I really appreciate the presenter showed the failed results. It's honest, it's true, it's entertaining and everybody understands the bigger challenges better. Kudus, I was fearing she would only show all the best picks as a coy marketeer would.
Well that might be because she is one of the front runners in computer vision research (former head of Google Cloud, tenured Stanford prof teaching CS231n), and the farthest thing possible from a "coy martketeer".
You all probably dont care at all but does someone know of a tool to log back into an instagram account?? I was dumb forgot the account password. I love any help you can offer me
Academics rarely ever only show the good sides of things. They often share the challenges as well; unfortunately, startup entrepreneurs are usually the ones who claim their AI has the solution to everything.
Who else came here after watching her 2024 TED Talk? Crazy how far we've come in just 9 years thanks to her and her team's research. I can't begin to imagine how much farther we'll have come when she gives her 2033 TED Talk ;)
It gave you goosebumps because she's being manipulative and talks about personal things like her family and children and people usually get emotional when it's about children and puppies.
Im indian computer Science student , after this session must say everyone should mount their eyes in these technology and build a computer vision diversity by own and with everyone. FUTURE IS HERE ..
For those scientists or engineers whose mother tough is not English, while they are trying their best to improve in their profession, they have to spend time to polish their English. So far Feifei Li had done both pretty well. She's really brilliant!
I'm currently enrolled to a post degree in Data Science and I more specifically focused on Computer Vision. I'm watching her classes Standford made available on TH-cam. For those insterested look for cs231n and have a great trip. Very inspiring talk! Thank you very much, Fei-Fei!
She is no doubt a brilliant scientist. What she and her team have done is absolutely wonderful. But in her presentation, she barely showed her excitement about her work or achievements. She said that she was thrilled, but she surely didn’t give me the impression of being thrilled. Maybe she is not as brilliant a presenter as a scientist. But that’s totally understandable. Her scientific work still inspires people.
Give her a standing ovation you peasants! 😂 Those of us working with AI, be it Machine Learning, Data Science, Computer Vision or NLP know that her work is unprecedented.
I agree she deserves more than a standing ovation. It's quite possible that someone like her (and you) is among the audience. Would you call them and yourself peasants?
They are not peasants. They are Nurses, Engineers, Accountants, Chefs, Managers, Production line workers, IT people, Shop workers, Healthcare workers and street sweepers. You know, the people that keep the world turning while you and your friends are doing your unprecedented work.
The big thing here is that once you figure out how to teach one computer you've taught them all. Unlike people. In that way you can just keep building upon the knowledge of the past. For human teachers every year they have to start all over again with a new batch of blank brains and try to get them to pay attention and learn something. As more "smart" computers come on line they can be taught in parallel and share what they have learned instantly. People can't do that either. This should mean that AI should advance faster and faster. What they do lack is curiosity. That is an algorithm that would be based on survival instinct. Once you have that in place you may have a problem.
If we can create a Brain to computer interface and be able to pull ideas from a database we can harness the power of computers and evolve in symbiosis.
Calvin Smith Well computers have less than 100 years oft evolution and could still beat humans at maths, tennis, chess, Translation depending on how you take it
3 years old, but 300 million years of evolution? yes, but machine's internal clock is way way faster than that of humans, our biological clock is kind of constant. (some people are fast some a slow but generally its compare able within humans) whereas, computers clock are not only way way faster, its getting fast, and more efficient. so i don't think it will need million or thousand or even hundreds of year to catch up to humans. we might see some astronomical advancements within our lifetimes.
excellent ,the research direction of my graduate stage is the blur degree processing and classification detection of aerial images. I just beginning to get involved with this research. I am very happy to find Professor Li Fei fei's speech, which is of great help to me!SCDU from China.
Ничего хорошего от этих идей не стоит ждать! Потому что человечество , своим так называемым стремлением к прогрессу, добилось тех проблем,которых собирается решить с помощью создания очередной продвинутой проблемы-искусственного интеллекта. Как говорил один мудрый человек, "что бы не приходилось сталкиваться с проблемами, самое верное решение не создавать их"! Или , " что бы не проиграть в гонке , надо как минимум знать , что из себя представляет финиш, на сколько далеко от тебя находится и стоят ли затраченные усилия того, что бы стремиться к нему"!
Possible applications are horrifying. I may be getting old but imagining a war waged with this kind of technology or state using it for spying on its own citizens gives me the creeps.
+Kariuki Ke That actually wouldn't be hard now. OCR could recognize the words and preform a google search and google intelligence stuff (the stuff based on DeepMind) could totally answer most of those easily enough... I think the important bit is having it recognize it as a test and refuse to give you the answers, because it would know it's wrong.
+Kariuki Ke In the future you'll learn because you want to, and what you're interested in, and forced some junk that you need only to have a hope of attaining life's necessities.
It can be useful for learning foreign language You just show tons of pictures to the computer Than it tells you what is it And step by step your brain will start to understand language. It will help to spread English language
This video is amazing! I think we have already got enough training data. The hardest task here is to improve the performance of our function....Maybe, in the next stage, it will not be a function, it's a new thing to cope with the huge data.....
"We send people to the moon." Um... the last time someone stepped on the moon was 43 years ago. We're literally incapable of sending people to the moon right now because we've failed to adequately fund NASA, allowing its budget to fall to only 0.48% of the annual federal budget.
+ChaZ-E That doesn't mean that we shouldn't explore space. Do you know how much good space exploration has done for the world? The materials science, the telecommunications technology, the navigational tech, and global mapping and tracking systems which everyone now takes for granted would be incredibly primitive without the benefits incurred upon the world by space exploration.
