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
|
|---|---|---|---|---|---|---|
_V-WpE8cmpc | plug this in into your stand the kind of action recognition data said a lot of those data sets have audio in them so we thought okay can we use this audio channel to improve our our method and we are actually definitely getting improvement over audio and we are improving other self supervised method we're still not as good as something that is kind of trained with lots and | 1,904 | 1,929 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1904s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | lots of relevant semantic data but again the hope is that as we get to harder and harder data sets that semantic labels will be harder and harder to get and so the self supervised methods hopefully will get better okay and finally kind of a cool fun thing we could do is to see if we can do off on screen separation of the audio sources so an example would be if you have a | 1,929 | 1,962 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1929s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | speaker speaking in on to the camera and then there's somebody else speaking that is not being seen our feature is only going to focus on the on the speaker that is being seen and so we can subtract away the speaker who is not being seen okay so let's let's see what I so unfortunately we thought oh this is such a cool idea nobody would ever think about it at the same time like four more | 1,962 | 1,992 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1962s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | groups who are basically doing the same thing luckily we heard about each other in time so we actually cross sighted in so it's all fine our method most of these folks actually that was the goal of their papers so they could have they basically work on that particular problem for us it's just basically one application of our feature that just kind of falls out of | 1,992 | 2,016 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1992s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | our method so we'd say that we're kind of a it's it's more of an application of our stuff but hopefully we are also able to to do other other things as well and the idea here is basically we take our representation and then we seed it into kind of a standard encoder/decoder representation that starts with a spectrogram and basically learns the separated into into things that part of | 2,016 | 2,048 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2016s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | the spectrogram that have evidence in the image or in the video and the part that does not have evidence in the video and then you can then you can play either one or the other right once you separate you can play either one and so here is an example then asking about it because they're not interesting facts to you that's not true I have a plenty of questions integration you a sample about | 2,048 | 2,070 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2048s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | your by talking about your flight now I know there's a bunch of people talking one is old screen one is off screen let's see what we can do so this is just the on screen been asking about it because they're not interesting facts to you and then people all about no I'm not I don't want to okay and this is the off screen that's not true I have a plenty of questions integration you have to | 2,070 | 2,096 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2070s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | disability Oscar by talking about your flight so there's there is a little bit of noise there but mostly does the right thing thank you so much thank okay and this is midnight is no second have day let's talk about digital omens there are some fears so much were able to show to the rest of the world the unshakable japan-us alliance okay we have both speakers here okay | 2,096 | 2,151 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2096s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | you're saying that they're not and we can we can hide one of the speakers and then gets away okay could even do something all right laughter okay all right all right okay most of it is gone but not all okay so so that is that is kind of a various ways of trying to get data to supervise itself and hopefully learn the representation that that is that is you know useful for other tasks the second | 2,151 | 2,213 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2151s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | topic I want to mention briefly is what we call meta supervision and the idea here is instead of telling what the correct answer should be we tell how that correct answer should behave so what do I mean by that so the kind of the direct supervision is a the direct supervision is is you have input X and you train a function f of X and you want that function f of X to produce Y's okay | 2,213 | 2,251 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2213s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | that's direct supervision you know from X's to eyes what are other ways we can we can set up this problem one way is we train a function of f of X that produces something in the domain Y in the set capital y so we don't tell it what particular y we want we just want it to be in the set of Y's okay and one example of this is generative adversarial networks I'll give a brief | 2,251 | 2,289 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2251s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | overview based on a paper we had a couple last year the colorization example that I showed it had a lot of kind of a ways of hacking it to make it look good and some what you want is you really want this white wall to be white on the back but because we kind of told it to all you need to be more colorful the wall becomes not white okay so it's kind of overshooting it and the annoying | 2,289 | 2,320 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2289s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | thing is that there is no way for the algorithm to look at this that's just not looking realistic you know do something better make it better have a try to optimize for things looking realistic so we don't know how