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
|
|---|---|---|---|---|---|---|
T35ba_VXkMY | the loss that you have at the end of that matching so this loss here would only then count wherever these connections are that loss is going to be your training loss okay so this solves the problems we had before it is not dependent on the order because if you reorder the things your minimum matching will simplify it will simply swap with it it is it is um if you output the same | 725 | 754 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=725s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | bird multiple times only one of these is going to be assigned so if if this here is that bird only one of them only this one maybe is going to be assigned to that one and the other ones can't be assigned to that one are forced to be assigned to a different one let's say this one here and are going to incur a loss so you encourage your model to output let's say diverse bounding boxes | 754 | 779 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=754s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | different bounding boxes for things okay so this D solves these problems and it's very clever and there are algorithms to compute these these minimum matchings and they use the Hungarian algorithm which will give you exactly such a matching again this is possible because you have n things on each side and the N is in effect here is a the maximum of objects that | 779 | 804 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=779s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | you can detect at once I guess if there is less you can simply pad right here and then the model of course is encouraged to come up with the equal number of no class predictions because if it outputs a prediction when it shouldn't right if it already predicts two things and these are assigned to these two things and then it outputs one more thing it is going to be penalized | 804 | 830 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=804s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | because it should output three things with no class but it has output one-to-many with a with a class is going to be penalized okay so the this is a pretty pretty cool thing it again it relies on the fact that you have n on both sides but you can make n so large that basically it covers all of the cases so you can make n like 50 so you can detect up to 50 things in a scene | 830 | 863 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=830s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | alright that's the algorithm in a high-level they do show their loss here you see the loss ultimately is going to be so it's going to be over this matching right here that's the minimum bipartite assignment that basically minimizes this total loss over your prediction and label matchings and the loss they come up with here I said you have to give the algorithm a loss is | 863 | 894 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=863s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | this one and they kind of go into how they do it I don't think it's super important so the the class algorithm sorry the loss on the class labels I think it's going to be a soft Max or a sorry a cross-entropy loss like in usual classification and the loss on the to say whether to bounding boxes agree is a mixture of the l1 loss that compares to bounding boxes and this iou loss which | 894 | 924 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=894s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | is not dependent on the scale of the bounding boxes it kind of computes how much fraction of the two bounding boxes overlap but in any case the lost base they consist of saying how eyeli how much do the labels agree and how much do the bounding boxes agree okay again this is only possible because after that you compute this matching otherwise you would have no clue which boxes to come | 924 | 949 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=924s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | which predictions to compare to which other predictions so let's look at this architecture a bit more in detail as we said you have this what they call the backbone which is a convolutional neural network and with that you put in some positional encodings now I already said the you should look at the these features right here as just smaller feature versions of the image but they | 949 | 977 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=949s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | still have some image nature then they are flattened so once they are put in the transformer encoder because the transformer is naturally a sequence processing unit okay so it takes in just a sequence of vectors right here and since an image is not a sequence what you'll do is if you have your image features and we said we have a bunch of channels let's say we | 977 | 1,005 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=977s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | have four channels and their height and width and see you're going to unroll and flatten that into one sequence so this is height times width you basically unroll across these axes right here into this axis and it's channels I so basically you have a sequence here of of C dimensional feature vectors that you then put into your encoder okay so your encoder will now transform this sequence | 1,005 | 1,047 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1005s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | into an equally long sequence yet again of features and the good thing about a transformer because why do you use a transformer the good thing about the transformer is that in such a sequence and I've done videos on transformers it you can basic mainly look at the video attention is all you need if you want to under than this more fully this thing can basically have a tension so it has | 1,047 | 1,077 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1047s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | attention layers it can attend from each position to each position in a one-shot manner so as it transforms this representation up the transformer layers at each step it can basically aggregate information from everywhere in the sequence to anywhere else and therefore it's very it's very powerful if you have a sequence and you need sort of global connections across the sequence this is | 1,077 | 1,108 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1077s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | very good for a language processing because in a sentence let's look at this sentence the input images are matched together all right applying blah blah blah blah blah blah blah blah blah blah and then there is they write the word they and you need you need to know that they refers to the input images okay and but you see this is very very far away in the sentence so | 1,108 | 1,135 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1108s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | you need a model that makes use of long range dependencies and they make the case that in such a task right here