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
|
|---|---|---|---|---|---|---|
YBlNQK0Ao6g | realized that I've forgotten to even read the name of the paper so it's called generative pre-training from pixels by March and Alec Radford round child jeff whoo he won't jus profile a dairy wall David Lewin and Ilya sutskever and since Henry AI labs has already made a video on this this video is going to be more of kind of a rumble rant about what I find interesting about | 185 | 211 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=185s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | the paper and some thoughts about it rather than like a classic explanation I hope you still enjoy that so what you saw on the right wasn't even though this isn't the final result the supposed result this is simply the pre-training task it's fun to look at it but the actual object objective of the paper is the following what if we train we pre train on a large | 211 | 237 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=211s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | data set to generate work good images like these or we to complete images like these and then we fine-tune on a classification task and the answer is here they say on C 410 we achieve 60 96.3% accuracy with a linear probe outperforming a super wide supervised the wide ResNet and the 99 cent accuracy with full fine tuning matching the top supervisor pre-trained | 237 | 272 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=237s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | models an even larger model trained on a mixture of imagenet and web images is competitive with self supervised benchmarks on image net achieving 72 top one accuracy on a linear probe of our features so the goal here is that you have a data set that you want to trend like train a classifier on so usually you have a data set and the data set has images and you put them through like a | 272 | 305 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=272s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | convolutional neural network and then you have to classify the image into one of I don't know how many classes on C for ten that's ten classes on image and it's a thousand and the data set is these images together with these labels now the idea of pre training is that you some where have a bigger data set that is sort of similar to the small data set but it's similar enough such that the | 305 | 332 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=305s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | network could learn something so what you want to do first is your first one it take the large data set terrain this network right here and then in a second step fine-tune the network on this smaller data set and you sort of hope that what you learned from the large data set right here transfers over a little bit of knowledge you already have a little bit of knowledge and you can | 332 | 354 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=332s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | make better use of the data that you have right here now the question is how do you do this pre training and of course this has a long tradition well long for maybe two or three years right now in the language community where people they pre trained these large models like we've just seen GP t3 or Bert was one of them they pre trained these large Transformer models on text | 354 | 382 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=354s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | and then to fine-tune them on classification tasks for text and that's what this paper is doing right here they pre trained a transformer that is a GPT to scale model they pre train it on image generation and then they fine-tune it or transfer learn it to classification tasks and the point of the papers to say that like in text data in text data we have made | 382 | 413 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=382s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | pretty good pretty good experiences with doing this with pre-training a generative model and then fine-tuning on a classification task while so far in images all we've ever done is we've pre-trained this pre training task he usually is a classification task or like a self supervised task with a contrastive loss or something like this what they're doing new is the generative | 413 | 443 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=413s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | modeling in the pre as a pre training and again this isn't like entirely new but they show that if you throw a lot of computers at it and lots of data and a big model then that can work equally well to these self supervised tasks so their model as I said is pretty pretty simple they take an image and they unrolled the image now an fully unrolled image on let's say image net has 224 | 443 | 474 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=443s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | squared pixels and that times three right because you have three color channels that's too large even for an open a supercomputer so what they do is first they down scale the image so they down scale it's not as drastic as here where you just get a three by three image but they do down scale it to like a 32 by 32 or a 64 by 64 then they unroll it which simply means they go | 474 | 502 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=474s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | through the image like this and make a sequence out of it because their models are naturally made for text sequences they simply put the image into a text sequence they further simplify this by reducing the three colour channels to a single one so they have their own color representation and basically yeah they reduce the three colour channels to one channel that simply indexes the color in | 502 | 532 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=502s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | their color representation and they say still pretty good it's pretty faithful so ultimately they end up with like 32 squared length representation of their image and then they do one of two things they either do auto regressive generative pre-training which is