to do that well actually we do so you know for any kind of a problem we do have like colic or is polarization or super-resolution or whatever it would be | 2,320 | 2,347 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2320s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | nice if we had this function that would tell us you know make something realistic enough it is this loss function this commune universal it says make make images look real we do have that function right now that's that's called a graduate student okay that's where the graduate student basically keeps hacking all algorithm until you know enough of the pictures look good | 2,347 | 2,369 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2347s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | and then then we send it to to publish but it would be better if the computer were doing it themselves okay so one way that we we can do this is we can basically have somehow the computer send the resulting images to Amazon Mechanical Turk ask a whole bunch of people if this is good looking or not good looking and use that signal to update the algorithm very very expensive | 2,369 | 2,397 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2369s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | okay but what what we can do is we can use this idea that that recently came out that kind of does something similar okay because remember we have a lot of real images so what we could do is we can have another network that can act as as an amazon turk are deciding if something looks good or not okay and that network is basically going to tell us if the image that we generated if it | 2,397 | 2,428 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2397s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | looks real that is does it can you distinguish that image from a set of real images and if their answer is no that means that we are doing well okay and this is the idea behind this giant material models in Goodfellow and colleagues that that has kind of really energized this whole field of image synthesis let's think about it you know what is it actually doing | 2,428 | 2,456 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2428s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | quickly so we have our function that translates say from greyscale to color okay and now what we want is we want to add another network on top of that and that network we want to decide how to decide if the image here if it looks real or it looks fake okay so what we want is we want G the network G to fool the network D so we want D to think that this is a real image whereas | 2,456 | 2,491 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2456s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | in fact it was generated by by by G okay and of course G who doesn't want to get fooled so do you really trust ooh to figure out if it can if it can if it can tell and the idea is to basically have these two networks battle it out in a kind of a duel like an arms race and the idea with as just any arms races when you have competition when you have a duel both | 2,491 | 2,518 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2491s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | get better and better and better that's that's that's the beautiful story Oh of gas right so here is a little bit of math so let's say that in this particular context if we're trying to see what G wants to do d wants to have a high probability for generated images if the image was generated by G we wanted to have a high probability of saying that it's it's a fake image okay whereas | 2,518 | 2,542 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2518s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | if the image is actually a real image then we want to have it say no this is a low probability that it was faked okay so we want to have a G such that it maximizes this quantity okay at the same time of course what we want G to do is to do the opposite it wants to minimize that same quantity right so d G is going to get back signal from D that says okay I figured you out | 2,542 | 2,578 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2542s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | and then G is gonna say okay let me see what I can do better now to improve my generator generated image so that G will have a hard time so maybe I'll something that looks a little bit better okay and try to fool it but of course I don't want to fool just this particular G I want to fool a the best possible D so this is where we get this whole minimax formulation where you want to | 2,578 | 2,607 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2578s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | have the best D and then minimize the the G to - to do the best of that okay so one way to think about it is that now this D you can think of it as a loss function it's kind of like l1 or l2 it basically tells what you need to do how do you optimize G such that it gets closer and closer to the goal the only difference is that instead of being nel - now this G is learnt the G learns what | 2,607 | 2,645 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2607s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | does that mean to get closer to the goal for that particular problem and what it means is that it basically means to be indistinguishable from the real samples from this from this data domain okay so we're almost done but not quite because here is an example imagine that my G went completely crazy and started producing cats for any input image it got it produced the cat okay now is this | 2,645 | 2,675 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2645s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | a real immature a fake image it's a real image it's a cat it's a it's my Student Union's cat named Aquarius very nice yet so it's a real image but it's not really what we met so we need to give it a little bit more constraints so basically what we want to do is we want to give D not just the generate image but also the input X so it can look at the pair of both of them together to say is it is | 2,675 | 2,706 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2675s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | the G of X the result of starting with X so that is is this a