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
|
|---|---|---|---|---|---|---|
SoODZ7tEN5Q | fuel ideas ideas more ideas and action chapter 8 belief makes things happen I have tried to make plain how this power through belief can be developed and to take you up the ladder as far as you wish to go it is necessary though to point out that it is easy to lose ones belief or faith thousands have risen to great heights of success only to stumble row or fall to undreamed-of depths | 3,145 | 3,188 | https://www.youtube.com/watch?v=SoODZ7tEN5Q&t=3145s | "The Magic of Believing" By Claude Bristol | |
SoODZ7tEN5Q | others seeking health have appeared to be more or less miraculously cured only to find that in later years or even months there is a recurrence of their ailments there are many weakening factors and influences all suggestive in nature which we in unguarded moments allowed to slip into our subconscious minds once these influences begin the destructive work they can undo all the good | 3,188 | 3,216 | https://www.youtube.com/watch?v=SoODZ7tEN5Q&t=3188s | "The Magic of Believing" By Claude Bristol | |
SoODZ7tEN5Q | accomplished by our constructive forces so step out in front head toward the Sun keep facing it and the dark shadows will not cross your path I know that it is difficult for the average person who knows nothing of the subject to accept the idea that all is within but surely the most materialistic person must realize that as far as he himself is concerned nothing exists on the outside | 3,216 | 3,246 | https://www.youtube.com/watch?v=SoODZ7tEN5Q&t=3216s | "The Magic of Believing" By Claude Bristol | |
SoODZ7tEN5Q | plane unless he has knowledge of it or unless it becomes fixed in his consciousness it is the image created in his mind that gives reality to the world outside of him happiness sought by many and found by few therefore is a matter entirely within ourselves our environment and the everyday happenings of life have absolutely no effect on our happiness except as we permit mental | 3,246 | 3,275 | https://www.youtube.com/watch?v=SoODZ7tEN5Q&t=3246s | "The Magic of Believing" By Claude Bristol | |
SoODZ7tEN5Q | images of the outside to enter our consciousness happiness is wholly independent of position wealth or material possessions it is a state of mind which we ourselves have the power to control and that control lies with our thinking Emerson said what is the hardest task in the world to think obviously this is so when one considers that most of us are victims of | 3,275 | 3,305 | https://www.youtube.com/watch?v=SoODZ7tEN5Q&t=3275s | "The Magic of Believing" By Claude Bristol | |
SoODZ7tEN5Q | mass thinking and feed upon suggestions from others we all know that the law of cause and effect is inviolable yet how many of us ever pause to consider its workings the entire course of a man's life has many times been changed by a single thought which coming to him in a flash became a mighty power that altered the whole current of human events history is replete with the stories of | 3,305 | 3,332 | https://www.youtube.com/watch?v=SoODZ7tEN5Q&t=3305s | "The Magic of Believing" By Claude Bristol | |
SoODZ7tEN5Q | strong-minded resolutely willed individuals who steadfastly holding to their inner convictions have been able to inspire their fellow man and in the face of tremendous and determined opposition have literally created out of nothing great businesses huge empires and new worlds they had no monopoly of thought power you and every man and woman have it all you have to do is use | 3,332 | 3,359 | https://www.youtube.com/watch?v=SoODZ7tEN5Q&t=3332s | "The Magic of Believing" By Claude Bristol | |
SoODZ7tEN5Q | it you will then become the person you envisage in your imagination know yourself know your power faithfully use the cards and the mirror techniques and you will get results far beyond your fondest expectations just believe that there is a genuine creative magic in believing and magic there will be for belief will supply the power which will enable you to succeed in everything you | 3,359 | 3,389 | https://www.youtube.com/watch?v=SoODZ7tEN5Q&t=3359s | "The Magic of Believing" By Claude Bristol | |
1DYmgoij4FQ | i have read literally thousands of books on modern psychology metaphysics ancient magic buddhism yogism theosophy christian science unity truth new thought and many other dealings it's what i call mind stuff many of these books were nonsensical others strange and many very profound gradually i discovered that there is a golden thread that runs through all the | 0 | 29 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=0s | The Secret Knowledge Of Believing | |
1DYmgoij4FQ | teachings and makes them work for those who sincerely accept and apply them that thread can be named in a single word belief it is the same element or factor belief which causes people to be cured through mental healing enables others to climb the ladder of success and gets phenomenal results for all who accept it why belief as a miracle worker is something that cannot be satisfactorily | 29 | 57 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=29s | The Secret Knowledge Of Believing | |
1DYmgoij4FQ | explained but have no doubt about it there's genuine magic in believing the magic of believing became a phrase around which my thoughts steadily revolved i've tried to put down these thoughts as simply and as clearly as i could so that everyone can understand my hope is that anyone who listens will be helped in reaching their goal in life so you begin with desire | 57 | 87 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=57s | The Secret Knowledge Of Believing | |
1DYmgoij4FQ | if you ever hope to achieve anything or gain more than you have now however as we shall see there is more to it than mere desire it has been said that thought attracts that upon which it is directed thought attracts that upon which it is directed it was job who said for the thing which i greatly feared has come upon me our fearful thoughts are just as creative or just as magnetic and | 87 | 119 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=87s | The Secret Knowledge Of Believing | |