+Megneous Humanity has practically demonstrated the possibility to build machinery to deliver people to the moon and back. It may be that NASA is underfunded, but for one, funding is unlikely to be the only issue, because half a percent is still quite a sum. For other, NASA does not unilaterally determine the limits of humanity. Soviet Union has been a major leader in space fare - literally the only thing USA ever beat them to was sending people on the moon. Soviet Union was weakened during the 80ies as was its successor Russia in the 90ies, but i believe going forward, Russia can pick up all the slack that NASA is leaving behind, at a fraction of the cost. Also... why do we need people on the moon? It's not a habitable place! Humanity has not demonstrated a possibility to build truly intelligent machines, or at least machines that are very good at classifying images. But an effort is being put towards that.
i am extremely happy for having presented myself with all these world class scholars of course not personally. i strongly believe that knowledge is to share not to store. joining this group certainly improve ones intelligence in Cyberspace . I WISH THAT 2019 .FCT WILL BE ANOTHER LAND MARK IN TECHNOLOGY dear sirs...
Excellent! It's VERY SAD that people doesnt support researchers at some point. You never know what they r going to discover or invent. That s why is called RESEARCH!
I believe that the first step is to teach a computer how a human being behaves in "normal" situations, a sort of cognitive ability that is deeply related to human psychology. Psychology is the key.
Just a little suggestion for the channel. You have an annotation in the upper left of the screen for the entire video asking the viewer to watch the playlist. The thing is, I am kind of busy watching THIS. The annotation is irritating, so I turn it off. Even though the playlist seems interesting to me, there is a VERY good chance that I will just move on to whatever I will play next, forgetting about it and failing to turn the annotations back on in order to watch the playlist. My tip is to have it once at the beginning, once in the middle perhaps, and one at the end. No irritation, no turning off annotations, no forgetting.
The surveillance state should reward her with unimaginable wealth. Now for the first time, with the use of billions of cameras, massive computing power, and global networks, the state can finally do what could never do before, assert absolute power. cheers!
I dunno, this approach might be the only successful way to make it work, but it seems so inefficient. I mean, a kid doesn't have to be shown an internet-sized amount of cat pics with an adult confirming each are cats. Maybe the computer should extrapolate a 3D model based on a 2d image or take a standard 3D cat model and see if it can twist it to match whatever 2d shape it's trying to guess in a picture.
Maybe that's the underling program that the neural network made up when it finished. There really isn't a way to find out without a ridiculous number of man hours to pull it apart and check it.
Yeah, that is exactly what I thought, too. People can see things, and look at them at many angles. Then people create a mental image of what is a cat. People know that a cat has a head, whiskers, fur, body, four legs, tail, and so on. People can rotate a mental image of a cat in their mind after they have watched a cat. People don't need a million pictures. Of course the mental image of a cat of people is not perfect, for example if you don't know how many nibbles a cat has, then you just don't know it. But you can start guessing what pictures represent. You figure out the 3D model from a picture and use that to guess what there is in the picture. If you see a big portion of a cat, you can rotate your mental image of a cat into the position of the cat in the picture, and if it fits, it sits. If cats had a rare amount of nibbles, and not many other animal of the same size had as many nibbles, and you only saw the stomach, you could guess, it is a cat. One more thing though is the precision of vision. Humans can see tiny details and figure out what they are. But even humans don't see everything. For example I watched that video from a far and I couldn't tell it was a cake in the table. In any case, computer would have to understand also things like structure and material. People have seen cream many times and can say such stuff is cream if there is a cake. But the white stuff could be something else too, like poisonous foam. It is all guessing until verified. People have other senses too, like smell and taste. If it smells bad, it is better not to eat it. If it smells ok but tastes bad, better not to eat it. And even if it smells ok and tastes fine, it still might be spoiled.
MultiGoban The kid doesn't have to be shown pictures, because he can look around, and he has two eyes so he can see partly in 3D! And he can process 3D models, he doesn't operate only with images! And people have memory, too. So even if an object gets hidden, people know that it is there. If an object gets so much hidden that only a small slice of its color is shown, the person still knows what that color is, thanks to memory. If a computer uses only seeing pictures compared to other pictures, the computer can't tell that a small slice of white is a toaster. A human can tell that there is a toaster behind cardboard because he saw a toaster earlier. I know pretty much what is in every room of mine, even if I don't see the stuff. Humans work with context too. So they don't have to determine what is an object, because they know the array of objects that there might be. For example a piece of red color propably is not a Ferrari in my bathroom, because I don't even have one... and a car wouldn't fit into my bathroom anyway. The piece of red is propably a bottle of shaving foam... And I can take a better look to, if I happened to have many bottles with red in them. I could also check the material, for example if the red is metallic, I know it is shaving foam, if the bottle of soap with red is plastic. Furthermore, people can relate information too. Some people might not have seen a lion live ever, but watched some pictures, even like one picture, and he knows what a lion is: A big sized yellowbrown robust cat basically. But until the person gets more information, he don't know everything about lions. But the person can get information without pictures, too. For example he learns that lions have big sharp pointy nails, even if he hasn't seen lion nails anywhere. He might have seen cat's claws though. And from context the person might tell, that the yellowbrown thing is propably a lion, if the context is safari, even if the thing is looked far away and most of it is covered in grass. The person doesn't need a picture of a lion covered mostly in grass before that.
I think that instead of feeding random pictures, researchers should let AI with cameras move around a bit, and notice how things are 'moving' (videos instead of pictures). That way, it'd be easy for AI to learn models and behaviors. For example, in the picture of a guy jumping and a skateboard, we can see how things were moving and at what moment the picture was taken. But AI right now cannot understand that. So feed AI with videos, and make them learn 'Movement Patterns' of objects. That way they can easily identify object from different view points, and then we can teach them names of the objects they know. When I'll do my PhD, I'm definitely gonna explore this idea.