you also need the long range dependencies because these bounding boxes as you see right here there can be quite large so if you have an image you need that this part here communicates with these and this and this and this part basically anywhere in the bounding box and these | 1,135 | 1,160 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1135s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | bounding boxes can be quite large so the transformer architecture actually makes sense here now I want to go a bit later into why I think it actually makes even more sense for a bounding box detection but right now I just want to keep going through this through this architecture right here so if my computer here decides to come back yes we can go on so what will get out is yet another so in | 1,160 | 1,191 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1160s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | here we put this thing we put down here we put into the transformer encoder and we get an equally sized equally shaped sequence out of the transformer encoder you see that this thing here goes as a side input into this transformer decoder so the transformer encoder here is just a bit more of a feature mapping technically just for the architecture you could think of just putting this | 1,191 | 1,217 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1191s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | into here but of course it's gonna go better with the transformer encoder the transformer decoder now does something similar but you see it has the encoder as a side input this is very much like this is not like Burt Burt is like a only encoder transformer whereas this is much like the original attention is all you need transformer that has an encoder and then the decoder as a side input | 1,217 | 1,246 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1217s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | basically as conditioning information has the encoder output what does the decoder do again since it's a transformer it's going to take a sequence and output a sequence the sequence it takes is right here is what they call object queries and this also is different from the attention is all you need papers and they don't do it autoregressive lee they just do it one | 1,246 | 1,268 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1246s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | shot what does it mean it means that you start with a sequence here of four things and this is these are the this is this big n and you out you output the sequence of a sequence of four things and it's important to see what they're going to end up so these things are then directly going through a classifier that now outputs the so these things here are these class label bounding box outputs | 1,268 | 1,299 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1268s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | okay so each of these things is going to after transformation end up being one of these bounding boxes either defining an object or saying that there isn't an object somewhere okay you see here this bounding box refers to this bird this bounding box refers to this bird so each of these things is going to to be one bounding box and the what they call object queries is the question of course | 1,299 | 1,329 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1299s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | is what do you input here right actually I want to transform this image information that comes from the left here I want to transform that into the bounding boxes what do I input here and the answer is you just input at the start you just input n random vectors because what's that gonna give you is basically n outputs you want and outputs because you want n of these | 1,329 | 1,353 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1329s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | bounding box classifications so you need n things and if I input n things into a transformer it's going to give me n things as an output and then in each step I can simply condition on the information that comes in the images and it it'll give me right then I can incorporate that information it's a very deep learning way of thinking about it actually that you just need the | 1,353 | 1,377 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1353s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | information somewhere in there and I need n things now they go more into detail into this transformer architecture help help in the helpful fashion in the appendix and will go there quickly so this I think here makes more sense so the image features come in here right and you see this is just a transformer stack an encoder stack of multi-head self attention and feed-forward in instants | 1,377 | 1,409 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1377s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | wise or like token wise feed-forward network and then that information is taken and is given as conditioning information over here now in here as I said you input these object queries which at the beginning are just n random vectors and what you're going to do you Argos are going to feature and code them and then you combine it with this image information so ultimately if you think | 1,409 | 1,438 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1409s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | of this one of these things one of these things is going to be a vector right and then that vector is going to be transformed and then it will have as it is transformed it will have the opportunity to basically look at features that come from here now the arrow is in the wrong direction so you have already taken the image and you've transformed it into a feature | 1,438 | 1,464 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1438s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | representation which is also a vector right you have the features of the image right here now as you transform this vector this object query queue you have the opportunity to look at the image features right and that's how do you get the image information in there so the image features will come in here transform that through attention so this is an attention mechanism on the image and | 1,464 | 1,493 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1464s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | then what you will output is a bounding box and a little class label it's really hard to explain I would guess you need to understand really what attention mechanisms are and of course the crucial part of of course is what what's this what do you input at the beginning and these object queries aren't actually random as I said they are learned so what you're going to do is you're going | 1,493 | 1,521 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1493s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | to learn independent of the input image you're going to learn n different object queries and these object queries now it's very it's very interesting because these object queries are sort of going to be different it's like you