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
|
|---|---|---|---|---|---|---|
NTz4rJS9BAI | Ludvig if I saw I saw Dave we incorporeal Ike like we recruited people from CMU for this I mean it's like a lot of work to label and we see the exact same plot man we see the exact same plot so that you yeah again like just using this metric you see exam this big shift this again it's again it's a small distribution ship is within four point three frames this is 10 this is PM 10 | 2,730 | 2,756 | https://www.youtube.com/watch?v=NTz4rJS9BAI&t=2730s | Training on the Test Set and Other Heresies | |
NTz4rJS9BAI | this is in 0.3 seconds of each other you see it's a small distribution shift and again everything you don't see adaptive overfitting you do see some sensitivity to distribution shift let me skip this go I'll talk about the end finally kaggle two more things Kaggle CAG will released a nice meta-analysis that were metadata set of all their competitions with a lot of information not all the | 2,756 | 2,778 | https://www.youtube.com/watch?v=NTz4rJS9BAI&t=2756s | Training on the Test Set and Other Heresies | |
NTz4rJS9BAI | information we would have liked but a lot of information everybody knows on cavil you have a public and a private leaderboard the nice thing is almost in every competition here those are a ID split from each other this is cool because now what we should see is that if you train the public leaderboard and you train and you somehow evaluating the on the private you should just see | 2,778 | 2,795 | https://www.youtube.com/watch?v=NTz4rJS9BAI&t=2778s | Training on the Test Set and Other Heresies | |
NTz4rJS9BAI | clustering around the y equals x and that's exactly what we see clustering around y equals x and it's not just on two it's on like B's how many do we do look I forgot the number 117 I don't have all of them in here but really key in this case because you have iid splits you just see clustering around the y equals x so now so again evidence that distribution shift because | 2,795 | 2,819 | https://www.youtube.com/watch?v=NTz4rJS9BAI&t=2795s | Training on the Test Set and Other Heresies | |
NTz4rJS9BAI | these are these are iid evidence here is that the distribution shift is really what's causing us to be ok I'll stop there um oh sorry the most important slide of course is that you because no one day right the data set to make all data sets and we see Leon Batu wrote this beautiful oral history of the M this data set and also managed to reconstruct a bunch more examples and | 2,819 | 2,844 | https://www.youtube.com/watch?v=NTz4rJS9BAI&t=2819s | Training on the Test Set and Other Heresies | |
NTz4rJS9BAI | it's very nice fun read and again we see the distribution shift again between these two tests we were doing accuracy so it should be negative right yeah yeah I didn't want to I just cut this out of their paper I should already made it that all right so I'll stop there it sounds like we've seen this before we know this we knew this was true in boosting the interpolating training data | 2,844 | 2,875 | https://www.youtube.com/watch?v=NTz4rJS9BAI&t=2844s | Training on the Test Set and Other Heresies | |
NTz4rJS9BAI | we knew that was fine and it did seem to always make you like building bigger models seem to have better test error making your models big doesn't hurt but definitely does seem to be some other issues that are going to really be the pressing ones that we need to deal with moving forward I'm not sure Jonathan and I will talk about how we teach this to our undergrad machine learning class is | 2,875 | 2,895 | https://www.youtube.com/watch?v=NTz4rJS9BAI&t=2875s | Training on the Test Set and Other Heresies | |
NTz4rJS9BAI | very gently I think so how but I do think that for us the researchers and and for anyone in industry the big issues the bigger issues are this distribution shift is a real dangerous thing right if you're putting it in a car or you're making health care decisions and like that we I can show you if you're interested afterwards I could show you a paper which which | 2,895 | 2,914 | https://www.youtube.com/watch?v=NTz4rJS9BAI&t=2895s | Training on the Test Set and Other Heresies | |
NTz4rJS9BAI | demonstrates that actually a lot of the there's huge distribution shift effects in medic with all of the radiology there's a deep learning for radiology huge things where you're basically overfitting to the machine that took the image so that's dangerous if that's a life-and-death situation we already know that Tesla cars kill people what's amazing about Tesla cars | 2,914 | 2,933 | https://www.youtube.com/watch?v=NTz4rJS9BAI&t=2914s | Training on the Test Set and Other Heresies | |
NTz4rJS9BAI | two people have died driving under trucks with their autopilot on in Florida and what's nice about that is that machine learning generalization you should never make the same twice I showed you your corner case and yet two years apart same errand same accident so this is a real thing that's something we have to think about but you know I think that's also cool there are | 2,933 | 2,951 | https://www.youtube.com/watch?v=NTz4rJS9BAI&t=2933s | Training on the Test Set and Other Heresies | |
fgwurrihq4A | you haven't heard this yet so I'm Gordon our facilitator is Florian so so today the paper we're going over is unsupervised data augmentation the primary author qi j xie and several other co-authors from google brain and carnegie mellon this is we're gonna go over motivation so deep learning is sorry typically requires a lot of labelled data in order to succeed and | 0 | 30 | https://www.youtube.com/watch?v=fgwurrihq4A&t=0s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | and label data is very expensive so that's one of the main motivations for this paper some of the less costly ways of applying improving deep learning are using unlabeled data which is much more abundant and and easy cheaper to accumulate and data augmentation which basically stretches your supervised labeled samples further and as well are we good with the sound data augmentation | 30 | 61 | https://www.youtube.com/watch?v=fgwurrihq4A&t=30s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | has mostly been applied in the supervised setting and so we want to see if it can be applied in the unsupervised setting as well the main contributions