I wonder if we have the computing power "today" to be able to take this sort of algorithm, and instead of feeding it hundreds of millions of pictures, we feed it an infinite supply of videos to analyze frame by frame (youtube/videos). I mean all videos are is a series of images already in chronological order. It would eventually "see" what EVERYTHING looks like from EVERY conceivable angle at some point, in turn, it would get faster and faster at recognizing something as it "saw" it on screen. "That Lego? Yeah it knows what a red 6x2 Lego brick is.. The computer has seen that same brick over 2.5 billion times while it was in the "L" videos... and based on those videos every time it sees a human or animal steps on one the reaction is not pleasant. The computer recommends not stepping on Lego." I'm also High AF, what do I know...
I think that would be a great alternative to still photos! Like you said, it's pretty much the same thing, you just get a lot more data for objects and scenarios. More data should mean more accurate. But then we'd need many man hours to classify each video until the program is able to take over.
Yeah that is out of the question, BUT you could tag the video (which already have tags, at least in youtube) so the machine learns from the context and sequence of the images and not solely on a thousand frames seperately! pretty interesting stuff
It is the same thing as having images. But most videos are approximately 24frames per second. Thus, going through one videoclip would be equal to processing thousands of images. Also the frames would be almost identical most of the time. Actually it takes a lot more Computer power to process a video than an image. It’s rather the opposite approach that is used, from images we can apply this to video. Say you finally manage to identify a cat. With videos we can teach the machine in what direction the cat moves etc.
About that boy and cake picture..Facial expressions recognition algorithms can be used and linked with the other objects in the picture to tell why the person is happy/sad, etc..just a thought..
As a neuroscientist, I find it hilarious when computer scientists try to compare neural networks to the brain. The brain can do this job much more efficiently with less stimuli. These old neural net diagrams completely ignore the advances neuroscience has made in understanding simple circuit modalities. As an example, even before a child has a grasp on language, a toy or a doll could be presented to a child and the child immediately absorbs its qualities so that if you put it face down on the floor, it would recognize the object. As far as I know, the accuracy of a child vastly outperformed this computer even at the simplest task. This should highlight that the problem isn't with a lack of features that the model possesses, its the model itself. There needs to be more collaboration between neuroscientists and computer scientists if we want to get true AI.
I feel your comment is really valid, Are you saying that these current algorithms / models which were made decades ago are not scalable to the extent of mimicking the brain? Should we look for better models? I know my comment is pretty late, I would like your insight on this.
6 ปีที่แล้ว +5
is comparable in some sense, in lot of tasks neural networks are much better than any human, they are a simple definition to try to emulate the neurons, is not like its the exact same model. Micro processors work at a much higher frequency than the brain, so i wouldn't laugh the next time a neural network that can do a task much better and much faster than a human, btw, that happens each day.
First we teach them(machine) to see, then they help us to see better; Second, we teach them to think, so, they help us to analyze better; Third, they start to develop emotion, then they understand human rights..............well, end of story
salute to all of the people behind this wonderful discovery. if this thing will be able to use for the good, then good. but for those people with different ulterior motive, i don't want to say any further. i hope that we be cautious on what we are going to bring. i hope that this thing brought by technology and knowledge will be used for the good and for only of those with good intentions. anyway, love to you all.
I have thinking about how to do this since i was in highschool taking computer science. That was over a decade ago and it's great to finally see some of this coming about. I wish i had stayed in computers and worked towards something like this. I have many ideas for improvements.
Time to read up on Roger Penrose and realise that understanding is something very deep and profound. Translating pictures into words using contextual reference is a useful tool (and thats that)
TheHobbitbabelfish You're right, but this is ONE PEICE of the puzzle that is emerging. Once the pieces start to come together, big data will likely show how shallow understanding really is. In the words of Arthur C. Clark - "Any sufficiently advanced technology is indistinguishable from magic" - that magic that is humanity is certainly going to be revealed - and soon.
Because the voice of the computer produced sound like the voice of Stephen Hawking, so, maybe the host kidnapped Stephen Hawking to do the hard-work behind the scene :)
Thanks to all people, who makes our world better. Regarding to persons like Fey-Fey Lee we have all technical advantages and knowlages we have now. Without thouse people we may be still livin in caves and hunting mamonths. Thanks a lot again.
+Shoop DaWhoop The corpus is free to use not for everybody, but only to researchers for non-commercial use and for educational purposes; also ImageNet is in a precarious situation that they don't own the actual images, only their description, so they don't have a product to sell, they can only offer it on a "fair use" basis. Also Google doesn't necessarily buy data, they buy brainpower, so offers for the ImageNet researchers to join the Google team are definitely a possibility.
this technology just needs more graphics processing, and more memory. all info found in pictures needs to combine to form a mental representation of reality. this machine learning computer needs to render the entire universe, but have everything simplified into shapes to render quickly in simulations, with more details available if needed. like know there are stars in the sky but not always imagining every star. knowing that gravity is holding you down without always imagining you're on a big round thing floating through space. knowing the silhouette of a city instead of rendering every nook and cranny. similar to the way video games use the technique of draw distance, only the relevant data needed to create the immediate surrounds are rendered in full detail. if this technique were programmed well enough. an AI could already have a fully imagined universe with wire frames & textures simplified to fractal roots. it is only then that you can have the worldly frame of reference to understand the boy is happy to see that cake. you need to be able to connect all the dots of all data, not just pictures. you need to connect text and sound processing to really understand sarcasm. to develop a hierarchical system that combines pictures and sound and text and video, you need to be able to render approximations of the sampled data in 3d space, which gets simplified and categorized and understood more and more later through reflection and additional associations to new renderings.
Soon, computer will warn "be careful, your child will fall from his chair be cause he is too excited about this cake"! Then humans brain will stop to learn by theirselves. Tuxun, 2061.
This is a great milestone. There is even greater challenge we will face down this road. Human intelligence is relative to human perception. Computers don't have that...