the sort of GPT - style pre-training and the the idea here is that you always want to predict the next pixel of a sequence so | 532 | 565 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=532s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | you can see right here that's the sequence that you are sorry that's the sequence that you input and you always want to predict what is the next pixel and in this case you've seen you see that we've already predicted everything here we've already predicted everything up to this red pixel so you want to know what's this next pixel this thing right here what's this going to be and the | 565 | 594 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=565s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | diagram here basically shows you how the attention flow so every position in this transformer and if you don't know what a transformer is I haven't made a video about attention is all you need where these are explained but briefly every position here can sort of send information can send information only in one direction as to so you train all of these in parallel and when you predict | 594 | 621 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=594s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | this pixel right here you only want information from whatever it was before that pixel otherwise the model could cheat right otherwise the model could simply learn to copy over the value but the attention pattern here is simply to show you that this is auto regressive and it's in one direction so you always want to predict the next pixel and then from all of this you want | 621 | 645 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=621s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | to predict the next pixel and from all of this you want to predict the next pixel this is in contrast to this objective here that comes from Bert and I've also made a video on Bert what you do in Bert is you simply take that image and you cross a block out two of the pixels or many of the pixels and you simply ask your network to reconstruct those pixels okay and now you can see the | 645 | 670 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=645s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | attention flows in all direction birth the B stands actually for a bi-directional so this is the contrast to the autoregressive pre training framework now the these two things have been applied in text both the autoregressive is usually it's easier to actually make it produce something like we saw producing these images because you can always just predict the next | 670 | 696 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=670s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | pixel and then the next and then the next and then the next whereas in Beart it's a bit more unclear how you would produce things in a consistent manner because the predictions of these two pixels right here they are independent it's one forward pass and then both of these are predicted but other papers have tried to solve this like this not excel net I forget I forget its name it's something | 696 | 724 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=696s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | with an X and yeah but but these are the two objectives they look at and it turns out they sort of trade off a bit they work equally well or a bit better and a bit worse depending on the task so once they have done this so they simply feed images and you will notice that you don't need any labels for this so what you'll do is simply input an image and then simply take away half of it like | 724 | 754 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=724s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | this and then predict that pixel and then you want to predict that pixel and then you want to predict that pixel right that's all like you do with text and invert you simply input an image cross out pixels and then predict them so you don't need labels for this and that's why you can do it with this big data set and you can do it in an unsupervised fashion so you can just | 754 | 778 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=754s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | crawl the internet for four images and just feed this in into there and it will sort of learn to produce these images now the question is if you produce if you learn to produce these images does that help you for classification and there they have two methods of of assessing this the bottom one here is the fine-tuning method where you simply so this is supposed to be the | 778 | 806 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=778s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | representation you learn in the different layers of the network so this is supposed to be this thing right here what you'll do is you'll simply fine-tune that means you on top of this representation you add a classification head that has two outputs cat or dog and you train this entire network on your small data set that we discussed before so you train the entire network all of | 806 | 833 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=806s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | the parameters this is called fine tuning in contrast to that what you can do is you can simply and this is the easy way you can simply add this classification head with two outputs and then only train this classification head and that is won't perform as well but it gives you sort of a better idea of how good is the representation that this network right here learned and on top of | 833 | 860 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=833s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | that so if you spin this idea further you can actually go and do this at any intermediate layer right here so you can forward propagate until layer two right here and then here you add your classification head into the two into the two classes and you only train the classification head that being said you can also do this with fine tuning but in this case this is called a linear probe | 860 | 888 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=860s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | and it is often used to assess how good the a