1DYmgoij4FQ | attracting troubles to us as are the constructive and positive ones and attracting positive results so no matter what the character of the thought it does create after its kind when this sinks into a man's consciousness he gets some inkling of the awe-inspiring power which is his to use i cling to the theory that while thoughts do create an exercise control far beyond | 119 | 145 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=119s | The Secret Knowledge Of Believing | |
1DYmgoij4FQ | any limits yet known to man they create only according to their pitch intensity emotional quality depth of feeling or vibratory plane in other words comparable to the wavelength and wattage of a radio station thoughts have a creative or controlling force in the exact ratio of their constancy intensity and power let me try to clarify that while many explanations have been | 145 | 176 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=145s | The Secret Knowledge Of Believing | |
1DYmgoij4FQ | offered no one knows whether thought is a form of electrical energy or something else yet to be defined but i have been an experimenter in that branch of electricity known as high frequency pioneered by the great genius nikola tesla and whenever i think of thought and its radiations and vibrations i instinctively link them up with electricity and its phenomena | 176 | 201 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=176s | The Secret Knowledge Of Believing | |
1DYmgoij4FQ | in this manner they become more understandable to me all persons living in high altitudes have felt and sometimes observed the electric spark resulting from walking across the room then touching some metallic substance that of course is a form of static electricity generated by friction it gives you an idea of how one kind of electricity can be developed through the | 201 | 224 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=201s | The Secret Knowledge Of Believing | |
1DYmgoij4FQ | body sigmund freud the famous austrian psychoanalyst brought the world's attention to the hypothesis that there was a powerful force within us an unilluminated part of the mind separate from the conscious mind constantly at work molding our thoughts feelings and actions others have called this division of our mental existence the soul some call it the super ego the inner | 224 | 251 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=224s | The Secret Knowledge Of Believing | |
1DYmgoij4FQ | power the super consciousness the unconscious the subconscious and various other names it isn't an organ or so-called physical matter such as we know the brain to be nevertheless it is there and from the beginning of recorded time man has known that it exists the ancients often referred to it as the spirit paracelsus called it the will others have called it the mind an adjunct to | 251 | 277 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=251s | The Secret Knowledge Of Believing | |
1DYmgoij4FQ | the brain some have referred to it as conscience the creator of the still small voice within still others called it intelligence and have asserted that it is a part of the supreme intelligence to which we are all linked no matter what we call it i prefer the word subconscious it is recognized as the essence of life and the limits of its powers are unknown it never sleeps | 277 | 304 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=277s | The Secret Knowledge Of Believing | |
1DYmgoij4FQ | it comes to our support in times of great trouble it warns us of impending danger often it aids us in what seems impossible it guides us in many ways and when properly employed performs so-called miracles perhaps the most effective method of bringing the subconscious into practical action is through the process of making mental pictures using the imagination perfecting an | 304 | 331 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=304s | The Secret Knowledge Of Believing | |
1DYmgoij4FQ | image of the thing or situation as you would have it exist in physical form this is usually referred to as visualization however before this visualization can work you must really believe i refer now to deep-seated belief a firm and positive conviction that goes through every fiber of your being when you believe it heart and soul as the saying goes now call it a phase of emotion a | 331 | 362 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=331s | The Secret Knowledge Of Believing | |
1DYmgoij4FQ | spiritual force a type of electrical vibration anything you please but that's the force that brings outstanding results it sets the law of attraction into operation enables sustained thought to correlate with its object this belief changes the tempo of the mind or thought frequency and like a huge magnet draws the subconscious forces into play changing your whole aura and affecting | 362 | 388 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=362s | The Secret Knowledge Of Believing | |
1DYmgoij4FQ | everything about you and often people and objects at great distances it brings into your individual sphere of life results that are sometimes startling after studying the various mystical religions and different teachings and systems of mind stuff one is impressed with the fact that they all have the same basic modus operandi and that is through repetition the repeating of certain mantras words | 388 | 418 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=388s | The Secret Knowledge Of Believing | |
1DYmgoij4FQ | formulas or just plain mumble jumbo is common with witch doctors voodoo high priest hexers and many other followers of strange cults they use them to evoke the spirits or work black magic one finds the same principle at work and chants incantations litanies daily lessons also the frequent praying of the buddhists and muslims alike the affirmation of the theosophists and | 418 | 442 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=418s | The Secret Knowledge Of Believing | |
1DYmgoij4FQ | the followers of unity the absolute truth new thought divine science in fact it is basic to all religions although here it is white magic instead of black magic this brings us to the law of suggestion through which all forces operating within its limits are capable of producing phenomenal results that is it is the power of suggestion and auto suggestion your own to yourself or hetero | 442 | 470 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=442s | The Secret Knowledge Of Believing | |
