work that helps other people and do important work and also learn a lot on the way, hopefully you can find positions like that, right? | 3,625 | 3,647 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3625s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | That let you set- set yourself up for long-term success but also do work that you think matters and that, and that helps other people. All right. Um, any questions while we wrap up? Yeah. [NOISE] I have a question about important work, what are some topics that you think you would include as important [inaudible]? What's important? You know, I don't know. Um, I think one of the most meaningful things to do in life is called [inaudible]. | 3,647 | 3,678 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3647s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | Either advance the human condition or help other people. But the thing is, I'm nervous. I don't wanna name one or two things because the world needs a lot of people who work on a lot of different things. So, the world's not gonna function if everyone works on computational biology. I think comp-bio is great but it's actually good that, where people work on comp-bio, | 3,678 | 3,696 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3678s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | my Ph.D students like you know, many work on the outside to healthcare. My team at Landing AI does a lot of work on the AI applied to manufacturing, to agriculture, to some health care and some other industries. Um,uh, I actually especially the California fire is burning you know, I actually think that there's important work to be done in AI and climate change, uh, um, | 3,696 | 3,717 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3696s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | but I think that there's a lot of important work in a lot of industries. Right, I actually think that, you know, I should think that the next wave of AI, excuse me I should say machine learning, is we've already um, transformed a lot of the tech world, right? So, you know, yeah, I mean we've already helped a lot of the Silicon Valley tech world become good at AI and that's great, right? | 3,717 | 3,743 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3717s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | Helped build a couple of the teams that wound up doing this, right? Google Brain, how Google become good at deep learning, the Baidu I grew up with, hope I do become, you know, good at one of the greatest AI companies in the world, certainly in China, and I'm very happy that between me and some of my friends in the industry we've made a lot of good AI companies. | 3,743 | 3,762 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3743s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | I think part of the next phase for the evolution of machine learning is for it to go into not just the tech companies like the, you know, like the Google and Baidu which I helped as well as Facebook, Microsoft which I had nothing to do with as well as what else AirBnB, Pinterest, Uber, right? All these are great companies. I hope they'll all embrace AI. But I think some of the most exciting work to be done still has also looked outside | 3,762 | 3,784 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3762s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | the tech industry and to look at all the sometimes called traditional industries that do not have shiny tech things because I think the value creation there as surprise you could implement there may be even bigger than if you, you know, uh, uh yeah. I'll mention one interesting thing, one thing I noticed is a lot of large tech companies all work on the same problems, right? | 3,784 | 3,808 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3784s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | So everyone works on machine translation, everyone works on speech recognition, face detection, and click-through rate and part of me feels like this is great because it means there's a lot of progress in machine translation and that's great. We do want progress in machine translation. Though sometimes when you look at other industries. Um, so, you know, when you look at manufacturing or um, some of | 3,808 | 3,829 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3808s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | the medical devices that you're looking at or sometimes on on these farms hanging out with farmers on, on, on. If you like, in my own work with my teams where sometimes we're stumbling across brand new research problems that the big tech companies do not see and have not yet learned to frame. So, I find one of the most exciting challenges is actually to be constantly on the cutting edge. | 3,829 | 3,851 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3829s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
733m6qBH-jI | Looking at these types of problems there's a different cutting edge than the cutting edge of the big tech companies. So, I think some of you will join the big tech companies and that's great. We need more AI in the big companies, in the tech companies, but I think a lot of exciting work to do in AI is also outside where we traditionally consider tech, right? | 3,851 | 3,869 | https://www.youtube.com/watch?v=733m6qBH-jI&t=3851s | Stanford CS230: Deep Learning | Autumn 2018 | Lecture 8 - Career Advice / Reading Research Papers | |
n1SXlK5rhR8 | hi there so you may have seen this already there's a cvpr paper called pulse and what it does is it's a method to up sample a pixelated image in a way that makes it look realistic but also that the again down sampled variant matches the original down sampled image so it's kind of a cycle consistency loss together with a again and all in all it's a method to demonstrate how you could do this now | 0 | 28 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=0s | [Drama] Yann LeCun against Twitter on Dataset Bias | |
n1SXlK5rhR8 | this has been trained on this face dataset among others there it was a user Bomb Z that made this into a collapse of people could try it out and tweet it this out and as you can see it works pretty nicely it gives pretty nice results on this particular data set but of course people started playing around with it and gave fairly funny results like this or that that gets more into | 28 | 57 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=28s | [Drama] Yann LeCun against Twitter on Dataset Bias | |
