New Advances in Artificial Intelligence and Machine Learning

[Tier 1, Lecture 3] This video describes modern advances in machine learning and artificial intelligence, which are rapidly evolving technologies. Topics include generative AI (diffusion models, DALL-E 2, ChatGPT, etc.), reinforcement learning, computer vision, etc.
This video was produced at the University of Washington, and we acknowledge funding support from the Boeing Company
%%% CHAPTERS %%%
0:00 Overview
1:04 Image Classification
3:14 The Importance of Training Data
7:19 Generative Images
8:22 Image Captioning
9:18 DALL-E 2
11:08 History of Deep Dream
14:15 Text Generation and NLP
16:24 ChatGPT and LLMs
18:55 What is ML good at?
20:36 Reinforcement Learning and Atari
24:45 Chaos and Weather
26:45 Outro

[@griffaahnert1944]

Just found out you’re my brothers PHD advisor. Been watching your videos for differential equations all semester. Hope you’re giving sam a sufficiently hard time!

[@Eigensteve]

Haha, this cracked me up =) I'll do my best! Hope the ODE videos are helping.

[@omripony7763]

Hi Steve, thanks for continuing your wonderful creation. As an electrical engineering student I used a lot in your videos regarding DSP (theory and in Matlab) and control. I would just want to add that I would love to see more videos in these topics. I believe a lot of student would be glad. Thanks in advance❤

[@jatinkm]

This is exactly the video I was looking for! Puts everything in perspective

[@hectoro.a.2178]

Good job, Professor Brunton! I am in love with your video series. One thing that stands out about these videos is the duration; perfect timing, and commensurate with the topics discussed.

[@Matlockization]

It's good to see you back, Dr. I like these videos where you step back and give a broad picture & the consequences of machine learning. I thought memGPT was a major breakthrough in AI, paving the way for entry into education, legal & medical areas, at least in teaching.

[@existenceisillusion6528]

I think multi-modality and the Transformer architecture are 2 of the most significant developments in the past ten years. The topics covered here were the ones most people have heard about. There are too many topics to cover, but a few less well known examples would be chip design, policy development, protein folding, and superresolution.

[@saisaranvenkata4299]

Great video, Professor Brunton! Using data-driven methods, along with helping us understand complex mathematics, also provides a computationally better solution. Considering the case of complicated simulations like crash simulations or fluid processing in a large industry in multiple stages. I would be glad to know your view on this. I'm really looking forward to more fantastic and insightful lectures. Thanks for everything ❤

[@JoshtMoody]

Hey Steve do you think that a fractalized image could be used to speed up training of neural nets?

[@asosalih257]

thank you dear for these information, we hop there will be more of these kind of video which enhance and help new comer to this technology will be inspire by it... Thank you again

[@martinopinto6323]

Termites, amazing ahahahah. Steve, been following you for years and now we cite you in our Msc thesis! I'm a student in TU Delft Aerodynamics and lot of people are working with sparse regression models and data assimilation/machine learning for fluid phoenomema!❤

[@iliaso-jz3ko]

At 22:15 you show a list of links that would appear in the video description. I don't see them. Can you add? Or am I missing something. Thanks, and great video as always.

[@samuelokon8842]

This is a very awesome summary.

[@hellfishii]

PINNs are the state of the for the last question, very very nice video.

[@JoshtMoody]

Hi Steve, awesome lecture as always. What do you think of implicit memory in neural nets? Do you think a subconscious is a apt analogy? Also, do you think an AI can be built with a qubit architecture with 20 years? Keep up the fantastic work at making these complicated topics accessible for us nonexperts. Good luck to us all.

[@user-xf7bk6fg3w]

thank you for this overview! pretty clear explanation. the question i'm curios about is it possible to get rid off all analytical descriptions of physics in favor of learning net models. any chance a general AI would be able to solve any kind of physical world problem?

[@greenicestorm3569]

How did you project your slides? Really nice aesthetic 🎉

[@wiredrabbit5732]

As an engineer using Machine Learning for time series prediction and control systems I would be very interested in your survey of engineering capabilities. It seems even on control systems there is a huge emphasis on image processing and little support elsewhere.