1DYmgoij4FQ | suggestion coming to you from outside sources that starts the machinery into operation or causes the subconscious mind to begin its creative work and right here is where the affirmations and repetitions play their part it's the repetition of the same chant the same incantation the same affirmations that lead to belief and once that belief becomes a deep conviction | 470 | 497 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=470s | The Secret Knowledge Of Believing | |
1DYmgoij4FQ | things begin to happen this is the same identical force and the same mechanics that hitler used in building up the german people to attack the world a reading of mein kampf will verify that dr renee favel a famous french psychologist explained it by saying that hitler had a remarkable understanding of the law of suggestion and its different forms of application | 497 | 523 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=497s | The Secret Knowledge Of Believing | |
1DYmgoij4FQ | it was with uncanny skill and masterly showmanship that he mobilized every instrument of propaganda in his mighty campaign of suggestion hitler openly stated that the psychology of suggestion was a terrible weapon in the hands of anyone who knew how to use it let's see how he worked it to make the germans believe what he wanted them to and once that belief took hold how they | 523 | 547 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=523s | The Secret Knowledge Of Believing | |
1DYmgoij4FQ | started their campaign of terror slogans huge signs posters masked flags appeared throughout germany hitler's picture was everywhere one reich one folk one leader became the chant it was heard everywhere today we own germany tomorrow the entire world the marching song of the german youths came from thousands of throats daily such slogans as germany has waited long | 547 | 577 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=547s | The Secret Knowledge Of Believing | |
1DYmgoij4FQ | enough stand up you are the aristocrats of the third reich germany is behind hitler to a man and hundreds others bombarded the people 24 hours a day from billboards sides of buildings the radio and the press every time they move turned around or spoke to one another they got the idea that they were a superior race and under the hypnotic influence of this belief | 577 | 603 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=577s | The Secret Knowledge Of Believing | |
1DYmgoij4FQ | strengthened by repeated suggestion they started out to prove it unfortunately for them there were other nations who also had strong national beliefs that eventually became the means of bringing defeat to the germans i know that it is difficult for the average person who knows nothing of the subject to accept the idea that all is within but surely the most materialistic person must realize that | 603 | 631 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=603s | The Secret Knowledge Of Believing | |
1DYmgoij4FQ | as far as he himself is concerned nothing exists on the outside plane unless he has knowledge of it or unless it becomes fixed in his consciousness it is the image created in his mind that gives reality to the world outside of him happiness sought by many and found by few therefore is a matter entirely within ourselves our environment and the everyday happenings of life have absolutely no | 631 | 661 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=631s | The Secret Knowledge Of Believing | |
1DYmgoij4FQ | effect on our happiness except as we permit mental images of the outside to enter our consciousness happiness is wholly independent of position wealth or material possessions it is a state of mind which we ourselves have the power to control and that control lies with our thinking emerson said what is the hardest task in the world to think obviously this is so | 661 | 693 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=661s | The Secret Knowledge Of Believing | |
1DYmgoij4FQ | when one considers that most of us are victims of mass thinking and feed upon suggestions from others we all know that the law of cause and effect is inviolable yet how many of us ever pause to consider its workings the entire course of a man's life has many times been changed by a single thought which coming to him in a flash became a mighty power that altered the whole | 693 | 717 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=693s | The Secret Knowledge Of Believing | |
1DYmgoij4FQ | current of human events history is replete with the stories of strong-minded resolutely willed individuals who steadfastly holding to their inner convictions have been able to inspire their fellow man and in the face of tremendous and determined opposition have literally created out of nothing great businesses huge empires and new worlds they had no monopoly of thought power | 717 | 744 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=717s | The Secret Knowledge Of Believing | |
1DYmgoij4FQ | you and every man and woman have it all you have to do is use it you will then become the person you envisage in your imagination know yourself know your power faithfully use the cards in the mirror techniques and you will get results far beyond your fondest expectations just believe that there is a genuine creative magic in believing and magic there will be | 744 | 774 | https://www.youtube.com/watch?v=1DYmgoij4FQ&t=744s | The Secret Knowledge Of Believing | |