n1SXlK5rhR8 | the horrible category these so you can see these ones I particularly liked from being made into the little child so you can see as soon as you get away from the original kind of dataset modality you are going to get these results that are off and people started to notice that so here you input Barack Obama and what comes out is a fairly standard Caucasian person someone tweeted out saying this | 57 | 89 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=57s | [Drama] Yann LeCun against Twitter on Dataset Bias | |
n1SXlK5rhR8 | image speaks volumes about the dangers of bias in AI I guess here is where the entire story starts so young Luca weighs in and says ml systems are biased when data is biased this face up sampling system makes everyone look white because the network was pre trained on flick face HQ which mainly contains white people picks train the exact same system on a date set from Senegal and everyone | 89 | 117 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=89s | [Drama] Yann LeCun against Twitter on Dataset Bias | |
n1SXlK5rhR8 | will look African so this is pointing out why this happens namely because the data set is mainly Caucasian people so the results of up sampling are going to be mainly Caucasian people I mean this is like a straightforward explanation of why we're seeing what we're seeing but of course this was not okay and here is where the piling starts as an interjection we have to talk about bias | 117 | 141 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=117s | [Drama] Yann LeCun against Twitter on Dataset Bias | |
n1SXlK5rhR8 | in machine learning technically there is a statistical notion of bias which has a very rigorous definition and there is the societal definition of bias and these two things even though they're the same word they're totally different a machine learning system mainly consists of four different parts there is a data set the model the loss function and the optimization procedure statistical bias | 141 | 163 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=141s | [Drama] Yann LeCun against Twitter on Dataset Bias | |
n1SXlK5rhR8 | means whenever the model the loss or the optimization procedure leads to a situation where the outcome doesn't reflect the distribution of the data that you input this for example is achieved when you regularize your model which means that you put some prior knowledge onto the model you introduce bias and therefore you choose to not accurately represent your data | 163 | 186 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=163s | [Drama] Yann LeCun against Twitter on Dataset Bias | |
n1SXlK5rhR8 | distribution regularize it to a more bias distribution that in turn has lower variance we know this as the bias-variance tradeoff it's actually very simple right you you have the Ferraris and the Lamborghinis and you want make a model that predicts the accident probability now it just so happens that the Ferrari drivers are a bit more reckless and they do slightly | 186 | 207 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=186s | [Drama] Yann LeCun against Twitter on Dataset Bias | |
n1SXlK5rhR8 | higher accidents and now I train my logistic regression and it tells me okay 6040 cool but now I train my logistic regression with an l1 penalty and I say I want my model to be you know explainable so I wanted to be sparse I want the least amount of variables to be contributing to it what's the model gonna say the models gonna say Ferrari drivers add Lamborghini drivers good | 207 | 228 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=207s | [Drama] Yann LeCun against Twitter on Dataset Bias | |
n1SXlK5rhR8 | societal bias and machine learning is way different an example for this is when face detection systems work well on Caucasian people but don't work so well faced with people from other Heritage's and these societal biases are in the dataset as young account points out here if you change the dataset you'll change these biases notably the societal biases can only be in the data | 228 | 253 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=228s | [Drama] Yann LeCun against Twitter on Dataset Bias | |
n1SXlK5rhR8 | set otherwise you'd have to argue something like logistic regression itself has a preference for white people or something like this now there is a considerable interaction effect between the two but as Jung Lacan points out the actual societal bias of the final system is a direct result of the bias in the dataset and he is very correct if you train that system on | 253 | 275 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=253s | [Drama] Yann LeCun against Twitter on Dataset Bias | |
n1SXlK5rhR8 | a different data set it will exhibit different biases societal bias cannot be in the other parts of the machine learning pipeline they can serve to exaggerate or mitigate that bias in the data set but they themselves can only be statistically biased and not societally biased but ya'know can make the terrible mistake of pinpointing the exact root cause of this problem and not addressing | 275 | 299 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=275s | [Drama] Yann LeCun against Twitter on Dataset Bias | |
n1SXlK5rhR8 | the I guess wider ranging problems in the field as some people perceive it and he shouldn't have to write he pretty clearly says this is why it happens we can solve it by swapping the dataset he doesn't say anything about anything else namely he doesn't say that general bias in the field is not a problem he doesn't say that this doesn't harm anyone none of that he simply suggests a | 299 | 325 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=299s | [Drama] Yann LeCun against Twitter on Dataset Bias | |
n1SXlK5rhR8 | solution Jonathan Peck says well yes that's the point ml researchers need to be more careful selecting their data so that they don't encode biases like this and Lacan responds with not so much ml researchers but ml engineers the consequences of bias are considerably more dire in a deployed product than in an academic paper which is also correct this paper was about the method showing | 325 | 349 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=325s | [Drama] Yann LeCun against Twitter on Dataset Bias | |
n1SXlK5rhR8 | that this method works on this dataset now assume if here makes a interesting point which I agree with saying that today ml researchers are inadvertently powering product of a lot of non-ai companies who ignorant lis start with a pre trained birth or ResNet or Yola from the internet probably ignoring the license readme and so on which is a valid point right there are | 349 | 372 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=349s | [Drama] Yann LeCun against Twitter on Dataset Bias | |