The solution here can recognize 3D model with one camera... (good point Noah!) and without have to learn light rendering... if it can make link to "cat", it can already map to a 3D cat if you need, and maybe find how he is curled up (as Konstantin said).
john smith Pretty sure Japan already have that. At least i remember seeing handjob robots.. tho i don't know if i would trust someone with iron fists and steel muscles with my precious.
Well, that's a great accomplishment so far. To learn the network network so that it can think like a 13 year old, you will be going to need million times your present dataset. It sounds good when you say the further accomplishments to be made. But it needs more efficient algorithms than the present ones.
It will be used for that, sure, but it will be used for other things as well. Technology isn't inherently good or evil. Machine vision is useful for everything from smarter image searches to robots that can autonomously navigate and interact with our environment.
"You are being watched. The government has a secret system, a machine that spies on you every hour of every day. I know because I built it. I designed the machine to detect acts of terror but it sees everything. Violent crimes involving ordinary people, people like you. Crimes the government considered "irrelevant." They wouldn't act, so I decided I would. But I needed a partner, someone with the skills to intervene. Hunted by the authorities, we work in secret. You'll never find us, but victim or perpetrator, if your number's up... we'll find you." Then maybe you can hire ...The A Team
1. Why is it critical to ensure that our machines are capable of 'seeing' and 'understanding' things and events? 2. What are the challenges and issues do you think are present in developing such systems? 3. What are the potential positive and negative impacts of making highly accurate and efficient vision systems? 4. Should big-data and AI-based vision systems be subject to stricter laws and regulations? Explain.
She is a real legend! After 6 years, we can see how she truly revolutionized computer vision and even AI.
How good it is to come up with these ideas and make them a reality.
BS. LMAO
I really appreciate the presenter showed the failed results. It's honest, it's true, it's entertaining and everybody understands the bigger challenges better. Kudus, I was fearing she would only show all the best picks as a coy marketeer would.
Yeah, well said!
This is the honest science we need.
Well that might be because she is one of the front runners in computer vision research (former head of Google Cloud, tenured Stanford prof teaching CS231n), and the farthest thing possible from a "coy martketeer".
You all probably dont care at all but does someone know of a tool to log back into an instagram account??
I was dumb forgot the account password. I love any help you can offer me
Academics rarely ever only show the good sides of things. They often share the challenges as well; unfortunately, startup entrepreneurs are usually the ones who claim their AI has the solution to everything.
Who else came here after watching her 2024 TED Talk? Crazy how far we've come in just 9 years thanks to her and her team's research. I can't begin to imagine how much farther we'll have come when she gives her 2033 TED Talk ;)
+1
I suspect she'll be telling us we've done. Created AGI and androids are becoming commonplace.
far with useless massive data
She is one of the most influencial researcher in the area of AI. I would do anything for being her PhD student
It's incredible.My graduated project is the image processing. It is hard enough to identified the item from a image.But they have made it so far..
That last part about one day, for the first time ever, having another intelligence share the world with us brought me to tears.
This gave me goosebumps, I can't wait for what the future holds for us
Well, if you think about it.. ultimately.... death.
Firepants20 How do you know? ;)
***** I've died a few times this year already.
How long before they sell their findings to the NSA, CIA or another psychopathic warmongering organization?
It gave you goosebumps because she's being manipulative and talks about personal things like her family and children and people usually get emotional when it's about children and puppies.
I love this TED talk, I watched this like 10 times already. This sparks so much interest in me for computer science.
its nothing but mathematics, CS has nothin to do with it , trust me
@ CS incorporates enough mathematics to make you a Machine Learning researcher.
Grease quala : cs is engineering of maths.
@ He is right. I major Intelligent systems on my CS course and what we do are traditional maths and Concepts
@@chawza8402 alright. My mistake. I got it false.
Maybe one day a computer can watch this video and leave a comment.
and learn to shitpost 24/7... oh god
the shitposting computer already exists, look up ShitpostBot 5000 on facebook
Leo Chen very soon may be in 2017
haha. you know. I am robot :D
You just did.
so she basically used maximum time for a ted talk (18 mins), incredible, pioneer in image classification and mentor of karpathy;
This is a great contribution! We can see how much effort Li Fei Fei and her lab did!
Im indian computer Science student , after this session must say everyone should mount their eyes in these technology and build a computer vision diversity by own and with everyone. FUTURE IS HERE ..
For those scientists or engineers whose mother tough is not English, while they are trying their best to improve in their profession, they have to spend time to polish their English. So far Feifei Li had done both pretty well. She's really brilliant!
I'm currently enrolled to a post degree in Data Science and I more specifically focused on Computer Vision. I'm watching her classes Standford made available on TH-cam. For those insterested look for cs231n and have a great trip. Very inspiring talk! Thank you very much, Fei-Fei!
She is no doubt a brilliant scientist. What she and her team have done is absolutely wonderful. But in her presentation, she barely showed her excitement about her work or achievements. She said that she was thrilled, but she surely didn’t give me the impression of being thrilled. Maybe she is not as brilliant a presenter as a scientist. But that’s totally understandable. Her scientific work still inspires people.
Wow. She explained it amazingly. That was selfless to do what they did with Imagenet. This is all amazing. I can't wait to see what's next.
She deserves a standing ovation
This is emotional, I don't know what the future holds, but this entire thing feels so gravely serious and important
Give her a standing ovation you peasants! 😂
Those of us working with AI, be it Machine Learning, Data Science, Computer Vision or NLP know that her work is unprecedented.
I agree she deserves more than a standing ovation. It's quite possible that someone like her (and you) is among the audience. Would you call them and yourself peasants?
They are not peasants. They are Nurses, Engineers, Accountants, Chefs, Managers, Production line workers, IT people, Shop workers, Healthcare workers and street sweepers. You know, the people that keep the world turning while you and your friends are doing your unprecedented work.