representation in intermediate layers is whereas what it actually does is assessing how linearly classifiable a representation is which isn't the same as how useful or how informative but it is one way to to assess these things okay so these are the two things they assess alright so for as for data sets for C 410 they use like C for 10 and c | 888 | 919 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=888s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | for 100 as data sets and the STL 10 and there you have to keep in mind the pre training is done on imagenet for those so that you pre train on imagenet without the labels and then you transfer learn or fine tune or or linear probe on these small data sets whereas later we're going to look at image net and they're the pre-training as I understand it is done on image net | 919 | 948 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=919s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | itself but also a wider collection of a hundred million or so images from the web from the internet okay so as you can see right here this is what happens if you do this linear probing and you can see it works pretty well so you get like a ninety-five ninety-six percent accuracy with linear probes this is very powerful so it's not easy to get 96 percent on C for ten I mean current | 948 | 982 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=948s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | state of the art is like ninety nine percent but still 96 percent is pretty good and this is the so the entire network there is this big giant network that you input your image into and then there is this one linear layer that does the classification and all of this right here has not been trained with classification in mind it simply has been trained to reproduce images it | 982 | 1,012 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=982s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | hasn't even been trained on C for ten as far as I understand has been trained on image net so the the this is to stress how cool or how significant this result is basically that just a linear probe on top of that will give you such a good accuracy and the second thing that is obvious right here is this bottom axis is the layer so this is the layer where they attach the linear probe and usually | 1,012 | 1,045 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1012s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | if you pre train a network with a classification task in mind so you pre train it with the labels or maybe even without the labels in a self supervised the way or something like this usually the last layer has the best representation for classification but here the special thing is that the intermediate layers in the middle have the best representation you can see that | 1,045 | 1,069 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1045s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | representation quality in terms of linear probing falls off as they sort of it falls off as they go into higher layers and this is consistent across the datasets as you can see and the the idea here is or the way they interpret it is that if you have an image right here Dada Dada Dada and they you've blocked part of it so you've blocked this and this wrong way around this | 1,069 | 1,106 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1069s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | so you've generated everything and now your task is to predict the next pixel right so you're you train to predict this next pixel right here and the idea is that as you put the image through the network what it will do is sort of since the first layers they're going to be if you're going to be similar to a CNN they're going to be doing some low level of feature transformation thing right | 1,106 | 1,142 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1106s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | but also the last layers they're going to really care about what's the exact pixel that goes here right since it's their job to to do that they're going to care what color does it need to have you know what exact luminosity and so on how does it fit in with the previous pixels and so on where as so that's that's also good but it's not just low level information and consistency with other | 1,142 | 1,170 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1142s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | pixels or something like this at some point if you want to generate consistent images and we saw that this model can generate consistent images at some point there needs to be some kind of a notion of the global information in the picture right such because the images are consistent throughout so there needs to be some some notion of what is in that image as a whole and that's the exact | 1,170 | 1,198 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1170s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | information that we need for classification and the only way that could actually be is here in the middle since you know that's the place so the hypothesis is that the these models somehow learn a higher-level representation of global information somewhere in the middle before they then specify that information again down to predict the actual pixel and that's why | 1,198 | 1,224 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1198s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | the best representations for classification are in the middle so this is one of the this is actually the interesting finding or one of the interesting findings of this paper means cool that they can reach a good accuracy but to recognize that maybe in these these generative models they have some intermediate stage where they represent the global information and that will | 1,224 | 1,249 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1224s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | actually make the best representation okay the second cool thing right here is that you can see they have different sizes of models so the IGP TL I believe is something like sixty layers then this is like 48 layers and this is 32 layers we don't really so these are on the olive on the scale of GPT to either a little bigger