which we'll get into are applying state-of-the-art data augmentation to semi-supervised learning a training technique called TSA training SiC leg kneeling that effectively prevents overfitting when you have much more | 61 | 83 | https://www.youtube.com/watch?v=fgwurrihq4A&t=61s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | unsupervised data than supervised data and they they achieve performance improvements on multiple text and vision benchmarks and then they also introduce a method to even the prediction distributions over across classes for unlabeled and labeled data semi-supervised learning I hope people online know what this is I probably won't explain it again so we'll get into | 83 | 110 | https://www.youtube.com/watch?v=fgwurrihq4A&t=83s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | smoothness in forcing okay so this is one approach to semi-supervised learning and the general idea here is you you try and regularize a models prediction to be less sensitive to small perturbations applied to to the the input data so that in potato can be label or unlabeled and when we say perturbations we're basically talking about adding some sort of noise to the | 110 | 139 | https://www.youtube.com/watch?v=fgwurrihq4A&t=110s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | inputs ampuls and so yeah you want your model to given a sample and an Augmented sample or perturbed sample you want the models predictions to be similar on both I think that's what I just said so enforce the predictions to be similar and in general you want a good model should be invariant to small perturbations on this input data that don't actually change the nature of the | 139 | 171 | https://www.youtube.com/watch?v=fgwurrihq4A&t=139s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | example and yeah so data augmentation is a technique to boost your training data size and the diversity of it so the general idea is you're augmenting in some way again adding some noise to your input samples so that you cannot that you can both get more training data and have more diverse training data and I guess what we'll see some examples of what diversity means so yeah basically | 171 | 208 | https://www.youtube.com/watch?v=fgwurrihq4A&t=171s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | you apply some sort of transformation you have a transformation function the apply to your input data and in data augmentation there's always this trade-off of diversity and validity that's being managed so so yeah you want to create novel and realistic training samples without augmenting them so much that you change their underlying inherent label so diversity it means | 208 | 234 | https://www.youtube.com/watch?v=fgwurrihq4A&t=208s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | growing the the reach of your data set or making your data set more broad and validity is making sure that you're not blowing up your samples so much that they are no longer recognizable or they're not related to the label that they should have assigned any questions so far so this this is what supervised data augmentation looks like here so here Q is is our transformation function so you | 234 | 268 | https://www.youtube.com/watch?v=fgwurrihq4A&t=234s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | can see it's conditioned on an input example X and then X hat is the Augmented data sample and so so basically we're trying to minimize the log likelihood the negative log likelihood of the true ground source ground truth label which is y star given an Augmented sample X hat and and so yeah you can see this as basically an additional training signal that's being | 268 | 303 | https://www.youtube.com/watch?v=fgwurrihq4A&t=268s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | sent to the objective function that that is hoping to yeah that is that is just using the the Augmented samples and then very similarly or actually slightly differently unsupervised data augmentation so this is when you have unsupervised unlabeled data this is a common way to to use that data for data on plantation you can basically take examine the output distribution | 303 | 336 | https://www.youtube.com/watch?v=fgwurrihq4A&t=303s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | prediction probability distributions for an unlabeled sample so here that's X and then an Augmented unlabeled sample X hat again and and you've got the same transformation function Q that you had in the supervised setting and so really what you're trying to do is in this case minimize the divergence or the the difference between these two probability distributions so you're trying to | 336 | 368 | https://www.youtube.com/watch?v=fgwurrihq4A&t=336s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | normalize or regularize the predictions on the Augmented samples to have similar class distributions to to the unag mented unlabeled data does anyone have any questions here doesn't minimizing this minimize they are therefore as well minimize this minimize the other in this case I'm sorry maybe I misspoke earlier in this case these are X's all labeled labeled samples here ordinary boys why | 368 | 414 | https://www.youtube.com/watch?v=fgwurrihq4A&t=368s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | is so why here is not the ground truth label y here so this is y star that's the ground truth annotated label and here we just have the the output prediction distribution or from the model for both the unlabeled and the Augmented unlabeled sample okay what's the difference oh yeah good question so theta it is implying that these are parameters that are being updated so the gradient is | 414 | 451 | https://www.youtube.com/watch?v=fgwurrihq4A&t=414s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | passing through these I'm pretty sure that the idea with theta tilde is those parameters are frozen so they're they're not they're not updated in the objective function data here yeah yeah so this is the instead of thing yeah these would be these are two different settings but but yeah in both cases they're the model parameters question here is the transformation function so so the | 451 | 497 | https://www.youtube.com/watch?v=fgwurrihq4A&t=451s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | augmentation function basically the