_V-WpE8cmpc | i-it's it's a wonderful to be here I have been remiss in that I have not been to Prague for for a decade so it was it's wonderful to be back in Prague and it's wonderful to be in this fancy new Institute and so I think because because there is various types of people here there are some vision people some graphics people and some some others that learning done learning I'm going to | 0 | 31 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=0s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | give you kind of a an overview of some of the stuff that we have been doing but don't go into too much detail because that might be too boring for others and and and maybe too much for a foursome but then I'm I'm here the rest of the day so I'm happy to to chat about these things more so of course you know how it is with with with with with professors you know all the work has done by the | 31 | 62 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=31s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | graduate students and the postdocs who are amazing and then the professor just pushed together in the slides and and and and presents the work and in this case actually even the slides most of them have been done by their by the Graduate since so I'm really just just a audio recording of really the all the wonderful work that that they have been doing so probably if you didn't live | 62 | 87 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=62s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | under a rock for the last few years you have heard of the deep networks and how they have revolutionised computer vision and kind of the standard classic way of doing this is it's basically a classic supervised learning problem you are giving a network which you can think of as a big black box a pairs of input images and output labels XY pairs okay and this big black box essentially you | 87 | 119 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=87s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | can think of it as memorizing these these the score currents or the the its memory its modeling the associations between the XS and the Y's okay and of course you need to have lot and lots and lots of these training pairs so you have lots of people clicking on a bunch of imager a lots of you know millions of images and what do they what is their labels and once you | 119 | 147 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=119s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | have trained on millions of these pairs of images and labels then given a new image this magic black box can tell you what label it is and this is what supervised direct supervised learning and this particularly deep learning has been all about ok but there are some problems that that I'll mention of this beautiful story one obvious problem is that this this labeling bit is very | 147 | 179 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=147s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | expensive millions of images don't come cheap you have to have people actually label them and for every new problem we need to label more images so that's that's that's a problem of cost but there is also another problem that is a little bit more subtle and here is an illustration of this so here is an image that is basically you can think of it as a texture synthesized version of the | 179 | 204 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=179s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | original image you basically some of the pixels have been moved around in certain ways but to kind of to preserve the kind of the statistics of that of the image it's it's it's actually work by by Leon Gattis ok so we change the input but the neural network is perfectly happy to still call it a collie in fact I can give you other random images like this right and it's still basically happy to | 204 | 236 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=204s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | call it by that same class collie ok and what this suggests is that this magical black box the convolutional neural network is not actually doing that much it's not doing what we think of a lot of computer vision should be all about it's not doing you know figure-ground detection it's not fine the the you know the the foreground region it's not finding the dog it's not | 236 | 267 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=236s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | segmenting the dog out from the background it's not doing any of the foreground background or occlusion reasoning none of that it doesn't need to do any of that because it probably just looks at the snout in a couple of eyes and then says oh yeah that's a collie okay and maybe some his called histogram so it doesn't need to work super hard to solve this problem okay | 267 | 289 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=267s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | and this is you know a cause for worry because in this particular case this is imagenet classification task so you have a thousand classes and so maybe this is not maybe you don't need to work that hard because you might not need to have to to to really worry about oops you might not need to worry about finding the boundaries of objects so even reasoning about objects with only a | 289 | 325 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=289s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | thousand classes but the issue is that we don't really have more than a thousand classes labeled and so in the end what this magical neural network is doing it's not really object detection it's really more like texture classification it's classifying dogcatcher collie texture so this if you know 1,000 weight classification tasks is the only thing you want to do maybe | 325 | 355 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=325s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | this is not so bad but here is an example of something that you know Joseph and and and a lot of us have been worried about for a long time action recognition okay action recognition the same thing but in time so you you have you have a video and you want to recognize what action is being performed and the one very weird result in action classification has been that giving more | 355 | 388 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=355s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | frames of video to the classifier did not seem to improve performance that just a single frame oh thank you perfect the single flame is is good enough all right look at that perfect okay thank you okay so for example so you basically for a single frame you do basically just as well and this was a big strange result that people don't know why it was so but if we look at for example here is an | 388 | 436 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=388s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | opening fridge action so you go to the fridge you extend your hand you pull on the fridge door you open the fridge