or a little smaller it's not like a GPT three scale where you | 1,249 | 1,280 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1249s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | need a ginormous supercomputer though they do do a lot of computation but it this still sort of fits within hardware of a standard size and not like exascale what's interesting right here is that you can see the larger models they reach a lower validation loss so here is the validation loss larger model if you train them on so these checkpoints here are always after the same amount of | 1,280 | 1,310 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1280s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | steps the larger models do reach a lower validation loss right here as you can see so this is the large this is the medium this is the small and also you can see that on this axis is the linear probe accuracy so this is whenever you you go and you find the best intermediate layer for linear probing you probe it and you record the accuracy so you can see a general trend | 1,310 | 1,337 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1310s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | as your validation loss goes down the linear probe accuracy goes up so there is a connection like it is in text models and text models there's a connection of the perplexity of your language model and the quality that of the representation you get for downstream tasks in this model it seems to be the exact same thing there is a connection between reaching lower | 1,337 | 1,361 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1337s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | validation laws and reaching a higher performance on classification so that's one interesting thing the general trend - up to the upper right corner the other interesting and even arguably even more interesting thing is that if you look at the same validation loss so at this point all of these models have the same validation all's yet still the bigger model is better right you can see right | 1,361 | 1,391 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1361s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | here the bigger model outperforms the smaller model even though they have the same validation loss on the image modeling task and this is also something that openly I in their in their text papers has stressed that the larger models they seem to be somehow more capable of forming good representations even you know even if they have the same loss so again this this could just be | 1,391 | 1,421 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1391s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | sort of a training data remember training data remembering thing and when I set that in GPT three I didn't actually mean explicit remembering of training data I meant a kind of a fuzzy remembering of training data of I formulate that in the in the comments but I I feel a lot of people have misunderstood me there here I think it's a much harder harder to estimate what's going on also since | 1,421 | 1,449 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1421s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | image pixels humans don't have a super good model of an on image pixels in their head as we have about text as you can see if you then fine tune so for now we've just do linear probing if you fine-tune these architectures then you reach like a 99% accuracy on C for ten which is on par with the best models that we have so G pipe is supervised pre trained on imagenet but also I guess | 1,449 | 1,481 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1449s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | uses a bunch of data element ation while these image GPT it uses minimal data augmentation I think they simply random crop a little bit and that's about it so they also experiment around with this Bert objective so until now this was all the this was all this autoregressive objective and I feel the open hi people are a bit more of a fan of the autoregressive objective just given what | 1,481 | 1,515 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1481s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | they've done so far in their papers and you can see here comparison of the two objectives on C for 10 and on image net again C for 10 is pre trained with image net and image net itself is pre trained with like a larger collection of images from the web all the pre training is done without labels now the blue is what you can reach with a linear probe and the orange is then on top of that what | 1,515 | 1,548 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1515s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | you can reach by fine-tuning okay so no linear profile tuning oh I have to say that the fine tuning is always done at the end so even though the linear probe even though the linear probe can be attached anywhere in between and it's often useful to do that as we saw because the in between layers are the best they say they tried fine-tuning also in from in between but it always worked out | 1,548 | 1,575 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1548s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | best whenever you fine-tune whenever you fine-tune you take actually the last layer so that kind of gives you an idea that the model is then it's sort of what seems to be important is this coming up with the higher level representation and then once you fine tune you're probably able to push that representation through to the end because of your training signal | 1,575 | 1,603 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1575s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | but if you hadn't done the pre-training you wouldn't even have that higher level representation and then the signal I guess is not strong enough to back propagate through the whole model it would be very interesting if they investigate if they do this linear probe analysis again after they fine-tune the model that and to see if then still it is the intermediate layers that have the | 1,603 | 1,629 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1603s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | best representation or if now the best representation in a