Why Computer Vision Is a Hard Problem for AI

Computer scientist Alexei Efros suffers from poor eyesight, but this has hardly been a professional setback. It's helped him understand how computers can learn to see. At the Berkeley Artificial Intelligence Research Lab (BAIR), Efros combines massive online data sets with machine learning algorithms to understand, model and re-create the visual world. His work is used in iPhones, Adobe Photoshop, self-driving car technology, and robotics. In 2016, the Association for Computing Machinery awarded him its Prize in Computing for his work creating realistic synthetic images, calling him an “image alchemist.” In this video, Efros talks about the challenges and changing paradigms of computer vision for AI.

00:00 Why vision is a hard problem
1:18 History of computer vision
2:01 Alexei's scientific superpower
3:14 The role of large-scale data
3:37 Computer vision in the Berkeley Artificial Intelligence Lab
4:15 The drawbacks of supervised learning
4:57 Self-supervised learning
5:33 Test-time training
7:08 The future of computer vision

[weinhardtadam]

I love that with 120.000 citations, he is regarding the grad students and the next generation of scientists as his biggest achievement.

[Alex-rhjo]

It's great that there are professors out there that value their students as their greatest achievement!

[joaoguerreiro]

As a computer scientist working in Computer Vision tasks (and other AI applications) for medical imaging processing, this video made me smile :)

[brianfunt]

I love how at 8:08 one of the students' phone falls out of their pocket and everyone turns and looks at it

[MiniLuv-]

All very interesting. I wonder if we are limiting computer vision by only considering human vision. Each other organism has vision selected to make the organism successful, and its not like ours. I wonder if there is something we can learn from this diversity of purpose for visual systems in all organisms. Alexei Efros has touched on this diversity of purpose with his own experience of vision.

[werwardas]

Thank you for the insights and this very well produced video!

[JZFeser]

Wonderful video! I love everything this channel has made!

[presence]

I had an idea when I was working on my thesis that if we have transformer for vision and a new embedding system that treat the visual data like human we can have a model that will understand the images of the universe that is beyond the computer ability of human brains such as the cosmic microwave background. But it’s an idea only😢

[.]

AI generated timestamps
0:00: 👁 Computer vision is a complex process that is difficult for computers to replicate, but advancements are being made.
2:56: 🌳 Visual data and its importance in machine learning and computer vision.
5:58: 🔑 Computers struggle to generalize in their machine learning algorithms, but test time training can help improve their performance.

[lilhaxxor]

This is a very good interview. I am glad to see that it's validating my intuition, about the fact that models should continuously learn instead to being frozen, and then retrained from scratch.

One of the biggest difficulties to improve the current techniques is reducing models size. I don't know how much data a real brain can store, but given the miniaturization of current chips, I suspect we are wasting resouces.

Anecdote: I have bad eyesight as well. 😂

[MichaelFergusonVideos]

Wonderful! Looking forward to the future!

[JayAyers]

1:55 Im not sure if showing birds was intentional here. In any case, it looks like a nod Gregor Mendel's genetics work.

[aldev]

Man.. I wish you were my CS professor. 👍

[johnyharris]

The next big step will be generalisation. For example, when AI can infer from its training data generally what a road looks like whether it has snow, leaves or sand on it.

[Fine_Mouche]

what about use analogue computing in the futur for AI ?

[timgabby]

Interesting to see the distribution of ethnicities along that outside shot bench.. humans are drawn to those with whom they assume they might have common ground. Just an observation. Might be wrong.

[autonomous_collective]

Computer scientist Alexei Efros suffers from poor eyesight, but this has hardly been a professional setback. It's helped him understand how computers can learn to see.

At the Berkeley Artificial Intelligence Research Lab, Efros combines massive online data sets with machine learning algorithms to understand, model and re-create the visual world. His work is used in iPhones, Adobe Photoshop, self-driving car technology, and robotics. In 2016, the Association for Computing Machinery awarded him its Prize in Computing for his work creating realistic synthetic images, calling him an “image alchemist.”

In this video, Efros talks about the challenges and changing paradigms of computer vision.

[_soundwave_]

5:28 he is so deep inside, he calls us 'agents'

[bharatjoshi]

So AI is just data with some selective results from that data ..is it ?

[ElParacletoPodcast]

Computers cannot see, and will never see, they only process information, but will never see.