and then you close it right and you want to recognize other actions that are opening fridge actions okay if you run a classifier for this task you label a whole bunch of opening the fridge actions as positives and then others as negatives you train your network and | 436 | 463 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=436s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | then if you look at what the performance is great by the way the performance is very very good but if you look at which frames did the classifier actually pay attention to it's just this one so it doesn't care it doesn't track the hand it doesn't care about the the fridge door really all it cares about is again it's the texture of an open refrigerator okay and once it sees an open | 463 | 491 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=463s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | refrigerator texture it knows oh this must be a fridge opening action what else could it be right so again it's it's taking the easy way out it's being lazy because it doesn't have to work hard okay and maybe if you asked me a year or two ago you know how do we deal with this problem I would say that this is all an issue of a data set bias that the only pictures of opening | 491 | 519 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=491s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | refrigerators are in this opening fridge action so let's add some negative images of just open fridges like from from you know maybe Amazon product search and then everything will be fine now I'm starting to think that while data set bias is a problem it's not the whole problem because in a sense this data set bias will never go away it's there is no way that the data the | 519 | 548 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=519s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | data is finite so they will never be able to fix all the holes there will always be a way to cheat because if there is a finite amount of data there is always a way to to find a path that that is that is somehow you know cheating through the data and and so it's it's there is it's kind of like playing a you know this game children's game of Wacom all you you you you push | 548 | 576 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=548s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | something down and something else pops up okay and and also if you think if you ask a machine learning people about it this this is not even their problem because the machine learning people say look you train on the training set and then you've evaluated on the validation set over your method or over your data set okay so you take your data set you split it into two the training and the | 576 | 603 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=576s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | test set and as long as it does well on the test set you're fine right and the test set comes from the same distribution as the training set so it's the same statistics what we have here is that we want to test our algorithms on something that's not really in the in the test set of the of the data set that's something else so we train on say detecting cars from from imagenet data | 603 | 629 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=603s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | set but then we want to go out on the street and and and detect cars there and the cars on the street don't really have the same distribution as the cars that you were trained on but we still want to do it so in a sense our problem is actually quite somewhat different than the problem with in machine learning that we actually do want to test on things we never really trained on so we | 629 | 650 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=629s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | wanted to really be general and so the way forward I see is that somehow we need to better use the data we have there is no hope to ever get all the data that we that that that that will make the problem perfectly can concrete but we need to somehow use the data that we have better I'll give you a couple of examples of the way I think about this so in in a you can think of it as as a | 650 | 685 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=650s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | way that a well around country is run compared to a badly run country's run it's not that in a well-run country you cannot cheat the laws of course you can but the system is set up in such a way that it's actually more expensive to cheat than to follow the law okay and so even though you can cheat you don't do it because it's not in your self-interest in a poorly run country | 685 | 716 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=685s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | it's cheaper to cheat so everybody cheats and there is no way you know there is nothing you can do about it okay so we need to somehow set up our problem is such a way that it's more expensive for the network to cheat let the easiest thing the network can do is to do the right thing that's that's the goal okay and if you think about the way that we do this direct supervised learning input | 716 | 741 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=716s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | image output label and just train on these pairs that just sets up sets up your your life for cheating because so in in in in in my class you know we have a whole semester worth of material and then the debt of the semester there is a final exam so of course most people don't do anything during the whole semester the night before the exam they they look at some exams from previous | 741 | 773 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=741s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | years and they try to memorize this you know the question and answer question answer question answer question alright and they basically memorize the whole all of this set of question and answer pairs and then they go to the examine actually they do pretty well right this is not a bad strategy to pass the final exam it's a bad strategy if you actually want to learn | 773 | 796 