have different people that can ask the input image different questions right and this they have so there n is 100 but they show 20 of these object queries that | 1,521 | 1,557 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1521s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | they learn and so did they have visualization of all bounding box predictions on all images so it's it's sort of like you have n different people at your disposal and you train these n different people to kind of ask different questions of the input image ok you say this person up here will always ask irrespective of what the input image is will always ask sort of | 1,557 | 1,584 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1557s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | hey input image what's what's on your bottom left right that's I'm really interested what's on your bottom left and sometimes I'm a bit interested in what's here but I'm mainly interested what's on the bottom left of the image whereas this person right here sorry this person right here is more interested in what's in the center that the different colors here is | 1,584 | 1,608 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1584s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | refer to different sizes of bounding boxes so this person is also interested so the person on the top-left is interested mainly in I think small bounding boxes that are on the bottom left and the person here is mostly interested in what I'm really interested what's in the center what's large in the center I want give me large things that are in the center right and then this | 1,608 | 1,636 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1608s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | person right here is really interested on stuff that's on the right side of the image so you see in order to get different sort of a difference in bounding box predictions you train n different people to ask different questions of the of the input image and this asking of questions is exactly what an attention mechanism is so this person right here let's let's take this this | 1,636 | 1,666 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1636s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | person and I'm saying person these are vectors these are learned object queries but this person first they will simply ask the question what's on what's on the right side and then the image features right I'm getting poor drawing the image features it will have an attention mechanism to this part of the image features and then it will get back some signal right and then it will transform | 1,666 | 1,697 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1666s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | that with its own signal up and then it will ask maybe again okay now that I know more because you see that person is interested in multiple things it's interested in those things and those things so at first it will focus on these things but then it says oh now I'm now I know more right there is there I know I see there is actually something on the right side so in the higher | 1,697 | 1,721 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1697s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | layers it can then go back and ask the image more questions by sending these cue vectors of the attention mechanism and it will get back the V vectors from the image features that correspond to these cue things so up and up the layers this person can ask more refined questions about what that particular person is interested in okay and since you have the different people here that | 1,721 | 1,748 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1721s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | ask different questions you basically learn the people in a way such that across the data set they all together they cover every possible image pretty well again these people what they're interested in initially is not dependent on the picture you simply learn this in a global manner all right this is the best way I have of describing it you basically learn n people that are each | 1,748 | 1,776 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1748s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | one is interested in different things different classes and different regions in the image and each one of these people is going to output their best guess of what is where based on what they're interested in so that person might say I'm you know I'm the person that's interested kind of in the left side of things so I am going to output that there is a bird right here now | 1,776 | 1,802 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1776s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | these people if this is a transformer right and everything can attend to everything they can actually communicate with each other as they incorporate information from the image so in each layer they can do both they can incorporate information from the image and they can communicate with each other and then in the next layer that can do it again and again and again and thereby | 1,802 | 1,825 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1802s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | they can sort of they can sort of say well you already got the left side I will take the right side you already got the bird class I will take the elephant class and so on so you see here how the the architecture of the transformer actually is also very conducive to doing this bounding box prediction in that these different things can sort of attend to each other and therefore | 1,825 | 1,851 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1825s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | communicate with each other all right I hope that sort of makes sense now before we get into the experiments I want to list a third reason of why the transformer especially the encoders might actually also make a giant amount of sense here since you on the image into height and width and you have to imagine what does the transformer do the transformer as we said here has this notion of a tension | 1,851 | 1,882 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1851s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | where from any point in the sequence it can gather information from any other point in the sequence and this that's usually one of the downsides of the Transformers is done via a quadratic attention mechanism so if I just list one feature channel go over here if I just list one feature Channel right here this is height times width of the image right this is this is the entire image | 1,882 | 1,910 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1882s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | unrolled in one vector height times width and here I unroll it again height times width then this this matrix that I can build right here which is called the attention matrix right here it will tell me which parts of the sequence attends to which other parts okay so if you have an image that has the let's say the number three and you really want to figure out whether or not this is a | 