real pair or a fake pair and now we're all good now this is the conditional gann case and now we're we're able to to get it to work and this is the the final thing that is being optimized now I'm of course hiding a whole bunch of things under the rug here this is not a pretty optimization as you might imagine it is very complex | 2,706 | 2,739 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2706s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | complicated to optimize this thing and and and there is a lot and a lot of work on trying to make it simpler so far it's still it's it's it's it's more art than a science how do you optimize this thing but my graduate students are really amazing at doing this so so we were able to get this working and so now we can we can you know plug in greyscale and call her pair and then we can colorize images | 2,739 | 2,764 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2739s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | we can do the exactly the same code we plug in Google Streetview and satellites and then we can you know we can basically hallucinate satellites from from maps or we can do it the other way around exactly the same code but because the G is getting optimized for every different pair it basically learns what is important for every domain we can generate from labels we can generate | 2,764 | 2,790 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2764s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | facades we can go from day to night we can go to from thermal imaging to to normal RGB imaging we can take edges image edges and produce images that could have come from those things okay this kind of looks cool but actually it's not that complicated H maps actually contain a lot of the information the cool thing is that we can then train on this and test on just | 2,790 | 2,822 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2790s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | you know human sketches and even there it's actually doing something reasonable which is which is quite kind of neat and then we we put this online the code online and and a lot of kind of artists decided to do cool things with this and so this is this is kind of a neat thing where you you don't even need to do results of your papers anymore you just kind of post the code online and then | 2,822 | 2,852 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2822s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | just download results from from that other peoples have done and and somebody even did a little edges to cats thing you can try to yourself you you you you draw something you hit the the pigs button and then it will get you and get your cat okay there you go the the the best the best yeah the best use of company computer technology in my at least for me I don't | 2,852 | 2,883 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2852s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | know this is yeah this is the the pinnacle of my research career I think I get lots of cats on the Internet so so this is an example of Gann so as a again we we talked about direct supervision again basically cell we supervise of not on the particular label but on a set why there are other types of metal supervision we can think about so one is one of my favorite ones its | 2,883 | 2,919 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2883s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | cycle consistency okay the idea is that we don't know the answer why we don't have label for that but what we know is that if we if we have our f of X which produces some Y and then we apply a G of that of that we should get back to X okay and this is a constraint that people have used a lot especially in tracking in computer vision you track forwards you get somewhere you don't | 2,919 | 2,953 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2919s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | know where you are then you track backwards in time along the video and the idea is that you you should end up where you started with and if you don't then something is wrong okay but we can use this as a constraint to again for optimization so for example let's say that we want to do this kind of a pics to pix image to image translation but we don't have labeled pairs so let's say we | 2,953 | 2,976 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2953s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | want to translate from horses into zebras right there is no possible label data for this so how do we do this well we can take an inspiration from actually Mark Twain and ID of back-translation in in in in in in language and the idea of back translation is that if you want something translated in a foreign language you don't know you hire one translator to translate to that language | 2,976 | 3,004 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=2976s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | and you hire another one to decide back into language you do know and then you double check that it kind of still make sense right and Mark Twain wrote this book jumping frog in English then in French then clawed back into a civilized language once more by patient and renew merited toil so here he was showing that in this particular case the translation was not a good one so he translated back | 3,004 | 3,030 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3004s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | and and sure that was it was it was not not looking good okay and so what we're gonna do here is we're going to basically do the same idea we're going to now have a translator G that goes from domain acts of the main why okay and then translator F that goes back okay and and and and that's all now because we one do we don't want it to cheat and just stay where it is | 3,030 | 3,057 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3030s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | we want it to also have this adversarial loss this this disg and loss it says that when you get to domain why you better be indistinguishable from a real thing and why and when you get to domain X you better be indistinguishable from something real in X okay so what we're doing here is we're since starting with an image X we translate into a zebra domain again we don't have