n1SXlK5rhR8 | going to be people that take this and think oh this is a face up sampler cool I can use that without noting that this is simply an example and implementation on an example data set so you can argue that there might be some responsibilities of the researchers right here that doesn't make yung Lacan not correct but I still consider this to be like a fruitful discussion between | 372 | 394 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=372s | [Drama] Yann LeCun against Twitter on Dataset Bias | |
n1SXlK5rhR8 | individuals right here but now we go on this person saying train it on the whole American population with an l2 Lawson almost everyone will look white or train it on the whole American population with an l1 loss and more people might look black stop pretending that bias does not also come from algorithmic choices young the car never says it doesn't right the | 394 | 415 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=394s | [Drama] Yann LeCun against Twitter on Dataset Bias | |
n1SXlK5rhR8 | car responds now saying the most efficient way to do it though is to equalize the frequencies of the categories of samples during training this forces the network to pay attention to all the relevant features for all the sample categories and training with an l1 instead of an l2 will not even begin to solve the problem I would pretty much argue training with an l1 loss here | 415 | 435 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=415s | [Drama] Yann LeCun against Twitter on Dataset Bias | |
n1SXlK5rhR8 | would exacerbate the problem because the l2 loss is much more sensitive to outliers drawl Sutton says serious question why do you feel that it's important to make this point are you worried that people are going to start suing cycle gang and Lacan says because people should be aware of this problem and know its cause so they can fix it how terrible yawn how terrible you dare | 435 | 457 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=435s | [Drama] Yann LeCun against Twitter on Dataset Bias | |
n1SXlK5rhR8 | pinpoint the exact cause of the problem so that people can fix it the correct thing to do is to point out that everything is problematic so Tim the giver says Jung I suggest you watch me and Emily's tutorial or a number of scholars who are expert in this area you can't just reduce harms to dataset bias for once listen to us people from marginalized communities and what we | 457 | 481 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=457s | [Drama] Yann LeCun against Twitter on Dataset Bias | |
n1SXlK5rhR8 | tell you if not now during worldwide protests not sure when so again I feel the argument here is that you can't simply point out that it's the data set bias you must point out the bigger problems which the on account does not ever deny he simply says this particular problem can be solved by switching the data set Nikola LaRue says Jung was in my PhD jury I am indebted for him for | 481 | 506 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=481s | [Drama] Yann LeCun against Twitter on Dataset Bias | |
n1SXlK5rhR8 | everything he taught me but this constant dismissal of the harms caused directly or indirectly by the m/l community is highly problematic where or when have I dismissed the harm caused by the m/l community I'm pointing out the cause of the harm so it can be fixed you can't fix the harm unless you know what causes it know the roux says causes of the biases are numerous only pointing | 506 | 528 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=506s | [Drama] Yann LeCun against Twitter on Dataset Bias | |
n1SXlK5rhR8 | out data set bias deflects the attention away from the other more pervasive ones that make the whole field of bias in ml many people try to your attention about these issues but you kept focus on the data set because the dataset is the problem right here he doesn't dismiss any of the other things he simply says here the data set is the problem if your problem is that | 528 | 549 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=528s | [Drama] Yann LeCun against Twitter on Dataset Bias | |
n1SXlK5rhR8 | it doesn't work as well for non-caucasian people which was never the intent of this the intent of this was to showcase the method I mean imagenet is like 60% dog species and still people trained on it to showcase their image recognition techniques no one training on image net makes a claim that they have solved computer vision for all the classes in the world in a fair manner | 549 | 572 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=549s | [Drama] Yann LeCun against Twitter on Dataset Bias | |
n1SXlK5rhR8 | Tim McGee Bru goes on saying I'm sick of this framing tired of it many people have tried to explain many scholars listen to us you can't just reduce the harms caused by ML to dataset bias doesn't do that doesn't do it so someone asks her is he engaging in any ways with you it's appalling to see that he answers to everybody but you yet maybe there is a conversation going on in | 572 | 595 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=572s | [Drama] Yann LeCun against Twitter on Dataset Bias | |
n1SXlK5rhR8 | private and I don't want to jeopardize it note that young Lacoste tweet has 500 retweets 1.9 K likes and comments as far as you can scroll to what she responds to with yep but I'm used to white men refusing to engage with black and brown women even on issues of bias that mostly affect us I mean he literally has ignored a whole body of work by people from that demographic hence the | 595 | 623 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=595s | [Drama] Yann LeCun against Twitter on Dataset Bias | |
n1SXlK5rhR8 | statement so not surprised I mean in absence of the fact that an argument should be independent of the person making the argument that is a low blow heart Meru says I respectfully disagree with Yun here as long as progress is benchmarked unbiased data such biases will also be reflected in the inductive biases of ML systems advancing ml with biased benchmarks and asking engineers | 623 | 650 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=623s | [Drama] Yann LeCun against Twitter on Dataset Bias | |
n1SXlK5rhR8 | to simply retrain models with unbiased data is not helpful I don't disagree with you here I don't think my tweet contradicts your statement which it doesn't people are reading into this because he doesn't conform to the orthodoxy of pointing out that everything and everything is problematic and the pinpoints a particular problem he must be thinking all the wrong things | 650 | 673 | https://www.youtube.com/watch?v=n1SXlK5rhR8&t=650s | [Drama] Yann LeCun against Twitter on Dataset Bias |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.