1,910 | 1,941 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1910s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | three then the bow up here must communicate with the bow down here right they need to share information you say oh there's a bow here there's a bow here and there is a spiky thing here that must be a three so you want something this is rather at the beginning of the sequence you want this to attend first of all it will attend itself so you get fairly high values along the diagonal | 1,941 | 1,964 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1941s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | maybe 1010 1011 1112 and I saw this all eg skated a hundred million nine nine but it also like this this part here at the beginning of the sequence let's say it's here because this is unrolled right needs to attend to the end so this needs to attend to the end which we will put an 11 here and the other way around doesn't always need to be symmetrical by the way | 1,964 | 1,992 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1964s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | okay but in any case this is going to be a H times W squared matrix because everything can attend to everything and that's the attention mechanism why do I think this is so good for bounding boxes because let's let's imagine you actually have a matrix that is like this okay height times width times height times width every single point in here actually defines a | 1,992 | 2,019 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=1992s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | bounding box because this point this point right here in this dimension corresponds to one location in the image and on this axis it corresponds to another location now in the attention matrix simply means these two points need to communicate but if you have two pixels you actually have defined a bounding box right here right you you were actually you're defining a bounding | 2,019 | 2,045 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=2019s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | box and the fact that this is happening in the exact same matrices could mean that the Transformers are uniquely well the Transformers across sequences of these height times with unrolled images are uniquely well conducive to these bounding box prediction tasks I'm actually a bit astounded because when I first just read the title this immediately popped to my mind I'm like | 2,045 | 2,074 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=2045s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | oh yes of course and they're going to predict the bounding boxes by simply training so what you would do what I thought this was gonna be as out you output an actual matrix like this and then you simply each point you can you can simply classify right so you can classify here whether whether or not like at in this direction there is a bird right and then if you have two | 2,074 | 2,099 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=2074s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | points like this for example you and you also classify whether in this direction there is a bird right and this naturally defines a bounding box or you could like take this matrix and actually just classify individual points in this matrix to be the bounding boxes because they already define bounding boxes so I just I think these these quadratic things are are uniquely I mean someone | 2,099 | 2,125 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=2099s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | must have thought of this or if not like the YouTube channel it would be funny first paper ever to actually have to cite the YouTube channel but again yeah so transformers seem to be a good idea for these kinds of things so how do they do of course they do well they are on par where with these other much much much more complex architectures these faster our CNN models they are apparently much | 2,125 | 2,155 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=2125s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | more complex but they are on par with this they do however train forever I think they train for like six days on eight GPUs is not that much if you compare to like language models on hundreds of TP use but still okay I don't want to go into the numbers of experiments but what is really cool is that they can now visualize this sort of attention and you can see right here | 2,155 | 2,182 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=2155s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | that if they look at a particular point in the image and visualize the attention it will actually attend to the instance itself so it will like these are usually the problems for these detection algorithms when things overlap and are partially occluded but you can see right here that the attention is on the part of the image that makes the instance in the back and the attention here is on | 2,182 | 2,207 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=2182s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | the part of this and it doesn't sort of overlap into the others so that is one thing that's pretty impressive about these architectures the other thing they show is for example it can generalize to many many instances so here it has never seen 24 giraffes in one image but yet it can absolutely do that and giraffe giraffe to rupture after of and the one of the coolest images I find are these | 2,207 | 2,239 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=2207s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | here where you can see right here again attention visualisation and you see that even within the bounding box of the front elephant here you see that the attention on this foot of the back elephant is is is assigned to this blue bounding box so this is the blue basically the blue bounding box person that is attending to that back foot that means they they these things really sort of | 2,239 | 2,273 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=2239s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | understand or they learn these things like occlusion and you know just hard I have a hard time describing it but you can see it visually here right like how it clearly learns that these are two instances that are sort of occluding each other but this this this instance can actually appear within the bounding box of the other instance and the same goes for the zebra here that are | 2,273 | 2,302 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=2273s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | partially occluding each other and you can see that the attention is correctly like even here this back foot of this zebra is correctly labeled so all in all that is pretty cool and they take it a step further and they say well with this architecture we can actually pretty easily do pixel wise classification so this is this cocoa stuff and things data set where I don't know which one is the | 2,302 | 2,334 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=2302s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | stuff and