linear probe sense shifted towards the end I'm gonna guess it's shifted towards the end but I sort of want to even see if the accuracy of the linear probe in the middle does it keep the same right so does the curve go like this this is the linear probe when you simply pre-trained right this is linear probe accuracy the | 1,629 | 1,658 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1629s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | question would be does it change to be like this or does it change to be like this this is supposed to be the same at the end so basically does it stay as good as it is but simply get better at the end or does the representation like in this curve does the good representation now shift towards the end and leave the lower layer with even more capacity to do some low level stuff yeah | 1,658 | 1,688 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1658s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | maybe they've done this I haven't seen it so and as you can see these Bert and autoregressive object if they sort of trade off so the birthda tends to do poorly in the linear probe setting but then it catches up during fine tuning in C for 10 almost being at the level of the autoregressive and in an image net actually outperforming it this this darker thing here it simply means that | 1,688 | 1,718 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1688s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | you a ver across different maskings of Bert because I guess even in classification it's not entirely clear how to get a signal out of Bert because they don't do this CLS vector with Bert what they do for classification and linear probing and it's written up here they simply take the they do an average pooling I think they do an average pooling of the of all the representations of the sequence and | 1,718 | 1,751 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1718s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | the last thing that I've also forgotten there's a lot of stuff when they fine tune while fine-tuning while fine-tuning the classification loss yields reasonable downstream performance we find empirically that the joint objective the generative objective and the classification objective works even better okay so even when you fine-tune with this model you have to keep the | 1,751 | 1,784 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1751s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | generative modeling part the generative loss around and then it performs even more better more well whatever that word is so that's also something to think about I think this this paper right here it kind of lays down a lot of cool things that you can think about and it gives rise to a lot of hypotheses of how does this stuff work why does this stuff work I don't even think that the the | 1,784 | 1,816 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1784s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | numbers are the most important thing it's mostly the fact of the effects and what does it mean okay so this was my take on it it was it's more kind of a my rant of what I find special about this paper then about the actual paper you can look at the paper their numbers are pretty good on imagenet they do not reach the same like SuperDuper performance as they do on c 410 and i | 1,816 | 1,847 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1816s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | guess that's probably because they have to downscale the image net images way more than they have to downscale to see 410 images because those are of course only 32 by 32 so because they have to downscale so much they lose probably a lot of information and I would be interested to see if there is a way to involve convolution in this in all of this so to do the | 1,847 | 1,874 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1847s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
YBlNQK0Ao6g | downscaling that in a learned manner with convolutions or something I'm sure this has all been done already I'm just lazy to look it up yeah so I invite you to look at their blog post where they have these his samples they look pretty pretty funny and these full samples up here look fairly you know fairly cool for what it's trained to do and that it has no spatial awareness whatsoever it | 1,874 | 1,900 | https://www.youtube.com/watch?v=YBlNQK0Ao6g&t=1874s | Image GPT: Generative Pretraining from Pixels (Paper Explained) | |
xp0O2vi8DX4 | Transcriber: Leonardo Silva Reviewer: Denise RQ So, we all have some behavior that we would like to change about ourselves. And we certainly all want to help someone else change their behavior in a positive way. So, maybe it's your kid, your spouse, your colleague. So I want to share some new research with you that I think reveals something really important about what gets people to change their behavior. | 0 | 38 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=0s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | But before I do that, let's zoom in on one strategy that I think you probably use a lot. So, let's say you're trying to stop yourself from snacking. What do you tell yourself? Well, most people, in a monologue, will say, "Beware. You'll be fat." And if this was your kid, you would probably tell him that smoking kills and, by the way, he's in big, big trouble. | 38 | 65 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=38s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | (Laughter) So, what we're trying to do here is we're trying to scare ourselves and others into changing their behavior. And it's not just us. Warnings and threats are really common in health campaigns, in policy. It's because we all share this deep-rooted belief that if you threaten people, if fear is induced, it will get them to act. And it seems like a really reasonable assumption, | 65 | 94 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=65s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | except for the fact that the science shows that warnings have very limited impact on behavior. So, graphic images on cigarette packets, for example, do not deter smokers from smoking, and one study found that, after looking at those images, quitting actually became a lower priority