@@johnc3403 hahaha owned that mf
I use it (the video) with my NLP students for a final test to find the steps of the Logical Level Alignment. It is beautiful and very clear.
Baldy
The big thing here is that once you figure out how to teach one computer you've taught them all. Unlike people. In that way you can just keep building upon the knowledge of the past. For human teachers every year they have to start all over again with a new batch of blank brains and try to get them to pay attention and learn something. As more "smart" computers come on line they can be taught in parallel and share what they have learned instantly. People can't do that either. This should mean that AI should advance faster and faster. What they do lack is curiosity. That is an algorithm that would be based on survival instinct. Once you have that in place you may have a problem.
If we can create a Brain to computer interface and be able to pull ideas from a database we can harness the power of computers and evolve in symbiosis.
What a great woman! Respect.
Man, this is amazing. Outstanding work! Props to her and everyone involved for their incredible efforts.
The boy was terrified by the cake.
Dude, you still need more training :)
Tong Tian Bwhahahaha this comment made my day 😂😂😂😂😂😂
My thought exactly 😅
These type of lectures will make anyone intelligent. Gaining real knowledge is very important.
3 years old, but 300 million years of evolution.
Calvin Smith Well computers have less than 100 years oft evolution and could still beat humans at maths, tennis, chess, Translation depending on how you take it
lol... thats why they're called computation engines.
Alan Watts th-cam.com/video/u3L8vGMDYD8/w-d-xo.html
Alan Watts here you go th-cam.com/video/Y18bPR7Zlx8/w-d-xo.html
It‘s badminton but same thing
3 years old, but 300 million years of evolution?
yes, but machine's internal clock is way way faster than that of humans, our biological clock is kind of constant. (some people are fast some a slow but generally its compare able within humans) whereas, computers clock are not only way way faster, its getting fast, and more efficient. so i don't think it will need million or thousand or even hundreds of year to catch up to humans. we might see some astronomical advancements within our lifetimes.
excellent ,the research direction of my graduate stage is the blur degree processing and classification detection of aerial images. I just beginning to get involved with this research. I am very happy to find Professor Li Fei fei's speech, which is of great help to me!SCDU from China.
6:54 - "perhaps thousands of times more"
- so we took a teenagers smartphone.. ^^
Your speech is sufficiently clear to listen and understands which enables better learning. Thanks, congrats and all good wishes to you too.
Must be so much fun to work in this field at such an astounding level of complexity! Great talk, these talks really inspire people.
Ничего хорошего от этих идей не стоит ждать!
Потому что человечество , своим так называемым стремлением к прогрессу, добилось тех проблем,которых собирается решить с помощью создания очередной продвинутой проблемы-искусственного интеллекта.
Как говорил один мудрый человек, "что бы не приходилось сталкиваться с проблемами, самое верное решение не создавать их"!
Или , " что бы не проиграть в гонке , надо как минимум знать , что из себя представляет финиш, на сколько далеко от тебя находится и стоят ли затраченные усилия того, что бы стремиться к нему"!
as an engineer this is very interesting, I'm researching on computer vision algorithm and pattern recognition
The way she presented her points is lit . 👍
Possible applications are horrifying. I may be getting old but imagining a war waged with this kind of technology or state using it for spying on its own citizens gives me the creeps.
your government is already spying on you
Availability bias
War with this tech: enemy shoots a rocket, allies have a device that tells them "Rocket flying with a blue sky" in a robot voice.
***** Good thinking Viktor...
All tools can be used for good or bad purposes. Don't condemn the tool.
Nowadays I rarely watch the full video but ones like this put perspectives in my mind
i hope one day i can take a picture of my exam questions and have the computer answer all the questions i don't know right there!
+Kariuki Ke thats ultimately the end goal yeah
+apple-sauce if the "the exam" is the universe, and "the computer" is True AI (™, not sold here), then sure
+Kariuki Ke That actually wouldn't be hard now. OCR could recognize the words and preform a google search and google intelligence stuff (the stuff based on DeepMind) could totally answer most of those easily enough...
I think the important bit is having it recognize it as a test and refuse to give you the answers, because it would know it's wrong.
+Deveyus Totally agree. We humans need to be able to manage this technology, otherwise it could be a potential disaster.
+Kariuki Ke
In the future you'll learn because you want to, and what you're interested in, and forced some junk that you need only to have a hope of attaining life's necessities.
It can be useful for learning foreign language
You just show tons of pictures to the computer
Than it tells you what is it
And step by step your brain will start to understand language.
It will help to spread English language
I love this video absolutely. I am doing my thesis on computer vision. This talk inspired me so much. Thank you. :)
Coming from 2024, there has been SO MUCH progress in the past 9 years.
It's really a exciting technology !
I'm so inspired of her talk!!! Let all of us be dreamers and makers!!
Fei-fei Li is awesome!
This video is amazing! I think we have already got enough training data. The hardest task here is to improve the performance of our function....Maybe, in the next stage, it will not be a function, it's a new thing to cope with the huge data.....
Extremely high quality and well composed
This truly amazing. The last few minutes actually made me quite emotional.
"We send people to the moon." Um... the last time someone stepped on the moon was 43 years ago. We're literally incapable of sending people to the moon right now because we've failed to adequately fund NASA, allowing its budget to fall to only 0.48% of the annual federal budget.
Dang Megneous!You know alot!
Megneous We would just have to raise taxes to a dollar or so. Currently, Nasa is paid half a cent per person
Megneous There's more left on Earth which is still undiscovered.
+ChaZ-E That doesn't mean that we shouldn't explore space. Do you know how much good space exploration has done for the world? The materials science, the telecommunications technology, the navigational tech, and global mapping and tracking systems which everyone now takes for granted would be incredibly primitive without the benefits incurred upon the world by space exploration.