which one is the things I think things is the objects and stuff is like sky and mountains and so on and so this is a classification task where you actually have to label every single pixel so what they do is they simply input this through their detector and they detect the instances they take the attention maps of the instances and then they scale it up this right here is just | 2,334 | 2,359 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=2334s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | a CNN sort of in Reverse that scales up the image because they have scaled it down as we said they scale it up again and then they scan simply classify each pixel where each of these you remember we had these different people here that it that cared about different things in the image each of these people will classify their respective pixels the pixels they feel responsible for and | 2,359 | 2,387 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=2359s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | then you simply merge all of these people's predictions together into this prediction and again this gives pretty pretty impressive results I am I mean this is this is fun this looks like it sort of works I haven't they do quantitative analysis of course but I'm just impressed by the examples right here alright that was sort of it I really enjoyed reading this papers the | 2,387 | 2,417 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=2387s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
T35ba_VXkMY | simplicity is pretty cool they do have not only do they have code in the paper to show how ridiculously easy it is to get this to run this is all you need in pi torch but they do actually have code and as I understand they also have pre trained models so they have this model Zoo right here where they give you the pre trained models so you can play with it and you can even load it from torch | 2,417 | 2,443 | https://www.youtube.com/watch?v=T35ba_VXkMY&t=2417s | DETR: End-to-End Object Detection with Transformers (Paper Explained) | |
Nj2YSLPn6OY | let me introduce you to mrs. mani she came to the emergency room she searched his fifty two-years-old she came to the emergency room with a foot sore doctors investigated of with saw and she ended up staying there in the hospital for 22 days here's what happened when she came to the emergency room for a foot sore they inspected her they saw no real reason for medical concern but they | 0 | 38 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=0s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | wanted to monitor in case her foot sore was infected so they put her in the general ward on day three she starts developing symptoms of what looks like mild pneumonia they give her the usual treatment of antibiotics and all's good but then her condition starts to worsen on day six she develops what's called tachycardia that means in medical speak her heart rhythm has accelerated | 38 | 68 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=38s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | dramatically she then has trouble breathing on day seven she experiences septic shock that means her body is in crisis incidentally mortality in shock is one in two now it's only at this point that the doctors get really concerned and they transfer her to the intensive care unit I see use other units where the most critically ill patients get cared for they here they give her every possible | 68 | 99 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=68s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | treatment to stabilize her but her condition only worsens first her kidneys start to fail then her lungs fail and on day 22 she dies mrs. mani did receive the right set of treatments the problem is she received them only too late what mrs. mani experienced was an infection that turned into sepsis let me tell you a little bit about what substance is sepsis occurs when infection releases | 99 | 132 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=99s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | chemicals in your blood to tackle the infection so your body releases chemicals to fight the infection now this chemical can trigger an egg of inflammatory response when this in when this inflammation triggers this negative inflammatory response what it can then do is cause a cascade of changes leading your organs to fail leading to depth sepsis is the 11th leading cause of death more than breast | 132 | 166 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=132s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | cancer and prostate cancer combined turns out substance is preventable if treated early okay so then what's the catch doctors find it very hard to recognize sepsis in fact a Harvard study shows with 93 leading academic experts that when they were given several cases of patients with and without sepsis they couldn't agree two years ago my nephew he was admitted to the best Hospital in | 166 | 197 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=166s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | India and he died of sepsis my family was devastated I'm a machine learning expert and what I do is study ways in which we can use large messy datasets to enable intelligent decision-making so natural question for me was could machine learning of health could machine learning of help mrs. Manny and my nephew so this led to a massive effort with my colleagues at Hopkins to design | 197 | 225 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=197s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | what we call the targeted real-time early warning system or trues based on machine learning I'll give you a sneak peek into what chooses and how we're using it to tackle sepsis let me take a step back and tell you a little bit about what machine learning is and what's AI artificial intelligence is a field of study very design very teach computers how to learn okay just like | 225 | 252 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=225s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | you teach your kids machine learning is one way of doing this by designing code or programs that teach computers stuff over time by by interacting with the environment or watching okay so I'm going to show you a video of some robots learning how to walk I find it funny how it shudders so you're probably now are thinking this is hopeless well so the question is how can | 252 | 298 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=252s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | we teach robots or machines how to walk intuitively you can think of it as designing a game the goal of the game is for the computer or the robot to learn how to walk for as long as possible without following ok so to do this first we have to design write down the goal in a language the computer understands for this we'll use math okay so now you're wondering well how do we write the goal | 298 | 328 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=298s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | of walking without falling as long as possible in math well that's often hard for different tasks but