a label for | 3,057 | 3,085 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3057s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | that but we know that it should look like a real zebra and then we translate it back and then if we didn't if we don't get exactly where we started with well that's our loss that's what we want to minimize that is exactly this the thing that we are going to want to minimize okay and if you kind of a step back and squint at this thing what does it look like it's | 3,085 | 3,110 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3085s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | an hour old friend out encoder right you have the input you reconstruct that input the only difference is that it's an out in Golder that instead of a bottleneck it just has a different domain in the middle so it basically just has go through something else it's forced to go from some other representation and then come back and then you also do it the other way around | 3,110 | 3,132 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3110s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | okay and so then we can turn horses into zebras and vice versa we can even do it in in in videos just one frame at a time the failures are kind of fun I showed this picture in Moscow last year and I thought that's it I've they'll not let me out but they did we could also do kind of nice things on going from images to paintings okay by just kind of matching from domain of | 3,132 | 3,167 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3132s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | photographs the domain of in this case Suzanne paintings and I'm particularly happy about those clouds and you've probably seen a lot of the stylization papers and results they usually look at a single single image that you want to stylize for a particular image here we can take a whole domain we can take all of Suzanne's painting all the thousands of them and learn the representation | 3,167 | 3,193 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3167s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | that kind of models the whole Cezanne okay and and the clouds look pretty nice we can also go the other way around we can go from a painting into something that kind of hopefully hard to distinguish from a real image now if this was a perfect talk this would also be Suzanne and Suzanne didn't work Monet is simpler so so I will show you Monet but we're still working on Suzanne | 3,193 | 3,218 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3193s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | hopefully we can get Suzanne - okay we can also apply this to translating between video games in the real world so this is Grand Theft Auto and this is making it look like kitty so now you can see it's like old German looking and you can go the other way around which is even cooler they kind of walking around with a with with like make your reality like a video game so here is an example | 3,218 | 3,242 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3218s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | of reality as a video game I can let me show you just because this is just such a cool such a cool result this is this is again people have been just playing around with with these things all right I know it's not alright never mind it's forgot this is like artists just taking our code and running with it and okay I don't have the okay nevermind I will I will I'll give give you the | 3,242 | 3,309 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3242s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | pointer but yeah finally I want to show a very cute example of you know I in in my you know in my group we have been playing around with a lot of making fake imagery from from early on and so now a lot of people are worried about you know all this fake news and you know put in screwing things up so we thought we could try to play on the other side of the fence as well and see if we can | 3,309 | 3,344 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3309s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | detect if image is not being realistic okay and the top one is actually from my old paper with James Hayes where we learn to kind of fill in holes and create these image composites and here we can finally you know detect this and and and recover the that this image was what's fake and we are also going to do it with this the same idea of of self supervision and metal supervision this | 3,344 | 3,376 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3344s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | is with a couple of wonderful former Berkeley undergrads and and under Owens so given this image it might not might look reasonable to you but in fact of course it is fake and how do we how do we detect this well if we had enough fake examples we can just again do this you know supervise direct supervision they're out but we just don't have enough of these positive | 3,376 | 3,405 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3376s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | examples and so what we're going to do is we're trying to think about this as a normally detection so we're going to find to see if we can learn if an image is consistent with itself okay and the idea here is the following we can look at a couple of patches of this image and we see is there some sort of some kind of fingerprint in these in this patches that tells us that they might have come | 3,405 | 3,434 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3405s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | from different imaging systems that they are not from the same camera or and not from the same image okay now if we had access to the actual images that this was taken from created with then we could actually look at the the metadata that comes with the image and then we can realize that it's actually you know the cameras are different the focal lengths and different cetera et cetera | 3,434 | 3,458 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3434s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | but of course in real life we don't have access to any of that and so what we're going to do is we're going to train an algorithm to see if for a pair of images if those images if we can