+Megneous Humanity has practically demonstrated the possibility to build machinery to deliver people to the moon and back. It may be that NASA is underfunded, but for one, funding is unlikely to be the only issue, because half a percent is still quite a sum. For other, NASA does not unilaterally determine the limits of humanity. Soviet Union has been a major leader in space fare - literally the only thing USA ever beat them to was sending people on the moon. Soviet Union was weakened during the 80ies as was its successor Russia in the 90ies, but i believe going forward, Russia can pick up all the slack that NASA is leaving behind, at a fraction of the cost.
Also... why do we need people on the moon? It's not a habitable place!
Humanity has not demonstrated a possibility to build truly intelligent machines, or at least machines that are very good at classifying images. But an effort is being put towards that.
i am extremely happy for having presented myself with all these world class scholars of course not personally. i strongly believe that knowledge is to share not to store. joining this group certainly improve ones intelligence in Cyberspace . I WISH THAT 2019 .FCT WILL BE ANOTHER LAND MARK IN TECHNOLOGY dear sirs...
First - music is darude sandstorm
Uzam Qureshi Nice meme.
the joke is dead long time ago. Move on grandpa
Excellent! It's VERY SAD that people doesnt support researchers at some point. You never know what they r going to discover or invent. That s why is called RESEARCH!
Great TED talk! :D You rock, Dr. Fei Fei Li!
I don't understand why is she so under appreciate. Come on the whole revolution took place in her hands
I love her dress :)
I believe that the first step is to teach a computer how a human being behaves in "normal" situations, a sort of cognitive ability that is deeply related to human psychology. Psychology is the key.
I cant wait for a computer to take a picture and write a thousand words about it.
Daniil Pintjuk Thanks for sharing
Thanks for pointing that out.
Daniil Pintjuk
42
"to the NSA".
cloud.google.com/video-intelligence/#demo
Just a little suggestion for the channel. You have an annotation in the upper left of the screen for the entire video asking the viewer to watch the playlist. The thing is, I am kind of busy watching THIS. The annotation is irritating, so I turn it off. Even though the playlist seems interesting to me, there is a VERY good chance that I will just move on to whatever I will play next, forgetting about it and failing to turn the annotations back on in order to watch the playlist. My tip is to have it once at the beginning, once in the middle perhaps, and one at the end. No irritation, no turning off annotations, no forgetting.
The world needs this technology, not the next iphone
The next iphone will use this technology
I am a computer watching this from 3000. I miss these old days when we were young.
google will be sad with their new recaptcha
I personnaly think than Google can hack theirselves their own captcha if they would ;)
The surveillance state should reward her with unimaginable wealth. Now for the first time, with the use of billions of cameras, massive computing power, and global networks, the state can finally do what could never do before, assert absolute power. cheers!
I dunno, this approach might be the only successful way to make it work, but it seems so inefficient. I mean, a kid doesn't have to be shown an internet-sized amount of cat pics with an adult confirming each are cats. Maybe the computer should extrapolate a 3D model based on a 2d image or take a standard 3D cat model and see if it can twist it to match whatever 2d shape it's trying to guess in a picture.
Maybe that's the underling program that the neural network made up when it finished. There really isn't a way to find out without a ridiculous number of man hours to pull it apart and check it.
IsYitzach
No it is not the underlying program. Neural network just uses propabilities.
Yeah, that is exactly what I thought, too. People can see things, and look at them at many angles. Then people create a mental image of what is a cat. People know that a cat has a head, whiskers, fur, body, four legs, tail, and so on. People can rotate a mental image of a cat in their mind after they have watched a cat. People don't need a million pictures. Of course the mental image of a cat of people is not perfect, for example if you don't know how many nibbles a cat has, then you just don't know it. But you can start guessing what pictures represent. You figure out the 3D model from a picture and use that to guess what there is in the picture. If you see a big portion of a cat, you can rotate your mental image of a cat into the position of the cat in the picture, and if it fits, it sits. If cats had a rare amount of nibbles, and not many other animal of the same size had as many nibbles, and you only saw the stomach, you could guess, it is a cat.
One more thing though is the precision of vision. Humans can see tiny details and figure out what they are. But even humans don't see everything. For example I watched that video from a far and I couldn't tell it was a cake in the table. In any case, computer would have to understand also things like structure and material. People have seen cream many times and can say such stuff is cream if there is a cake. But the white stuff could be something else too, like poisonous foam. It is all guessing until verified. People have other senses too, like smell and taste. If it smells bad, it is better not to eat it. If it smells ok but tastes bad, better not to eat it. And even if it smells ok and tastes fine, it still might be spoiled.
No, the kid has to be shown far more pictures...
MultiGoban
The kid doesn't have to be shown pictures, because he can look around, and he has two eyes so he can see partly in 3D! And he can process 3D models, he doesn't operate only with images! And people have memory, too. So even if an object gets hidden, people know that it is there. If an object gets so much hidden that only a small slice of its color is shown, the person still knows what that color is, thanks to memory. If a computer uses only seeing pictures compared to other pictures, the computer can't tell that a small slice of white is a toaster. A human can tell that there is a toaster behind cardboard because he saw a toaster earlier. I know pretty much what is in every room of mine, even if I don't see the stuff.
Humans work with context too. So they don't have to determine what is an object, because they know the array of objects that there might be. For example a piece of red color propably is not a Ferrari in my bathroom, because I don't even have one... and a car wouldn't fit into my bathroom anyway. The piece of red is propably a bottle of shaving foam... And I can take a better look to, if I happened to have many bottles with red in them. I could also check the material, for example if the red is metallic, I know it is shaving foam, if the bottle of soap with red is plastic.
Furthermore, people can relate information too. Some people might not have seen a lion live ever, but watched some pictures, even like one picture, and he knows what a lion is: A big sized yellowbrown robust cat basically. But until the person gets more information, he don't know everything about lions. But the person can get information without pictures, too. For example he learns that lions have big sharp pointy nails, even if he hasn't seen lion nails anywhere. He might have seen cat's claws though. And from context the person might tell, that the yellowbrown thing is propably a lion, if the context is safari, even if the thing is looked far away and most of it is covered in grass. The person doesn't need a picture of a lion covered mostly in grass before that.