way that you're adjusting your input samples so the first version sees X's or actual the data that you have that you have labels for in this one yeah so in sorry in this case the unsurprised a documentation you're assuming that you have so yeah I guess you can see here you've got a so in U and U is an unlabeled set whereas in the | 497 | 529 | https://www.youtube.com/watch?v=fgwurrihq4A&t=497s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | first one you have x and y star in in a labeled set so yeah there's no labels in this data at all so this this actual approach there are a few different ways you can do this this specific approach of using the KL divergence between the unlabeled or started the Augmented and unadmitted was from a paper in 2018 VA t believe was the mall so so they borrowed that | 529 | 556 | https://www.youtube.com/watch?v=fgwurrihq4A&t=529s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | approach here and really the main difference they applied in this case is the the transformation function so what they did with Q and so that's what we're going to talk about now so this idea of targeted data augmentation so so over conventional methods such as adding Gaussian noise or affine transformations perturbations like that if there are a few advantages applying targeted | 556 | 586 | https://www.youtube.com/watch?v=fgwurrihq4A&t=556s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | augmentation so one is that they give a valid and realistic perturbation so so the idea is when you apply some of these state-of-the-art data augmentation methods the the output augment example it is still very much in in the same distribution as the the sample it was transformed from so so these are sort of realistic augment augmented examples whereas when you just apply | 586 | 621 | https://www.youtube.com/watch?v=fgwurrihq4A&t=586s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | some random Gaussian noise it can often make the the data point if you apply too much make it sort of very unrecognizable and not realistic so so not really as we say valid it also applies a diverse perturbation so again if you want to use those other methods just adding Gaussian noise you're usually not going to change you're not going to be able to change your input samples significantly enough | 621 | 652 | https://www.youtube.com/watch?v=fgwurrihq4A&t=621s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | and so you just end up with sort of local changes to the samples whereas with targeted augmentation you can really generate diverse diverse samples that are much more useful and in growing your training set and then as well we'll see what some of the methods they had a targeted targeted inductive bias so so yeah you can actually apply approaches that are optimized for the tasks that | 652 | 686 | https://www.youtube.com/watch?v=fgwurrihq4A&t=652s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | you're that you're solving in the particular data set and so so we'll see we'll see an example of that in the augmentation strategies they applied yep so this is the the training set up that they applied so on the left hand side we have the labeled data and it's so split up to x and y star part of me x is fed in through m that's the model and then we just have the standard supervised | 686 | 724 | https://www.youtube.com/watch?v=fgwurrihq4A&t=686s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | cross entropy loss being calculated here that's feeding up into a final loss on the right hand side is where they take the unlabeled data and and they do two things so so one they feed it through the same model and and then as well they take that sample the unlabeled sample and then they apply some augmentation to it so we'll get into all these different segmentation | 724 | 750 | https://www.youtube.com/watch?v=fgwurrihq4A&t=724s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | strategies so that results in X hat so this would be their their Q the transformation function so they get X hat and then they feed that through the model as well and then they take the the output of the model on X and on X hat and they feed that into the unsupervised loss function and that's the the KL divergence that we saw here so basically they just take a weighted sum of the | 750 | 785 | https://www.youtube.com/watch?v=fgwurrihq4A&t=750s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | supervised and unsupervised loss and and combine those together question yeah what is supervised course and complete loss this is just the standard loss function for okay across the so you take the the ground truth label and and compare that to the output prediction probabilities yeah Joey so none of these data seem to have two tilde over it does that mean that oh yeah that's a great | 785 | 811 | https://www.youtube.com/watch?v=fgwurrihq4A&t=785s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | point I pretty sure they forgot to apply a theta on this this prediction distribution so the idea is when you when the gradient propagates back through the through the model it will run through the supervised portion and it will also run through the Augmented application up here but but I'm pretty sure it's not flowing through just the unag mented unlabeled data good question | 811 | 846 | https://www.youtube.com/watch?v=fgwurrihq4A&t=811s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | I can't remember if there was a justification for that what's up and okay we're in this flow does the targeted or the diversity augmentation where does that come to play in this flow the target okay so so the idea is that based on the type of input data that you have for the unlabeled data they'll apply a specific augmentation to that sample so so here they've listed the ones that they've | 846 | 878 | https://www.youtube.com/watch?v=fgwurrihq4A&t=846s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | used but the idea is they only use one at a time so based on the type of data you're using and the particular data set they'll apply a particular augmentation strategy so so I think in the next slides it'll show the particular policies they just they just sort of listed them all here on the left hand side is a leap of data so you have the ground truth the white star the right I | 878 | 906 | https://www.youtube.com/watch?v=fgwurrihq4A&t=878s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | know if you don't have one right so the lesson1 the left hand side is your doing the prediction training and the right hand side you do the prediction infant yeah so so on on the right side yeah the loss function does look different than then on the left but in both cases you have you have a version of the model and with its parameters that that can be updated with signal from from the loss | 906 | 941 | https://www.youtube.com/watch?v=fgwurrihq4A&t=906s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | function so so in the unsupervised case this is the this is the function that is leading to the unsupervised or signal flowing through the unsupervised portion yeah so there's just two different loss functions one looks like this and then the other is the center oh my oh my okay scheduling of training do they train all supervise first then on label or are they training all mixed together that's | 941 | 981 | https://www.youtube.com/watch?v=fgwurrihq4A&t=941s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | a great question and that relates to the big contribution that I mentioned earlier TSA the training signal and annealing so we'll probably wait until that two to explain that yes so for the unscrew right side it doesn't matter what the label and what the model what with the label is not about the model is right like the lost isn't it yeah really what you care about here is | 981 | 1,005 | https://www.youtube.com/watch?v=fgwurrihq4A&t=981s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | that the the amount that you're changing or sample by is not differing significantly from the the same model prediction on the sorry the model prediction on the same unlabeled sample okay so so going into the augmentation strategies the first one we'll talk about is Auto and all so Auto augment learning augmentation strategies from data the general idea here is they have input | 1,005 | 1,047 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1005s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | data on the left and they basically have a model that will automatically search through multiple different augmentation strategies so you can see here in policy one there they're making a transformation rotation to the input data the other ones are mostly changing the color of the samples and and so the idea is the the model will automatically select the policy that is adding the | 1,047 | 1,078 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1047s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | most novel signal to to the yeah to the training so for example if and they find that in different in different data sets different policies are optimal so in in one data set you might need to modify the color of your images a lot to get more diversity and improve your training training set and in another you might need to rotate the images etc so basically it's a again coming back to | 1,078 | 1,110 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1078s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | the idea of being a targeted policy and this is something that you can vary on a task by task basis so even on a particular data set you can see what's the optimal augmentation approach any questions on this one so that was the one that they applied for vision and then they have to for text one is back translation so the general idea here it's pretty intuitive | 1,110 | 1,140 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1110s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | you take a question in one language in their case they used English to French so they trained a machine translation model between those two languages then they translate the English sentence into French they then take that French translation from the model and they translate that back into English and and then they use that as the Augmented sample in the model so obviously any | 1,140 | 1,170 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1140s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | machine translation model is going to have some some loss and it's not going to exactly translate something the same way back and forth so in this case you can see a lot of it has stayed the same this is Google Translate but but this word crankily was previously grin jingly and I think spoof spoof gets translated to tragic travesty so so this is one way that you can augment your samples and | 1,170 | 1,202 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1170s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | then the other one that they have for text is this tf-idf based word replacement so and they use this for text classification so so the idea here is sometimes in back translation the the Augmented transformation might actually miss translate some of the key words for that sample and in the classification tasks and so here they basically assign an IDF score to each word in the in the | 1,202 | 1,237 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1202s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | sample and then they randomly sample or randomly swap out words and giving a higher likelihood to swap words that have a low IDF score so here you can see I've just sort of created this example and so in this case the words this in decides to etc are transformed or swapped but but the words that are a little bit more rare and therefore have a higher IDF score such as pathetically cringing | 1,237 | 1,269 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1237s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | lease poof those are more likely not to be swapped out and so that's yeah based on the intuition that certain keywords sometimes are really useful for text classification any questions on the strategies those are the three okay so so now we come back to the question of how do we balance the need for having a large model so so when we're dealing with unlabeled data of which we have a | 1,269 | 1,303 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1269s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | really often you have much higher volume you have much more unlabeled data than say labeled data and so you generally