learn if a pair of patches comes from the same image or not now that by itself you could also do that that doesn't work as well because then again you don't have enough enough data for it for that so | 3,458 | 3,493 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3458s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | instead what we're gonna do is we're gonna train if a pair of patches have the same is it metadata tag there's a many different metadata exif metadata tags like camera brand focal length JPEG compression etc etc and for each one we can predict not the value of that tag but is it the same or is it different okay and so the idea is that we have a whole bunch of we train | 3,493 | 3,525 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3493s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | with a whole bunch of real images we don't need any fake images for this so for every pair of real images we take a couple of random patches and then we look at the exif tags that are similar in this images and we train those things to say okay yes those are similar and the different ones we say no those are those are different so basically we train something like 80 different | 3,525 | 3,550 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3525s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | classifiers that says for every different for every single active tag is this going to be the same or is it going to be different okay and so now we have a kind of a way to establish if a pair of images if pair of patches are consistent and more along one of the dimensions the dimension being you know are they do they come from the same camera do they have the same resolution | 3,550 | 3,575 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3550s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | do they have the same jpg etc etcetera okay so here is the different task tags and how well the how well we can predict if they if they come from the same image or not so you can say that the lens make is one of top-performing months so it is basically like the the what who produced the lens then you have a custom renderer is basically some Apple iPhone thing it | 3,575 | 3,608 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3575s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | basically says iPhone then then then you have a bunch of various things that really code for a different processing that's done by different cameras so they're all kind of things that different cameras do differently whereas things like image date and time or GPS coordinate are basically a chance level as you would expect okay and then what we do is we combine them all together oh | 3,608 | 3,632 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3608s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | and we also have some other consistencies that are like try to do it like blurring and and our Erie JPEG in etc etc and then a test time here is an image here is a manipulated image and what we do is we just have a pair go and find a whole bunch of different pairs of images and for every single active tag we could predict a map of if those two images are consistent or | 3,632 | 3,660 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3632s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | not consistent with each other these two patches okay and so we have a kind of consistency map for every single tag like camera or or focal length etc and then we combine them together into an overall consistency heat map and then once we have a heat map we can we can run normalized cuts and actually cut it into inter and thing or not and so here we can predict that all look at this | 3,660 | 3,687 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3660s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | this is this is a this is the inconsistent part and here is here is what it found and consistency is here here actually we didn't even notice it but it detected that the shadow on the on the floor was also painted in not just the guy on the top not sure nice nicely it works it you know for normal images it doesn't fire usually which is good and you know we're beating most of | 3,687 | 3,719 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3687s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | the pretty pretty much all all the other methods that are have been supervised we are beating them with this kind of self supervised method and for some images we don't have the ground truth so we don't know so I don't know who knows maybe this is how conspiracy theories are are born so I think I'm over time there so I I will I will I will skip the last part | 3,719 | 3,747 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3719s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | but I'm happy to talk about curiosity 101 so I think oops there you go thank you very much [Applause] any questions actually all know on the last part detecting detecting fake images so manipulated images so I would I wonder what would happen if you actually plug it into this game cycle where you're trying to actually make make wadding image look like it was taken with | 3,747 | 3,803 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3747s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | different objects with different camera and then having this discriminatory trying to discriminate it it which of the sides would win eventually if any so yeah so in I think in in you know longer so I think what they were saying is what happens if you have the the the critic the guy who is trying to find fakes connect with the generator that tries to make them better | 3,803 | 3,837 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3803s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | right and then have them battle with each other with this because there's an ongoing thing where people are trying to find fake images and then on the other hand improving that's right that's right that's right no I think this is this is actually that that's that's what we are talking you're thinking about doing doing doing next this kind of to have the the the fake detector not be a kind | 3,837 | 3,863 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3837s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | of a static one but to to to learn by having a generator generate