for smokers. So, I'm not saying that warnings and threats never work, but what I'm saying is, on average, they seem to have a very limited impact. | 94 | 120 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=94s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | And so, the question is: why? Why are we resistant to warnings? Well, if you think about animals, when you induce fear in an animal, the most common response you will see is freezing or fleeing; fighting, not as much. And so, humans are the same. So if something scares us, we tend to shut down and we try to eliminate the negative feelings. So, we might use rationalizations. | 120 | 146 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=120s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | For example, you might tell yourself: "My grandpa smoked. He lived to be 90. So, I have really good genes and absolutely nothing to worry about." And this process can actually make you feel more resilient than you did before, which is why warnings sometimes have this boomerang effect. In other times, we simply put our head in the ground. (Laughter) Take the stock market for example. | 146 | 171 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=146s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | Do you know when people pull their head out of the ground to look at their accounts -- not to make a transaction, just to log in to check their account? So, what you're seeing here, in black, is the S&P 500 over two years, and in gray, is the number of times that people logged in to their account just to check. And this is data from Karlsson, Loewenstein & Seppi, | 171 | 192 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=171s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | it's control [data] for all the obvious confounds. So, what do we see? When the market is high, people log in all the time, because positive information makes you feel good, so you seek it out. And when the market is low, people avoid logging in, because negative information makes us feel bad, so we try to avoid it altogether. And all this is true as long as bad information can reasonably be avoided. | 192 | 221 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=192s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | So, what you don't see here is what happened a few months later, in the financial collapse of 2008, when the market went drastically down and that was when people started logging in frantically, but it was a bit too late. So, you can think about it like this -- it's not just finance: In many different parts of our life, (Laughter) we have warning signs and bad behaviors now. | 221 | 246 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=221s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | And they could potentially lead to all these bad outcomes later, but not necessarily so, because there are different routs from your present to your future, right? It can go this way, it can go that way. And, as time passes, you gather more and more information about where the wind is blowing. (Laughter) And, at any point, you can intervene and you could potentially change the outcome, | 246 | 273 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=246s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | but that takes energy and you might tell yourself: "What's the point about worrying about something that might happen? It might not happen." Until we reach this point, at which time you do jump into action, but sometimes it's a little bit too late. So, we wanted to know, in my lab, what type of information does leak into people. So, we conducted an experiment | 273 | 295 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=273s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | where we asked approximately 100 people to estimate the likelihood of 80 different negative events that might happen to them in the future. So, for example, I might ask you: "What is the likelihood that you'll suffer hearing loss in your future?" And let's say you think it's about 50%. Then, I give you the opinion of two different experts. So, expert A tells you: | 295 | 320 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=295s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | "You know, for someone like you, I think it's only 40%." So, they give you a rosier view of your future. Expert B says: "You know, for someone like you, I actually think it's about 60%. It's worse." So, they give you a bleaker view of your future. What should you do? Well, you shouldn't change your beliefs, right? Wrong. What we find is that people tend to change their beliefs | 320 | 349 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=320s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | towards a more desirable opinion. In other words, people listen to the positive information. Now, this study was conducted on college students, so you might say: "Well, college students are delusional, right? We all know that." (Laughter) And surely, as we grow older, we grow wiser. So we said: "OK, let's test that. Does this really generalize? Does it generalize to your kid, to your parent? | 349 | 375 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=349s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | Does it generalize to your spouse?" And so, we tested people from the age of 10 until the age of 80, and the answer was yes. In all these age groups, people take in information they want to hear -- like someone telling you you're more attractive than you thought -- than information that they don't want to hear. And the ability to learn from good news remained quite stable throughout the life span, | 375 | 401 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=375s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | but the ability to learn from bad news, that changes as you age. So, what we found was that kids and teenagers were the worse at learning from bad news, and the ability became better and better as people aged. But then, around the age of 40, around midlife, it started deteriorating again. So, what this means is that the most vulnerable populations, kids and teenagers on the one hand, and the elderly on the other hand, | 401 | 431 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=401s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | they're