would need a very large model to to train on that data so but but you may have a small amount of labeled data so you wanna the question is how to balance a need for a large model while preventing overfitting at the same time and so they're their answer to that | 1,303 | 1,332 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1303s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | question is to gradually release the training signals of supervisors examples as the model is trained on more and more of the unsupervised examples so I'll show the the equations for for all this okay so here let's see so B here is the batch that is just a renormalization constant and the key is is this part over here so again this is sorry this is the objective function and | 1,332 | 1,369 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1332s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | and so really what they've introduced is this portion on the right hand side where they say and this is all with regards to the labeled samples so n or what's this constant in Ada Adah yeah so so a dub G is is a threshold that varies with your training progress so we'll see that in the next slide but basically it's a threshold and and if the if the models prediction | 1,369 | 1,407 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1369s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | probability on the on the correct class for a labeled sample is above that threshold then then that this will evaluate to zero and therefore that sample signal will not get propagated through to the loss function at that time and yeah so so the constant will will change over time but generally the idea here is is that one yeah when ADA is is small then you'll | 1,407 | 1,445 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1407s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | pretty much be rejecting most most signal from labeled samples from going to the loss function so you'll be preventing the model from overfitting on on say a small set of labeled data so that was the that was the reason they introduced this any questions question it is I the indicator function yes sorry I didn't say that I is the indicator function here and and and Zed is just to | 1,445 | 1,474 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1445s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | rebalance or renormalize the effect from that during this rain ah ADA goes from it's always increasing okay so I'll show the schedule so they they introduce a few different schedules for for ADA so here's the equation for ADA on the right K is the number of classes in the classification example and lambda of T you can see in in the plot so lambda of T varies with the training progress and | 1,474 | 1,514 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1474s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | and as lambda of T increases or sorry as training progress increases lambda increases and as well the threshold also increases so at the beginning lambda will be zero and ADA will be one over K so 1 over K is 1 over the number of classes which is just the random chance of predicting a sample and then at the end ADA will be 1 and so coming back to this equation if ADA is 1 that means | 1,514 | 1,550 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1514s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | every single sample practically will be will be carry forward towards the the loss function and and at the beginning when it's 1 over K only the predictions that only the samples that the model is very unconfident on will actually be used in the loss function and so so the idea here the intuition is that if you have a small labelled set so you have only a few examples that have labelled | 1,550 | 1,582 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1550s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | labels you want to avoid your model overfitting on those at the beginning of the training so they suggest using this exponential schedule for for the case where you have a low number of labeled examples and so yeah the idea is at the beginning you you won't be feeding as much signal from those samples to your loss function but by the end you can start releasing more and more of it once | 1,582 | 1,610 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1582s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | more of the unlabeled data has been incorporated and then conversely if you have say a large number of labeled samples then you can use a log schedule the green line here and so that will release a lot of the supervised signal at the beginning and less at the end any questions on this think that's it before the break so we'll just take a 5-minute break and then afterwards we'll go over the | 1,610 | 1,642 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1610s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | results and have some discussion come back people online the sound is working sounds good okay great so we're gonna do a five minute break now we just did a five minute break we're gonna do experiments so basically they they applied this method to two different types of tasks Texas classification and then image recognition so a couple of vision benchmarks these are the actual data | 1,642 | 1,684 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1642s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | sets that they used so there's a mixture of binary and five class text classification most of it is sentiment I except for dbpedia which is I believe categories and I think dbpedia actually has fourteen classes or ten classes the two image benchmarks both have ten classes they also use image net they test on image net and they do some ablation studies for TSA and for the targeted | 1,684 | 1,714 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1684s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | augmentation so for text classification and the settings for the labeled data so the goal here was to use a small number of labeled samples so for binary classification that means twenty just twenty labeled samples and the rest of the data coming from the unlabeled and for the five class classification they found they needed to use a bit