better and better fakes and so hopefully then the detector becomes better and better and better technology to improve the detectors of face is exactly the same technology that you will use to improve the producing of the images right that is that is true but at least for now with these grand formulations in the end so that it this | 3,863 | 3,896 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3863s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | is kind of a weird thing that began really the Khmers converges when the detector cannot know the difference between the real and the fake right so that would be when the the thing actually converges in reality though the detector always wins so we we cannot fool the detector so those gaps never really converge so for now and it kind of may be reasonable because it's always | 3,896 | 3,924 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3896s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | easier to criticize them to create right always easy to be critic than it would then than a than a painter right so for now that might be okay but but in generally I think it's a it's it's it's an arms race it's an arms race and and and yeah there is no there is no gonna be a perfect solution there is always going to be something that the generator can do that that defeats the the | 3,924 | 3,953 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3924s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | defenses that's why I think it needs to be active that's why the the the this the the the fake detector needs to keep thinking of some other ways that the generator could potentially be fooling it and and be prepared for that if I may just smoke before so many of these fake images are created by copying you know patches of the same image somewhere else in the | 3,953 | 3,979 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3953s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | image is that something that this detector will be able to detect no no we actually we looked at this and most of those most of the fakes are not copied from from the same image at least the ones that we have looked at they're usually you know you go you find something on the internet you put you know picture of the Pope or Putin or whatever or Trump yes yes so in this thing we're basically | 3,979 | 4,013 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=3979s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | where we're focusing on what's called image splicing where you have two sources and then you create an image out of those two sources yeah so if you if you move things within the same image it might still possibly detect things we have seen a couple of examples where that happens of kind of the copy/move thing because it sometimes it screws up the the JPEG compression for example but | 4,013 | 4,039 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=4013s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | in general it's not it's not trained for that it's not it's that's not what it's looking for it's looking for really different imaging pipelines so it's not it's not supposed to be working for that example for that thing yeah mm-hm allow me I don't hear it that close doing videos I will I will I will repeat yes so the question is if we have tried it on on videos with people imitating | 4,039 | 4,100 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=4039s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | other people different voices we haven't but I think this is this is yeah we can or any of the the papers online the code is online so anyone can actually run it and see that would be an interesting thing to to run and see we have played with them very very ventriloquist's so when you have a puppet and you talk as a puppet and then you talk as as yourself and it kind of works it kind of works | 4,100 | 4,131 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=4100s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | for that so hopefully it should be it should be doing something like this yes it's not it's not meant to be kind of a detective detecting detecting fakes it's it's really mad it's meant to be fooled by the same things humans are fooled by so if the impersonator is a good one hopefully our method will work too yeah yeah question about this domain to domain mappings and you have to like one | 4,131 | 4,164 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=4131s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | directional and cycling example Google Maps in the wandering back to them the paintings were in the cytosine in the cytosine you care about the consistency that you come back to the center and you say the painting were actually working only in one direction the pay they the the paintings work in both directions paintings two two two photos they just don't look so good so | 4,164 | 4,199 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=4164s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | III didn't well I didn't show the results that didn't didn't work very well so remember in them in the Google Maps you go to the Google page and you can you can copy you can you can you can you can copy the the Google map and then you switch to the satellite and you can copy the satellite of that same thing so you have aligned inputs and outputs right so this is a much simpler problem | 4,199 | 4,241 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=4199s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | so there you don't need any consistency because you have the X in the Y's given to you and so then you can just do it directly with the paintings and them the photographs there is no alignment you have the suzanne paintings and you have photographs and you don't have a one-to-one correspondence and so for that the cycle is is is imported so that's the only thing we have basically | 4,241 | 4,268 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=4241s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | we there is no other constraint so then it's very portable and we come back to the polarization for example okay then the colored image is not strange today oh it is it is because there is a there is a hold on the the gun yeah so so let me see let me see zouri so in this in this setup you are given the input and the output okay and you are