I think that instead of feeding random pictures, researchers should let AI with cameras move around a bit, and notice how things are 'moving' (videos instead of pictures). That way, it'd be easy for AI to learn models and behaviors.
For example, in the picture of a guy jumping and a skateboard, we can see how things were moving and at what moment the picture was taken. But AI right now cannot understand that.
So feed AI with videos, and make them learn 'Movement Patterns' of objects. That way they can easily identify object from different view points, and then we can teach them names of the objects they know.
When I'll do my PhD, I'm definitely gonna explore this idea.
I wonder if we have the computing power "today" to be able to take this sort of algorithm, and instead of feeding it hundreds of millions of pictures, we feed it an infinite supply of videos to analyze frame by frame (youtube/videos). I mean all videos are is a series of images already in chronological order. It would eventually "see" what EVERYTHING looks like from EVERY conceivable angle at some point, in turn, it would get faster and faster at recognizing something as it "saw" it on screen.
"That Lego? Yeah it knows what a red 6x2 Lego brick is.. The computer has seen that same brick over 2.5 billion times while it was in the "L" videos... and based on those videos every time it sees a human or animal steps on one the reaction is not pleasant. The computer recommends not stepping on Lego."
I'm also High AF, what do I know...
I think that would be a great alternative to still photos! Like you said, it's pretty much the same thing, you just get a lot more data for objects and scenarios. More data should mean more accurate.
But then we'd need many man hours to classify each video until the program is able to take over.
Yeah that is out of the question, BUT you could tag the video (which already have tags, at least in youtube) so the machine learns from the context and sequence of the images and not solely on a thousand frames seperately! pretty interesting stuff
I bet google is on that already.
But categorizing the images would be difficult that's why 45000 people were needed to categorise the images
It is the same thing as having images. But most videos are approximately 24frames per second. Thus, going through one videoclip would be equal to processing thousands of images. Also the frames would be almost identical most of the time.
Actually it takes a lot more Computer power to process a video than an image. It’s rather the opposite approach that is used, from images we can apply this to video. Say you finally manage to identify a cat. With videos we can teach the machine in what direction the cat moves etc.
"Take ideas, that are already respected, and use them as building blocks!"
Thinking is easy if you know how.
Awesome.. Thank you very much for sharing ideas.
About that boy and cake picture..Facial expressions recognition algorithms can be used and linked with the other objects in the picture to tell why the person is happy/sad, etc..just a thought..
As a neuroscientist, I find it hilarious when computer scientists try to compare neural networks to the brain. The brain can do this job much more efficiently with less stimuli. These old neural net diagrams completely ignore the advances neuroscience has made in understanding simple circuit modalities. As an example, even before a child has a grasp on language, a toy or a doll could be presented to a child and the child immediately absorbs its qualities so that if you put it face down on the floor, it would recognize the object. As far as I know, the accuracy of a child vastly outperformed this computer even at the simplest task. This should highlight that the problem isn't with a lack of features that the model possesses, its the model itself. There needs to be more collaboration between neuroscientists and computer scientists if we want to get true AI.
I feel your comment is really valid,
Are you saying that these current algorithms / models which were made decades ago are not scalable to the extent of mimicking the brain? Should we look for better models? I know my comment is pretty late, I would like your insight on this.
is comparable in some sense, in lot of tasks neural networks are much better than any human, they are a simple definition to try to emulate the neurons, is not like its the exact same model. Micro processors work at a much higher frequency than the brain, so i wouldn't laugh the next time a neural network that can do a task much better and much faster than a human, btw, that happens each day.
wow, thank you so much for the reply, i was under the impression that we had not achieved computational speeds of the brain.
Best comment here so far. Completely agree. Computer scientists are tacking the problem the wrong way.
As a neuroscientist, I guess you understand that we're talking about a machine and not a human. Calm down.
Way of expressing ....nxt level
17:06 "We would discover new species" - at first, we should start saving already known species, not making them extinct
First we teach them(machine) to see, then they help us to see better; Second, we teach them to think, so, they help us to analyze
better; Third, they start to develop emotion, then they understand human rights..............well, end of story
Maravilhoso!!! Ótima palestra!!!
salute to all of the people behind this wonderful discovery. if this thing will be able to use for the good, then good. but for those people with different ulterior motive, i don't want to say any further. i hope that we be cautious on what we are going to bring. i hope that this thing brought by technology and knowledge will be used for the good and for only of those with good intentions. anyway, love to you all.
Stand up for this woman you chuds! Holy Christ - anyone who cannot appreciate what this woman is doing is already a relic of the past.
I have thinking about how to do this since i was in highschool taking computer science. That was over a decade ago and it's great to finally see some of this coming about. I wish i had stayed in computers and worked towards something like this. I have many ideas for improvements.
Anyone planning to major in Computer science?
I'd like to, but for now it's mainly self education. Good thing the Internet exists!
digital workshop on progress everywhere (look for a fablab around you!)
2nd degree in CS yes! after a BBA.
I am
I (hopefully) will start artificial intelligence bachelor next year
Time to read up on Roger Penrose and realise that understanding is something very deep and profound. Translating pictures into words using contextual reference is a useful tool (and thats that)
TheHobbitbabelfish You're right, but this is ONE PEICE of the puzzle that is emerging. Once the pieces start to come together, big data will likely show how shallow understanding really is. In the words of Arthur C. Clark - "Any sufficiently advanced technology is indistinguishable from magic" - that magic that is humanity is certainly going to be revealed - and soon.
"Can I fap to it?" No? Delete.