more so here they use twenty five hundred total | 1,714 | 1,745 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1714s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | samples and and so that's five hundred per class so when speaking to the author the goal here was to find out how low could they go so I was asking him did you experiment or how did you stumble across these numbers and really they wanted to determine how few examples they could labeled examples they could keep while still achieving really strong results from the unlabeled data for IMDB | 1,745 | 1,776 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1745s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | which is one of the the binary classification they use the concatenation of the training set that they didn't use so the training data that was not used for super as supervised they use that as unlabeled and then they use the rest of the unlabeled set so I believe the total training set is 25,000 and the unlabeled set is about 50,000 so that's how much unlabeled data they're using there and | 1,776 | 1,802 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1776s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | for Yelp and Amazon they obtained some really large review datasets and I believe there was something like 6 million samples in in the unlabeled and so for the most part again another choice was that I think for the most part they used one Augmented sample per unlabeled sample so so but but he said that that would be something that might be a task specific parameter and that | 1,802 | 1,832 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1802s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | you could adjust so for some tasks for each unlabeled sample you might want to make a couple of augmentations and use both and as for the model they try a few different initialization schemes so all are working on the transformer architecture applied in Beart so they have just a random initialization then they have bird base bird large and Bert large fine-tuned on on the unlabeled | 1,832 | 1,865 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1832s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | data the in domain on little data so the same unlabeled data that they're using and and for each setting they compare the performance for each of these settings with and without the unsupervised data augmentation method UDA so here are the results for for the text benchmarks so you can see at the top the data set name and then below that the number of supervisors examples that | 1,865 | 1,899 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1865s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | exist in that data set the the top two results are the pre bert state-of-the-art so that's that's how influential Burt was there's a before bird and after bird so they report both in some cases Burt was better than the state of the art in other cases I I guess it's always okay and then their results the results from their paper are in the bottom in the semi-supervised | 1,899 | 1,929 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1899s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | setting again they've got the different initialization strategies on the on the left and you da X or check indicates whether or not they use u da and and so these are error rates here so lower is better and below the the data set name in the bottom you can see the number of labeled samples that they used so for IMDB they literally only use 20 examples and and when they apply when | 1,929 | 1,961 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1929s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | they use an initialization from fine-tuned Burt model they're they're able to beat the state of the art and the pre bird state of the art so that literally means they're just using 20 labeled examples along with Burt which is fine-tuned on unlabeled data and and then they use augmented samples from the unlabeled unlabeled data set so that's that's one of the most significant | 1,961 | 1,995 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1961s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | improvements and they found across the board they they got very close it or actually beat the the the state of the art for these tasks I think the one that they found the most difficult and they perform the the worst on was the de 5 class classification for both Yelp and Amazon you can see that their results are still a bit a bit off from from the baselines so yeah a really big finding | 1,995 | 2,026 | https://www.youtube.com/watch?v=fgwurrihq4A&t=1995s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | here is that obviously just with using Burt and know you da you still get 6.5 error with just 20 labeled examples so a lot of this is just indicating how much information is being contained in Burt but but it's quite significant that this shows that you can use pre-trained language models along with UDA so it's it can be complimentary to pre-trained language district models yeah that's | 2,026 | 2,067 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2026s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | right on the 20 labeled samples plus fine tuning for the last case but but yeah that's it yeah they do they do do some studies on the vision benchmarks that investigate whether or not the whether or not doing augmentation on the unlabeled data is actually more advantageous than doing it on the labeled data so but they don't think they did that for for this so yeah it | 2,067 | 2,106 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2067s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | could be a good unless I'm wrong that could be a good feedback they might have their I should say that they did do something speaking with you out there they did do other experiments that they didn't report here and I think I might have asked about that I just can't remember if if that was something they did but if they did it wasn't as as performant so for the vision benchmarks yeah so | 2,106 | 2,137 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2106s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | they wanted to use the the same model that