you can think of it as writing down a formula and what this formula does is it scores so in the case of walking it'll score every move the robot makes if the move it makes helps the robot walk it gets a high score if the move that the robot speaking makes the robot unstable it | 328 | 356 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=328s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | gets a low score and now the robots goal is to experiment with the sequence of moves in order to be able to maximize its score so how does it know which moves to try right well there are two strategies for doing it first it expected learns by interacting with the environment okay so here the robot will just make a guess it guesses it makes a move if the move gets a high score | 356 | 384 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=356s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | that's positive feedback and the robot builds on it okay the second strategy is by watching other robots in other words the robot finds data from past robots that are similar to this robot it watches what moves that robot did when it was in very similar positions and now it emulates or replicates those moves okay so those are the two strategies so I'm going to show you a video of robot | 384 | 411 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=384s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | learning how to walk using the strategy I just described okay so in the beginning it's going to look hopeless but I promise you it gets better and just to be clear this is not so this is the skeleton off the robot and so this is not a human animator going there and just moving or animating this video this is really the robot the algorithm choosing which moves to make by moving | 411 | 437 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=411s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | the joints of the skeleton that you're seeing and you can see it's already getting better now suddenly the robots a will blue walk and run for a lot longer than it was doing right so essentially the basic principle is as follows you figure out a game that the computer can play you write it down using a language it understands and then we train it to optimise the score right this is how we | 437 | 464 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=437s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | teach cars how to drive computers how to play the game of go an Alexa to understand say your preference of coconut water so let's go back in our case stir the problem of sepsis so the goal here is to identify sepsis as quickly as possible right and for this truths learns by watching in other words using data from past patients this avoids the need for tools to have to | 464 | 493 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=464s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | experiment on new patients right so to do that what are the pieces truth needs to do so one big change that has happened in medicine that's interesting to note is in the past five years the introduction of electronic health records in EHRs every single measurement every single lab test that is ever done when you walk into the clinic or you're in the hospital gets collected trues | 493 | 518 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=493s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | analyzes this data from thousands of patients to identify subtle signs and symptoms that appear in patients with sepsis than those without okay but that's not alone what truths also needs to do is to figure out how to think about every signal in the context of every other signal let me give you an example let's look at the example of creatinine in SOL creatinine is a waste molecule | 518 | 545 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=518s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | okay and your kidneys filter it out okay but here's the catch so when your body is septic it affects your kidneys it deteriorates your kidneys ability to filter out creatinine so creatinine level rises but there are many other things that can affect your kidneys ability to filter out creatinine for example if you have chronic kidney disease you're very likely to have high | 545 | 573 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=545s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | creatinine levels so now what truth has to do is to figure out is your creatinine high because of sepsis or because of chronic kidney disease or the numerous other factors that need to high creatinine levels but that's not enough it needs to do this for every single signal that exists in the electronic health record and chooses think about every signal in the context of every | 573 | 598 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=573s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | other signal to identify signs and symptoms that occur more often in patients with sepsis than those without let's return to mrs. Manny research by Kumar and colleagues have shown that for every hour treatment is delayed mortality goes up by seven to eight percent so timing is critical we went and took mrs. Manny's data and we ran throughs on it and here's what we found | 598 | 627 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=598s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | truth would have detected mrs. Manny's sepsis 12 hours before doctors currently did as my clinical colleagues would say that is the difference between life and death last year we showed using data from 16,000 patients that throughs on average would have detected on most patients on average 24 more than 24 hours prior to the shock onset that's not 24 hours in two-thirds of these | 627 | 661 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=627s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | patients their sepsis was detected prior to any organ dysfunction whatsoever and to put this result in context that's 60% increase in performance over state of the art so what truth is really doing is doctors a much longer window to come in and intervene in order to prevent organ dysfunction and mortality this year the independently validated truths in data from Howard County | 661 | 690 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=661s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | General Hospital in Maryland and now we're working to do real-time integration in order to make something like truths available to every doctor at Hopkins I'm also really excited because after we've published our papers several other health systems are now already implementing the published version of truths in order to be able to develop it in their own environment so I'm going to | 690 | 715 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=690s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | highlight like a few perhaps three salient characteristics that I think makes a strategy like truths very powerful ok first truths runs 24/7 what it does is it gives doctors a second pair of reliable eyes right - it's hard to scale-up doctors it's easier I think much easier to scale up computers and what truth is really