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=773s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | but it's not the best strategy to pass the exam because you know I'm lazy I'm going to make the exam this year to be not that different from the exam last year in any case there is very small set of problems you can ask that is easy to grade etc etcetera right and so this kind of memorization of question answer question answer it's actually the correct thing to do if your goal is to | 796 | 818 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=796s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | pass the exam but of course our goal is not to pass the exam our goal is to to actually learn the material so how do you learn the material how do you actually learn you know arithmetic for example when you're a little kid to really learn it what you do is you don't get yourself question-answer pairs you look at the question you try really hard to solve it and then once you solve it | 818 | 839 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=818s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | you got some answer you go to the back of the book to compare it with them with the right answer and then that's how you kind of try to update yourself okay and so basically that's the kind of idea that that we want to try to push our computers to do to try to study harder to try to learn things that are more generalizable that are not just good enough to pass the test but to actually | 839 | 867 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=839s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | understand the world okay so that's kind of the the preamble and the the way we have been working on this in in in my lab is we have been doing it in three different paths and I'm just gonna kind of quickly show some of their some of the results of it so the first is self supervision the idea of not having a expert tell you the correct answer but let the computer figure out the | 867 | 899 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=867s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | correct answer second is what we call meta supervision it's I'm not sure how standard this term is I think we might just have came up with it and the idea here is that you don't supervise the data the correct answer you supervise how the answer supposed to behave okay and finally if there is time I also want to mention a little bit about you know what if there is no correct answer what | 899 | 925 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=899s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | if you're just learning by just playing around you know if you don't have a goal there is no there is nothing to cheat there is no need to cheat because you're just playing around there is no goal and so the idea here is to see if we just get removed the goal remove the you know the whatever whatever we're trying to optimize and see if we can just just play and be curious can that get get us | 925 | 949 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=925s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | some representation that's more generalizable okay so I'm gonna show some examples of all of these in the next oops alright okay so first is self supervision here is a evocative drawing by Asher of what we mean here and this is something that actually kind of had been classic in in in in deep machine learning under the heading of representation learning so we basically | 949 | 975 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=949s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | want to somehow have a compact representation of an input image and we want to compute this representation maybe without any labels and the kind of classic way to do this is what's called an out encoder which says let's have a representation that is small okay so there is a bottleneck here but that if we unpack it and compress it we can reconstruct the original input okay and | 975 | 1,004 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=975s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | then you train this kind of a out encoder set up for many many many images but you don't have any don't need any labels here right it's just just pairs of just just a single image and this is this a very influential idea unfortunately it doesn't actually work in practice the representation that you learn here if you're doing it for any kind of real data like a big image for | 1,004 | 1,030 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1004s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | example not a tiny you know third you do way third you image but a big image that doesn't actually work okay and the reason it doesn't work is that this is you can think of it really as data compression right you're compressing your data but and data compression is related to machine learning but it's not quite the same because data compression doesn't care about how do you perform on | 1,030 | 1,054 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1030s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | new images it only cares about how you compress the training images that you got and so what we propose to do is to think about this in terms of not data compression but data prediction to make the computer try to work harder and say instead of just compressing the data let's see if we can train it to predict the data so one way to do this very simple way is you only give it half | 1,054 | 1,081 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1054s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | of the input signal and you train it to predict the other half so now we not just the compression it's not just that you keep taking the pixels and compressing them you need to think a little bit more you need to think about context and what should go well with with the input that you got ok and one very simple way of doing this is to split the data in terms of color and and | 1,081 | 1,108 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1081s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | and luminance so this was our paper a few years back where we said okay let's take let's take a color image separate it into