the least likely to accurately learn from warnings. But what you can see here is that it doesn't matter what age you are. You can be 20, 30, 40, 50 or 60; everyone takes in information they want to hear more than information that they don't. And so, we end up with a view like this of ourselves. (Laughter) Our mistake as teachers, as mentors, as employers | 431 | 465 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=431s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | is that, instead of working with this positive image that people so effortfully maintain, we try and put a clear mirror in front of them. We tell them: "You know, the image is just going to get worse and worse and worse." And it doesn't work. It doesn't work because the brain will frantically try to distort the image, using Photoshop and fancy lenses, until it gets the image it's happy with. | 465 | 491 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=465s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | But what would happen if we went along with how our brain works and not against it? Take handwashing, for example. We all know that handwashing is the number one way to prevent the spread of disease, and this is really important in hospitals. So, in a hospital here in the United States, a camera was installed to see how often medical staff do, in fact, sanitize their hands | 491 | 515 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=491s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | before and after entering a patient's room. Now, the medical staff knew a camera was installed. Nevertheless, only one in ten washed their hands before and after entering a patient's room. But then, an intervention was introduced: an electronic board that told the medical staff how well they were doing. Every time you washed your hands, the numbers went up on the screen | 515 | 544 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=515s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | and it showed you your rate of your current shift and the rate of the weekly staff. And what happened? Boom. Compliance raised to 90%, which is absolutely amazing. And the research staff were amazed as well, and they made sure to replicate it in another division in the hospital. Again, the same results. So, why does this intervention work so well? It works well | 544 | 576 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=544s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | because, instead of using warnings about bad things that can happen in the future, like disease, it uses three principles that we know really drive your mind and your behavior. Let me explain. The first one is social incentives. In the hospital study, the medical staff could see what other people were doing. They can see the rates of the shift, the rate of the week. | 576 | 604 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=576s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | We're social people, we really care what other people are doing, we want to do the same and we want to do it better. This is an image from a study that we conducted, led by PhD student Micah Edelson, and what it's showing you is a signal in the emotional center of your brain when you hear about the opinion of others. And what we found was that this signal can predict | 604 | 628 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=604s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | how likely you are to conform at a later time, how likely you are to change your behavior. So, the British government are using this principle to get people to pay taxes on time. In an old letter that they sent to people who "forgot" to pay taxes on time, they simply stressed how important it was pay taxes, and that didn't help. Then, they added one sentence, | 628 | 655 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=628s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | and that sentence said: "Nine out of ten people in Britain pay their taxes on time." And that one sentence enhanced compliance within that group by 15%, and it's thought to bring into the British government 5.6 billion pounds. So, highlighting what other people are doing is a really strong incentive. The other principle is immediate rewards. So, every time the staff washed their hand, | 655 | 686 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=655s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | they could see the numbers go up on the board and it made them feel good. And knowing that in advance made them do something that they, otherwise, may not want to do. Now, this works because we value immediate rewards, rewards that we can get now, more than rewards that we can get in the future. And people tend to think it's because we don't care about the future, | 686 | 710 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=686s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | but that's completely wrong, we all care about our future, right? We want to be happy and healthy in the future, we want to be successful, but the future is so far away. I mean, maybe you'll behave badly now and you'll be fine in the future, and maybe you'll be altogether dead. (Laughter) So, the here-and-now you would rather have that tangible drink, that tangible T-bone, | 710 | 735 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=710s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | rather than something that's uncertain in the future. If you think about it, it's not altogether irrational, right? You're choosing something sure now rather than something that is unsure in the future. But what will happen if you reward people now for doing actions that are good for them in the future? Studies show that giving people immediate rewards make them more likely to quit smoking, | 735 | 763 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=735s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | more likely to start exercising, and this effect lasts for at least six months, because not smoking becomes associated with a reward, and exercising becomes associated with a reward, and it becomes a habit, it becomes a lifestyle. So, we can reward ourselves and others now for behaving in ways that are good for us in the future and that's a way for us to bridge the temporal gap. | 763 | 787 