was used for prior semi-supervised work and this was the wide residual networks with depth 28 and with 2 and and so yeah they use the same exact labeled samples that Auto augment use to find its optimal policy so for C far 10 which is a 10 10 Way image classification task they have 4000 samples and for this is Street View | 2,137 | 2,170 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2137s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | house numbers it's a digit recognition data set they used a thousand labeled examples and so they do 10 10 runs with this model and and calculate the average and the standard deviation so here are the results here these on the left you see the the fully supervised setting at the top so with no augmentation and then following that you see different different augmentation techniques that | 2,170 | 2,208 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2170s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | that that are applied to to the unlabeled data so I think for I can't remember how many samples but there are roughly maybe 50,000 unlabeled samples for 4c far 10 yeah I'm not sure of the numbers right now but I think you might remember earlier I mentioned V 80 which was the the paper that they got the the idea to take the KL divergence between the distribution of unlabeled and an | 2,208 | 2,243 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2208s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | Augmented unlabeled sample so really what you're seeing here so obviously they performed the the previous state of the art method but between UDA and v80 the only real difference is the perturbation or transformation function that they applied since they're both using the same KL divergence technique and so so this is indicating that that targeted data augmentation strategy was was | 2,243 | 2,275 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2243s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | helpful so is there any metric or measure that we can look up into in order to see like they are saturated and why not using hundred-thousand right so we would say morning that's better but there's a point of a saturate there like the diversity that they don't were like the distribution or being the new thing that you're defining in our data we are like many in the data which doesn't | 2,275 | 2,311 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2275s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | matter after that right so for example Wendy chose to see a thousand there's a reason for that certain metric for us to see that's level oh you know I don't think they provide our metric there might be a good discussion for at the end though when we get to the discussion points like so so yeah keep that out of minds but but as far as I remember there wasn't anything | 2,311 | 2,333 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2311s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | provided Hockman Auto admins will find the optimal policy but I didn't read the paper in depth I'm not sure if it optimizes anything else such as the number of samples as far as I know it's just finding the optimal policy for a particular data set the optimal augmentation transformation and okay so they also perform some experiments on imagenet and the the motivation here was | 2,333 | 2,374 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2333s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | they the first the initial data sets they use all had between two and ten classes so they wanted to and and and they all had a low number of supervised examples four thousand or less so they wanted to use a data set that had a hot much higher number of classes it's a bit of a harder task and much more supervised examples and see if this approach was still applicable or if it | 2,374 | 2,402 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2374s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | was really only sort of a niche improvement for for smaller data sets and then they also wanted to see if they could make use of out of domain unlabeled data that had different class distributions so keep in mind in all of the previous examples the unlabeled data sets that they were using largely were just the the actual labeled data without the labels being used so so it's very | 2,402 | 2,433 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2402s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | much in domain literally samples that have a true label they just didn't tell the model what the label was so they wanted to see if this would be applicable if you didn't have yeah if your labels coming from out of domain and so imagenet overall has almost 1.3 million images and and about a thousand different classes so the settings they do a couple of settings so one is image | 2,433 | 2,468 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2433s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | they call image net 10% and this is where they they take roughly 10% of all image net data and use that as labeled samples and they use all of the the rest of image net as unlabeled data so ten percent would be I guess one hundred and thirty thousand samples and and then they would have over a million unlabeled or about a million and then the the other one is the fully supervised | 2,468 | 2,503 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2468s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | scenario so this is where they use the entire image net data is supervised data and they obtain extra unlabeled data from data set called jft it's another image I believe it was automatically generated image data set so they essentially train a model on image net and they use that model to to source out the most relevant samples from the jft data set for each class in the image net | 2,503 | 2,534 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2503s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | so they basically for each class they take the I think it was thirteen hundred most most relevant examples from jft and they and they use that as the unlabeled set for for their experiment any questions okay the baseline model that they used was ResNet fifty here so they did encounter