a luminance channel and a chrominance channel and then train a network to predict from the grayscale to the the color and then you know you can get a nice beautiful image but hopefully also you'd learn the representation that is actually meaningful and somehow captures | 1,108 | 1,135 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1108s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | something about the natural world okay so of course you need to show some pictures first so it actually does learn a reasonably good representation of color but the cool thing that suggest that maybe is also learning something else is some of the failures so here is a couple of instructive failures can somebody see what the failure is it's let me what that year the year is a | 1,135 | 1,168 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1135s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | little bit off but it's a but somebody said the tongue do you see the tongue there is no tongue and yet it's it's point it's coloring it pink why could this why would this be well we were confused too but then we looked at the training data and in the training data these poodles they all have their tongues out so if this was just a stupid compression it would not this this air would not | 1,168 | 1,199 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1168s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | happen but here it seems like the network is actually recognizing that this is a dog it's recognizing the breed of the dog it's remembering the the similar dogs that has seen before and then it's making and mistake but it's a reasonable mistake that maybe if all the dogs have had their tongues out in the training set maybe that's also true in a test ok and indeed so if we could you know we | 1,199 | 1,229 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1199s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | did a various tests but I will just show you one way to to see what's what's been learned in this representation is to to do what we call deep net electrophysiologist so we kind of probe the different represent features of that compressed representation and see what they wear with a fire think of Amalek neurons firing so where do they fire and so what we found is that there is a we found a | 1,229 | 1,256 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1229s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | neuron that fires only on faces there another neuron that only finds and fires and dog faces another one that fires on flowers ok so it's basically it was able to kind of disentangle from massive pixels of the input it was able to to find the kind of specialized neurons for different different parts of the visual world even though the labels we had were just the color it was it's a very very | 1,256 | 1,285 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1256s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | weak label and a way label that that that didn't have any semantics in it and yet we are basically getting something out that is semantic okay and so this is this is kind of a hopeful direction it's their presentation is not as good as all the kind of semantic train representations yet but I still feel that it is it is optimistic direction because hopefully it might be more | 1,285 | 1,313 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1285s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | general in in the long run but this is still to be to be to be found out originally actually this concept of self surprise learning one of the early papers ended by psychologist Virginia dasa was on thinking about it intra of the different modalities of of the sensory signal so instead of saying okay Kohler versus grayscale although that that true is kind of biologically | 1,313 | 1,340 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1313s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | plausible you have the rods and the cones and yeah rods in the cones and you can say that the rods in the codes kind of code train each other but it's much more reasonable to think about it in terms of different modalities for example sight and sound okay so the idea here is that you know you learn about cows by seeing the cow hearing the Moo associating those two together and using | 1,340 | 1,369 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1340s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | that as a kind of learning signal okay and so just recently we decided we're going to try to try to try to use self surprise learning in this in this domain and this is one motivation for why this kind of this kind of thing is is is is a need at any one moment we are being bombarded by sensory information our brains do a remarkable job of making sense of it all | 1,369 | 1,402 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1369s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | [Music] it seems easy enough to separate the sounds we hear from the sites we see but there is one illusion the reveals this isn't always the case have a look at this what do you hear ba ba ba bi yes ba ba ba ba but look what happens when we change the picture and yet the sound hasn't changed in every clip you are only ever hearing bar with a bee ah it's an illusion known as | 1,402 | 1,462 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1402s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | the McGurk effect take another look ba concentrate first on the right of the screen ah now to the left of the screen ba the illusion occurs don't because what you are seeing clashes with what you are hearing in the illusion what we see overrides what we hear so the mouth movements we see as we look at a face can actually influence what we believe we're hearing if we close our eyes we | 1,462 | 1,493 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1462s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | actually hear the sound as it is if we open our eyes we actually see how the mouth movements can influence what we're hearing so did the people here I wonder how it works for non-english speaker so so yeah so I think that some some of it is specific to English speaker that the first sound is you you make your mouth like