doing is allowing us to get expertise from the best | 715 | 747 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=715s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | doctors everywhere here's the third one which i think is very interesting in many cases like we see in substance we might not need new measurements the signs and symptoms were already in your data and what truth is really doing is discovering these signs and symptoms to learn something that we couldn't see by eye finally there's been a lot of buzz about big data and I want to make a | 747 | 777 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=747s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | little subtle point about a technical problem that I think truces solving that is very interesting it truths would be able to learn learn much faster if it had a lot of data on you or it could get more data by experimenting on you but we don't want that right so what truths really has to do is leverage your limited data to figure out what's right for you right so in other words what | 777 | 805 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=777s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | truths really has to solve ism is a challenging small data problem in other words it has limited data on you and has to figure out what is the right treatment for you and for that it has to let it leverages vast amounts of data from other patients and figures out what information to borrow in order to make these assessments reliably and precisely so I also want to tell you a | 805 | 833 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=805s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | little bit about how the strategy is not unique to sepsis so very broadly if you think about it in many diseases essentially where you have profile of symptoms and the response to treatments varies a great deal across individuals you can use the strategy like truth in order to target treatment so you're wondering like for example if you consider cancer diabetes multiple | 833 | 859 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=833s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | sclerosis Parkinson's lupus so there are many such diseases on which a strategy like Truths is amenable in fact in our own lab with experts in rheumatic diseases or immune diseases in particular we're looking at how in scleroderma for instance we can use strategies similar to tools to avoid giving strong immunosuppressants to patients who don't need them other | 859 | 887 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=859s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | colleagues this is William Pelham Susan Murphy and their team they're studying kids with ADHD and looking at how using similar data-driven strategies they can identify when kids can be benefit from behavioral therapy and we can avoid the need for giving them psychostimulants altogether so the strategy is very very powerful so I was speaking about substance so let's go back to substance | 887 | 912 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=887s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | again so I said it was sepsis Awareness Month and the CDC has declared substance to be a medical emergency rightfully so remember 750,000 people annually are affected by sepsis a patient's family recently asked me what will it take to bring this to a hospital near us I think that can be done in fact it can even be done within a year but we don't want to stop there we want it to be possible to | 912 | 942 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=912s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | bring strategy like truth - hospitals everywhere and so the question is to do that what will it take right so I think the three key things we need your help for one we need super smart engineers to be working in healthcare we need your help in building and scaling up such technologies don't go to wall street healthcare needs you right we need policymakers to create | 942 | 973 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=942s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | incentives to open up electronic medical records as an expert at a leading health institution it's taken me more than a year because the EMR is so closed in order to be able to figure out how to implement rules against the EMR it really should be easier than this three we need a healthcare system that's based on quality our current healthcare system is incentivized to optimize volume | 973 | 1,006 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=973s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | rather than quality right now you can choose which restaurants to go to based on the quality of food should you be able to choose the hospitals you go to based on quality of care part of the problem is that quality data at the moment is not very visible to consumers and we really need to make a bigger effort to make this quality a visible so that you can choose based on quality so | 1,006 | 1,029 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=1006s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
Nj2YSLPn6OY | to summarize sepsis is one preventable killer in many pressing medical problems like we saw in sepsis the answers for knowing whom to treat when to treat and what to treat with might only be in your data sometimes I wonder if we had done this work to use earlier if I could have prevented my nephew Nick induced depth I can't wait for this to be the way medicine is practiced thank you | 1,029 | 1,070 | https://www.youtube.com/watch?v=Nj2YSLPn6OY&t=1029s | Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston | |
SaJL4SLfrcY | all right so I'll talk these two parts to my talk about three parts so one is about a little bit of kind of state of the art of supervisor on you in France Maroni and and then the second part about supervised running and that's the title really of the talk and with an introduction to something I call energy based running which is sort of a way of sort of a general framework or paradigm | 0 | 24 | https://www.youtube.com/watch?v=SaJL4SLfrcY&t=0s | Self-Supervised Learning | |
SaJL4SLfrcY | if you want to approach to approach learning in general should I use this all right much better oh this is just for recording I guess okay so we all know what supervised running is about I'm told you all know what supervised learning is about and this is the situation where you train a machine by telling it where the correct answer is for a bunch of training samples and this | 24 | 50 | https://www.youtube.com/watch?v=SaJL4SLfrcY&t=24s | Self-Supervised Learning | |
SaJL4SLfrcY | works really well if you have lots of data it works for image recognition translation natural language processing speech recognition you're all kinds of applications but those are applications where the economics are such that it's worth actually labeling a lot of data by hand and and of course you know with that context machine learning is basically comes down to finding a good | 50 | 75 | https://www.youtube.com/watch?v=SaJL4SLfrcY&t=50s | Self-Supervised Learning |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.