your basically where is the picture your | 4,268 | 4,326 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=4268s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | training on the grayscale and the color okay so you're given the input and the output and the Gann is just making sure that the output that the generator produces is similar in the perceptual space to what we humans expect okay because if you don't have the Gann then then you get weird results like this one right this one right this is what happens if you just kind of do | 4,326 | 4,365 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=4326s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | regression with the sake something like l2 loss yes yeah I I did not mention that I I'm sorry yes it does have the kind of this standard l2 loss plus the gamblers yes you're right I'm sorry I yeah I dropped out that that's like yes so for this you have the Delta loss plus the Gann loss in the second one with the cycle we don't have in the l2 loss because there | 4,365 | 4,395 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=4365s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | is no there is no data for that yes yes I'm sorry yes this good good point yes mm-hmm no it'sit's true I think I think it's it's it's not it's not clear what heat map what do you want to call her because again the the the producer of the sub for example when you're when you're playing the the organ the producer of the sound is the big the big pipe that's the thing produces the sound but of | 4,395 | 4,471 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=4395s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | course what what it's being it's it's coloring is the guy player pressing the keys and so I think it's it's not a very clear what is it that we want you know do we want the actual physical thing that produces the sound or do we want the the actuator and so here I think we are not we're not we just wanted to see what what would it be sighs so we are happy with anything that connects with | 4,471 | 4,502 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=4471s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | with sound production in some way or the other but yeah it you're completely right it's not I think if I think it's really it's it's going to be looking at correlations not causations and I think actually I would be happy if the dance party thing if it shows the people dancing I think that would be actually pretty cool but yeah you're completely right it's not a it's just kind of a | 4,502 | 4,527 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=4502s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | type of visualization that that shows okay this these are the pixels that connect [Music] so the question is I think this is a very very very good question that that there was a little bit of a slide of hand I said that you know the the tasks were too easy and the computers were easily cheating on the kind of the classification tasks and then I ended up making tasks harder but also changing | 4,527 | 4,600 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=4527s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | the tasks to something like you know pretty pictures or connecting audio and visual etc etc what about going back to you know detecting car detecting dogs and and and classification so I think I personally am NOT a big fan of the classification task to begin with I think it's a it's a it's a task that is designed to to be cheated on because it's a task that basically assumes that | 4,600 | 4,636 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=4600s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | you have a close do you have a close the world your world is a thousand classes and you're basically deciding one of the thousand things right so your your your your your chance performance is actually not that bad your chance performance on on something like image net is is one in one and two hundred right it's it's actually the chance is pretty high in the real world we're in the open world | 4,636 | 4,664 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=4636s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | where the potential number of that you need to recognize is almost infinite okay and so I think that actually a lot of the problems with with with the with with the come with these networks cheating is because we're testing them on tasks that are very constrained we're cheating is actually actually the right thing to do so we're testing them on something that is that is not that that | 4,664 | 4,702 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=4664s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | that that that is a kind of a very it's a it's a it's a it's a specialist task right what I think these methods will excel at is the generalist tasks something where you train on something and then you you you apply to something completely different and hopefully it will work better and we have already seen this like for example the colorization feature that we have | 4,702 | 4,726 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=4702s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | trained it does better than imagenet if the task is for example then to predict depth from a single image right so if the task is very different from the task it was trained on the cell supervised features work better if the task is similar to what it was trained on then the semantic tasks work better so I think that the the big goal is that we want to produce a generalist computer | 4,726 | 4,752 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=4726s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | something that is able to deal with novel situations in a reasonable matter not something that we already have if your goal is to just learn a specialist that will tell you no different types of of you know of Viennese pastries from each other you know you have a thousand different pastries and you want to tell name all of them then I think the current direct supervision methods are | 4,752 | 4,780 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=4752s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | exactly the thing you need to do give in this kind of closed world but if you want to have a general algorithm that can do the pastries it can do the flavors of gelato or it can you know tell cooker from from Malevich then the hope is that these kind of more general methods should get you a better of doing this so it's the difference between doing well on the exam versus | 4,780 | 4,808 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=4780s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