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=763s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | And the third principle is progress monitoring. So, the electronic board focused the medical staff attention on improving their performance. This is an image from a study that we conducted, that shows you brain activity suggestive of efficient coding of positive information about the future. And what we found was that the brain does a really good job at this, | 787 | 811 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=787s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | but it doesn't do such a good job at processing negative information about the future. So, what does this mean? It means that, if you're trying to get people's attention, you might want to highlight the progress, not the decline. So, for example, if you take that kid with the cigarette, you might want to tell them: "You know, if you stop smoking, you'll become better at sports." | 811 | 837 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=811s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | Highlight the progress, not the decline. Now, before I sum up, let me just share this small anecdote with you. A few weeks ago, I got home and I found this bill on my fridge. And was really surprised because there's never any bills on my fridge. So, I was wondering why my husband decided to put that on our fridge. And so, looking at the bill, I could see that what this bill was trying to do | 837 | 859 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=837s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | is get me to be more efficient with my electricity use. And how was it doing it? Social incentives, immediate rewards and progress monitoring. Let me show you. Here are the social incentives. In gray is the energy use on the average energy use of people in my neighborhood. And in blue is my energy use, and in green is the most efficient neighbor. And my reaction to this was -- | 859 | 886 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=859s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | my immediate reaction was: "I'm a little bit better than average" (Laughter) -- a tiny bit, but still... and my husband had exactly the same reaction -- and "I want to get to the green bar." And then, I got a smiley face. That was my immediate reward and it was telling me, "You're doing good," and it made me want to put this on my fridge. (Laughter) And although I have this one smiley face, | 886 | 911 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=886s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | I can see an opportunity there to get two smiley faces. (Laughter) So, there's an opportunity for progress and it's showing me my progress throughout the year, how my energy use changes throughout the year. And the last thing this bill gave me: it gave me a sense of control. So, it gave me a sense of I was in control of my use of electricity. And that is a really important thing, | 911 | 937 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=911s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | if you try to get people to change their behavior, because the brain is constantly trying to seek ways to control its environment. It's one of the principles of what the brain is actually doing. And so, giving people a sense of control is a really important motivator. OK. So, what am I not saying? I'm not saying that we do not need to communicate risks, and I'm not saying that there's one-solution-fits-all, | 937 | 964 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=937s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
xp0O2vi8DX4 | but I am saying that, if we want to motivate change, we might want to rethink how we do it, because fear, the fear of losing your health, the fear of losing money, induces inaction, while the thrill of a gain induces action. And so, to change behavior in ourselves and in others, we may want to try these positive strategies rather than threats, which really capitalize on the human tendency | 964 | 994 | https://www.youtube.com/watch?v=xp0O2vi8DX4&t=964s | How to motivate yourself to change your behavior | Tali Sharot | TEDxCambridge | |
7agK0nkiZpA | well I'm honored to be here and my hat is my favorite hat I think it makes me look rather handsome this hat is made from a mushroom called Amadou Amadou is a birch polypore and a Madhu is a hardwood conch and this mushroom is responsible for human survival not too long ago there's no doubt that we all came from Africa we went north we discovered something new called winter | 0 | 45 | https://www.youtube.com/watch?v=7agK0nkiZpA&t=0s | Mushrooms as Medicine with Paul Stamets at Exponential Medicine | |
7agK0nkiZpA | oops this mushroom allowed for the portability of fire moreover you can hollow this mushroom out put embers of fire inside and carry fire for days and the fire keeper of our clans thousands of years ago were absolutely critical for the clans survival well this mushroom has other properties and when you boil this mushroom it delaminates and becomes mycelium a fabric and since | 45 | 69 | https://www.youtube.com/watch?v=7agK0nkiZpA&t=45s | Mushrooms as Medicine with Paul Stamets at Exponential Medicine | |
7agK0nkiZpA | some ladies in Transylvania have kept this tradition alive so this threat of knowledge has carried forth over thousands of years and so many threads of knowledge have been interrupted because of famine disease and war well this mushroom is first described by Hippocrates in 450 BC II as an anti-inflammatory as well as for Carter izing wounds another mushroom I brought | 69 | 94 | https://www.youtube.com/watch?v=7agK0nkiZpA&t=69s | Mushrooms as Medicine with Paul Stamets at Exponential Medicine | |
7agK0nkiZpA | mushroom friend of mine also is a polypore wood conch and this is a Garak on a gerakan is the longest living mushroom in the world rose exclusively in the old-growth forests now presently only known from Northern California Oregon Washington and British Columbia in a sky island or two in Central Europe it was described by ascribe ease in the very first materia medica as Alexei reom | 94 | 119 | https://www.youtube.com/watch?v=7agK0nkiZpA&t=94s | Mushrooms as Medicine with Paul Stamets at Exponential Medicine |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.