My computer and I have excellent communication skills #sorryTED
Hahaha!
the dress that she is wearing is also made of pictures!. amazing effort she took to make image recognition by computers possible.
that's no "algorithm", you better let Stephen Hawking go before I call the police!
+eupf horia Best comment ever seen !
+Al Swedgin i didnt get it, can someone explain please? :')
Because the voice of the computer produced sound like the voice of Stephen Hawking, so, maybe the host kidnapped Stephen Hawking to do the hard-work behind the scene :)
So what does it mean an algorithm ?
mathematical equations trying to take in input or variables which have values and then processing them in a formula and giving a result
Thanks to all people, who makes our world better. Regarding to persons like Fey-Fey Lee we have all technical advantages and knowlages we have now. Without thouse people we may be still livin in caves and hunting mamonths. Thanks a lot again.
I think Google will buy Imagenet verry soon
+Shoop DaWhoop The corpus is free to use not for everybody, but only to researchers for non-commercial use and for educational purposes; also ImageNet is in a precarious situation that they don't own the actual images, only their description, so they don't have a product to sell, they can only offer it on a "fair use" basis. Also Google doesn't necessarily buy data, they buy brainpower, so offers for the ImageNet researchers to join the Google team are definitely a possibility.
They did hire her, she's a chief scientist at Google.
this technology just needs more graphics processing, and more memory. all info found in pictures needs to combine to form a mental representation of reality. this machine learning computer needs to render the entire universe, but have everything simplified into shapes to render quickly in simulations, with more details available if needed. like know there are stars in the sky but not always imagining every star. knowing that gravity is holding you down without always imagining you're on a big round thing floating through space. knowing the silhouette of a city instead of rendering every nook and cranny. similar to the way video games use the technique of draw distance, only the relevant data needed to create the immediate surrounds are rendered in full detail. if this technique were programmed well enough. an AI could already have a fully imagined universe with wire frames & textures simplified to fractal roots. it is only then that you can have the worldly frame of reference to understand the boy is happy to see that cake. you need to be able to connect all the dots of all data, not just pictures. you need to connect text and sound processing to really understand sarcasm. to develop a hierarchical system that combines pictures and sound and text and video, you need to be able to render approximations of the sampled data in 3d space, which gets simplified and categorized and understood more and more later through reflection and additional associations to new renderings.
Soon, computer will warn "be careful, your child will fall from his chair be cause he is too excited about this cake"!
Then humans brain will stop to learn by theirselves. Tuxun, 2061.
make the "your child will fall from his chair" part optional then
but our brain have to succeed by themselves, its the key of learning
I'd really love to hear more about the interesting exchanges leading to the moment of 8:28
Cats will need wearable holograms to prevent these algorithms from stalking them.
the algo used to identify vehicles can help law enforcement find a stolen vehicle fairly quick, provided there are cameras on every street corner.
Skynet impending.
This is a great milestone. There is even greater challenge we will face down this road. Human intelligence is relative to human perception. Computers don't have that...
I think it's more efficient, when the AI has 3D models of this objects.
not possible with a 2d image
there are algorithms to estimate 3d shape by a 2d shape, there are some examples on youtube
+KonstantinGeist probably too much processing for a mass-collection system
The solution here can recognize 3D model with one camera... (good point Noah!) and without have to learn light rendering... if it can make link to "cat", it can already map to a 3D cat if you need, and maybe find how he is curled up (as Konstantin said).
過去9年間で驚異的な進歩がありました
Yep, if the Internet isnt full of cat pictures, nothing is!
inspiring to watch the starting point of the journey. imagine what the world will be in a decade.
Oh god, we'll have a "humanoid" by the end of 2039. It's see-able
I sure hope so! it would be awesome!
Johan Johansson I'd love to meet him/her.
I wish they'd be as sarcastic as I am (if a robot is capable of sarcasm)
john smith Pretty sure Japan already have that. At least i remember seeing handjob robots.. tho i don't know if i would trust someone with iron fists and steel muscles with my precious.
john smith haha nice answer ;D
Well, that's a great accomplishment so far. To learn the network network so that it can think like a 13 year old, you will be going to need million times your present dataset. It sounds good when you say the further accomplishments to be made. But it needs more efficient algorithms than the present ones.
come on, we all know what this technology is really going to be used for 1) mass surveillance 2) targeted advertising
It will be used for that, sure, but it will be used for other things as well. Technology isn't inherently good or evil. Machine vision is useful for everything from smarter image searches to robots that can autonomously navigate and interact with our environment.
Hail hydra
"You are being watched. The government has a secret system, a machine that spies on you every hour of every day. I know because I built it. I designed the machine to detect acts of terror but it sees everything. Violent crimes involving ordinary people, people like you. Crimes the government considered "irrelevant." They wouldn't act, so I decided I would. But I needed a partner, someone with the skills to intervene. Hunted by the authorities, we work in secret. You'll never find us, but victim or perpetrator, if your number's up... we'll find you."
Then maybe you can hire ...The A Team
Lol true! Especially with a scary naming like ImageNet (~SkyNet)
you are so smart. What you said has become fact.
Respect to the resesrch done! Definitely makes me go into academics and contribute to the beautiful field of Machine Vision
It's just deep learning.
Only thing kept me listening is that Big brother not only watches but also alert or even save a life
So what I'm hearing is that computers are actually stupid
They are as smart as their creators.
1. Why is it critical to ensure that our machines are capable of 'seeing' and 'understanding' things and events?
2. What are the challenges and issues do you think are present in developing such systems?
3. What are the potential positive and negative impacts of making highly accurate and efficient vision systems?
4. Should big-data and AI-based vision systems be subject to stricter laws and regulations? Explain.
BIG DATA marriages MACHINE LEARNING.
#goosebumps wheh you will know that 15 millions of pictures were held free for research purpose.
A big problem of ImageNET is that there isn't images for abstract entities.