qG_YmIPFO68 | I want to talk to you about the future of medicine. But before I do that, I want to talk a little bit about the past. Now, throughout much of the recent history of medicine, we've thought about illness and treatment in terms of a profoundly simple model. In fact, the model is so simple that you could summarize it in six words: have disease, take pill, kill something. | 0 | 43 | https://www.youtube.com/watch?v=qG_YmIPFO68&t=0s | Soon We'll Cure Diseases With a Cell, Not a Pill | Siddhartha Mukherjee | TED Talks | |
qG_YmIPFO68 | Now, the reason for the dominance of this model is of course the antibiotic revolution. Many of you might not know this, but we happen to be celebrating the hundredth year of the introduction of antibiotics into the United States. But what you do know is that that introduction was nothing short of transformative. Here you had a chemical, either from the natural world | 43 | 68 | https://www.youtube.com/watch?v=qG_YmIPFO68&t=43s | Soon We'll Cure Diseases With a Cell, Not a Pill | Siddhartha Mukherjee | TED Talks | |
qG_YmIPFO68 | or artificially synthesized in the laboratory, and it would course through your body, it would find its target, lock into its target -- a microbe or some part of a microbe -- and then turn off a lock and a key with exquisite deftness, exquisite specificity. And you would end up taking a previously fatal, lethal disease -- a pneumonia, syphilis, tuberculosis -- | 68 | 97 | https://www.youtube.com/watch?v=qG_YmIPFO68&t=68s | Soon We'll Cure Diseases With a Cell, Not a Pill | Siddhartha Mukherjee | TED Talks | |
qG_YmIPFO68 | and transforming that into a curable, or treatable illness. You have a pneumonia, you take penicillin, you kill the microbe and you cure the disease. So seductive was this idea, so potent the metaphor of lock and key and killing something, that it really swept through biology. It was a transformation like no other. And we've really spent the last 100 years trying to replicate that model over and over again | 97 | 130 | https://www.youtube.com/watch?v=qG_YmIPFO68&t=97s | Soon We'll Cure Diseases With a Cell, Not a Pill | Siddhartha Mukherjee | TED Talks | |
qG_YmIPFO68 | in noninfectious diseases, in chronic diseases like diabetes and hypertension and heart disease. And it's worked, but it's only worked partly. Let me show you. You know, if you take the entire universe of all chemical reactions in the human body, every chemical reaction that your body is capable of, most people think that that number is on the order of a million. | 130 | 155 | https://www.youtube.com/watch?v=qG_YmIPFO68&t=130s | Soon We'll Cure Diseases With a Cell, Not a Pill | Siddhartha Mukherjee | TED Talks | |
qG_YmIPFO68 | Let's call it a million. And now you ask the question, what number or fraction of reactions can actually be targeted by the entire pharmacopoeia, all of medicinal chemistry? That number is 250. The rest is chemical darkness. In other words, 0.025 percent of all chemical reactions in your body are actually targetable by this lock and key mechanism. You know, if you think about human physiology | 155 | 188 | https://www.youtube.com/watch?v=qG_YmIPFO68&t=155s | Soon We'll Cure Diseases With a Cell, Not a Pill | Siddhartha Mukherjee | TED Talks | |
qG_YmIPFO68 | as a vast global telephone network with interacting nodes and interacting pieces, then all of our medicinal chemistry is operating on one tiny corner at the edge, the outer edge, of that network. It's like all of our pharmaceutical chemistry is a pole operator in Wichita, Kansas who is tinkering with about 10 or 15 telephone lines. So what do we do about this idea? | 188 | 220 | https://www.youtube.com/watch?v=qG_YmIPFO68&t=188s | Soon We'll Cure Diseases With a Cell, Not a Pill | Siddhartha Mukherjee | TED Talks | |
qG_YmIPFO68 | What if we reorganized this approach? In fact, it turns out that the natural world gives us a sense of how one might think about illness in a radically different way, rather than disease, medicine, target. In fact, the natural world is organized hierarchically upwards, not downwards, but upwards, and we begin with a self-regulating, semi-autonomous unit called a cell. | 220 | 251 | https://www.youtube.com/watch?v=qG_YmIPFO68&t=220s | Soon We'll Cure Diseases With a Cell, Not a Pill | Siddhartha Mukherjee | TED Talks | |
qG_YmIPFO68 | These self-regulating, semi-autonomous units give rise to self-regulating, semi-autonomous units called organs, and these organs coalesce to form things called humans, and these organisms ultimately live in environments, which are partly self-regulating and partly semi-autonomous. What's nice about this scheme, this hierarchical scheme building upwards rather than downwards, | 251 | 278 | https://www.youtube.com/watch?v=qG_YmIPFO68&t=251s | Soon We'll Cure Diseases With a Cell, Not a Pill | Siddhartha Mukherjee | TED Talks | |
qG_YmIPFO68 | is that it allows us to think about illness as well in a somewhat different way. Take a disease like cancer. Since the 1950s, we've tried rather desperately to apply this lock and key model to cancer. We've tried to kill cells using a variety of chemotherapies or targeted therapies, and as most of us know, that's worked. It's worked for diseases like leukemia. | 278 | 306 | https://www.youtube.com/watch?v=qG_YmIPFO68&t=278s | Soon We'll Cure Diseases With a Cell, Not a Pill | Siddhartha Mukherjee | TED Talks | |
qG_YmIPFO68 | It's worked for some forms of breast cancer, but eventually you run to the ceiling of that approach. And it's only in the last 10 years or so that we've begun to think about using the immune system, remembering that in fact the cancer cell doesn't grow in a vacuum. It actually grows in a human organism. And could you use the organismal capacity, the fact that human beings have an immune system, to attack cancer? | 306 | 329 | https://www.youtube.com/watch?v=qG_YmIPFO68&t=306s | Soon We'll Cure Diseases With a Cell, Not a Pill | Siddhartha Mukherjee | TED Talks |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.