this right and so even though the sound is exactly the same the lips move | 1,493 | 1,527 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1493s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | very differently and so your brain how many people got the the effect how many people heard the five okay very good very good so it seems to be working very good right so this motivates the idea that if you want a representation for example a video representation you to combine the visual and audio and you probably want to combine it pretty early this is this is a very kind of powerful | 1,527 | 1,556 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1527s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | effect even if you know about the effect if you're very well aware you you steer this thing all the time you you cannot turn it off it is very very powerful low level effect and so said it this suggests that the coupling of audio and video probably happens pretty early on and so the idea that we had was to recreate an video representation that takes in audio and | 1,556 | 1,584 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1556s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | visual features at the same time okay so kind of classic video representations you basically have some some way to go from a series of frames to representation and then also the same thing for for audio and what we propose is to to combine it together and then have a joint audio and video representation you know for for a number of layers at the top so that there's | 1,584 | 1,614 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1584s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | this information can kind of a cook together and get us a join out your video representation okay but how do we train this representation and again we want to train it and without supervision it will self supervise training so one thing that was kind of the obvious first idea that we had was why don't we train say a binary classifier where the positives are videos with the correct | 1,614 | 1,641 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1614s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | sound and negatives are videos with a wrong sound of sound from some other video for example okay and then we train this classifier and hopefully it will learn to to figure out that the correct sound correct video correct sound there is a correspondence okay this unfortunately doesn't work well at all and the reason it doesn't work well at all is because again it's a problem of | 1,641 | 1,669 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1641s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | cheating because if you have if you take a random video and a random audio the video could be me giving a talk and the audio could be you know some sort of a you know sound from the restaurant for example right you just look at the picture you look at it but people sitting and listening and you know that this this is not a restaurant so just just to look at overall picture | 1,669 | 1,700 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1669s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | just a single frame will be enough to tell you you know this is a presentation that is a restaurant that is a rock concert so you don't even need to listen to anything you can just have a kind of a 1:1 label which says okay this is a restaurant or a rock concert so we need again to try to make the computer work harder and so here is our also simple idea - and the idea - is that we're | 1,700 | 1,726 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1700s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | going to have as positives again videos with the correct sound but as negatives we're going to have that same video that same audio but we're going to displace it in time a little bit okay now this becomes a much harder problem to solve because it's the router is correct the video is correct the only thing that's not correct now is that there is a little bit of a time lag so this | 1,726 | 1,753 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1726s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | representation really needs to very carefully pay attention to to the connected to the registration between between those two okay so here is the idea so the correct the correct samples are line samples are just you know how would you and video together okay and then and then incorrect once all was the same except couple of seconds is placed and we have to careful because because if you displace | 1,753 | 1,793 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1753s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | for maybe like a second or less than a second humans are actually not even that sensitive to it so it needs to be a little bit of displacement okay so now we trained this representation for a long long time you know weeks of GPU time and then we have a representation that has the whole thing working right here hope the whole thing working and then we can think we can look at what we | 1,793 | 1,827 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1793s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | can do with that representation so one thing we can do is we can now visualize the source of the sound because what we can do is we can say okay given this task of you know is it is it aligned or not aligned we can actually just use the kind of classic class activation visualization maps to see what pixels is it using to tell if things are aligned or not because those are the pixels that | 1,827 | 1,851 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1827s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder | |
_V-WpE8cmpc | are the sound producing pixels okay so here is an example of some of the places where it thought those were the important places that it's that it decided where the sound producing places and here is some some some visualizations of this over time [Music] okay so this is again completely automatic no labels of any kind okay another thing we can do is we can just | 1,851 | 1,904 | https://www.youtube.com/watch?v=_V-WpE8cmpc&t=1851s | Alexei Efros: Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.