some some errors some issues with with image net so they observed that they had flat class | 2,534 | 2,572 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2534s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | distributions so the prediction probability distributions across classes was pretty flat or uniform for the unlabeled and the Augmented unlabeled samples so there really wasn't much signal coming through from from the unlabeled part of the the training set up and yeah so there's there's a probably to do with the fact that there are so many more classes and and there's also | 2,572 | 2,601 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2572s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | so much supervised data available here actually sorry no but the amount of supervised data wasn't an issue I think this was more of an issue even for the image net 10% so where they only had 10% of the training data as supervised this was this was an even larger issue so this led to the unsupervised training signal being pretty much dominated by the the supervised signal and so that | 2,601 | 2,632 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2601s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | their solution they another one of the contributions I mentioned in the beginning was to sharpen in a few different ways the prediction the predicted distribution produced on unlabeled samples so that there would be basically to encourage the model to use the training signal from from the unlabeled samples so the specific techniques that they use were entropy minimization so | 2,632 | 2,659 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2632s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | they added an over an entropy term to the overall objective to regularize the predicted distribution on the Augmented examples to have a low entropy so again to discourage these uniform distributions of probabilities they also did soft max temperature control and here this is to control the temperature of soft max when they're computing the prediction on the original example and | 2,659 | 2,690 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2659s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | then confidence based masking so this was where they basically removed any samples that the unlabeled samples that the model was not very confident on so all of these approaches were were to try and sharpen the probability distribution of the on the unlabeled Augmented samples for these sessions but the car if the actual distribution of your training data your related training data is | 2,690 | 2,724 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2690s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | uniform and then you try to sort of force it to take a different shape on your unlabeled a of what we shouldn't that prepared for that why should that work so the the number of samples is uniform across the classes but here this is this is saying the problem was that the the models predictions prediction probabilities across the different classes for an individual example were | 2,724 | 2,755 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2724s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | more uniform so so the example has some category let's say it includes a giraffe so that should be the true label but the actual output prediction distribution was was pretty no uniform so they wanted to find ways to encourage it to be sharper to - yeah to be less less prone to just being killed divergence per batch no I believe it would be calculated per sample pretty sure it | 2,755 | 2,797 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2755s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | would be each sample would look at the figure so their first average error and then the unsupervised error and then it will do back propagation right hmm I did do it like separately but the way I add necessarily do it all together so the mass calculation kind of patchwork L the Virgin's okay yeah so I suppose that they do these are the results here that the left | 2,797 | 2,827 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2797s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | is the image net 10% again with the baseline its resonant 50 and the right is is the fully supervised image net setting and top 1% is the the models accuracy on for its first top prediction top 5 is its accuracy looking within the first 5 samples and and yes so you can see that it improved on on the baselines in both cases I guess the smaller improvement for the fully supervised | 2,827 | 2,862 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2827s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | setting where they're using an additional 1.3 million samples from jft yeah over the the previous Auto augment policy I think for image net 10 was it for this one yeah I think they also so someone was asking earlier about why not using the why not use the labeled examples for augmentation for the baseline and I think they also did run that experiment here so they they used | 2,862 | 2,895 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2862s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | the 10% labeled examples as for augmentation and so that I think got something like 58 accuracy so still significantly below and yeah moving on to the ablation studies so they want to do they did a couple the first one is for TSA so to determine if this training signal annealing actually made a difference and so they did it for Yelp 5 and see far 10 in the Yelp case they | 2,895 | 2,938 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2895s | Unsupervised Data Augmentation | AISC | |
fgwurrihq4A | didn't use Bart pre-training just to really make sure that there was all the information was coming from from the data use and not from the pre train distribution or language model and yes so you can see in the first case where there's an X that's where they're not using any tsa and so you can see in both cases applying some schedule does improve the results and you can also see | 2,938 | 2,967 | https://www.youtube.com/watch?v=fgwurrihq4A&t=2938s | Unsupervised Data Augmentation | AISC |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.