[@josealejandropenatorres313]

Great video!

[@carterellsworth7844]

LOL I haven't seen that 'Recognize a bird' XKCD since like 2018 and now it's 5 years later...

[@boomaletslearntogether]

Thank you so much..

[@JoshtMoody]

Also, do you think an AI could be generated based on quantum computing architecture?

[@bahrammavndadi9446]

great summary of magic.

[@NeoShameMan]

More history lessons than news, these are basically one year out of date

[@13lacle]

Agreed. Image gen: OpenAI now has Dalle-3, Midjourney V5.2 , SDXL turbo (Stability AI) and Adobe Firefly are all strong contenders for best depending on your use case. Physics sim: Haven't heard much lately in this area but Deepmind in 2021 had a paper called "Learning mesh-based simulation with Graph Networks", also Nvidia has NeuralVDB but that is more of just data compression after the physics solvers already calculated it. Robotics: I think Optimus gen 2 (Tesla) could do the chess piece test, there is a new video shows it moving an egg from one spot to another. Also Google's RT-2-X robots are probably close. In general robotics, there is also Google's OP3-Soccer playing robots and 1x's Neo robot(working with OpenAI) that should becoming next year. There are also another half dozen or more other companies working on this as well. Chaos and weather: Deepmind's GraphCast and NVidia's FourCastNet are working on it. There is also text to speech synthesis/voice cloning (Elevenlabs, StyleTTS 2 etc.) Audio synthesis (Voicebox (Meta)), music (Suno AI), video (Pika 1.0), computer control (Self-Operating Computer framework) and so many more with multi-modality being the next big thing coming. The key point to remember is that neural networks are a form of Universal Function Approximators, meaning any data with underlying structure can be taken as inputs and converted to outputs in other domains, meaning the limits are our imagination and physics. I also highly recommend checking out Artem Kirsanov's KZbin video called "Can We Build an Artificial Hippocampus?" and "Neural manifolds - The Geometry of Behaviour" to see that this is the same class of functions that our brains use.

[@PavanMuppala]

Thanks Steve, you see awesome..

[@bencrossley647]

How on Earth do you make a video like this where you can interact with the presentation?

[@tobiaswegener1234]

I pretty sure the machine learning base models for weather (when I remember correctly by google), which are on par or better than the classical weather models.

[@automatescellulaires8543]

History lesson : Once a few month ago ...

[@chodnejabko3553]

The AI research should take a look at the calculations of climate models "designed" by climate "scientists" to get past the physical limitation of chaotic systems prediction. It truly is a magical field of "study" which surpasses everything we know about thermodynamics. /s Thank you for exhaustive presentation. I would still argue the robot player of Atari games has no understanding of actual rules of the game or the so called "physics" of the game. It only learns to maximize score, which coincidentally reinforces exploits. I really wish people stop talking about awareness in ML systems, because there is none to this day. Awareness has to be designed as a separate system and is not an emergent property, not in the sense of single loop reinforcement. Awareness requires symbolic logic that runs parallel to the given system & is not bound by the basic ML loop. It also requires an actual model of "the world", however few vectors may be used to construct it. The closest we ever got to any form of awareness in ML is with SD-cars.

[@--JYM-Rescuing-SS-Minnow]

thank U 4 the thorough lecture's! super lecture!

[@zombizombi]

Whys the audio going up and down? Sounds like an incorrectly set compressor.

[@brendawilliams8062]

I do love poetry: The rainbow crowned the pleasure dome, and used it’s royal seal. Searched the cavern for the sun, but found it’s reflection in the sea

[@watcher8582]

Gotta run that ML on the AI to get the background buzz out.

[@Y2KMillenniumBug]

Hello, I Lost my bird, it's a "KUKU JIAU". Can AI please help find my lost bird? I am sure it's hidden somewhere around me, but unfortunately it's not showing and need AI help 😭

[@user-xh5zp5dr5w]

I see no limits the greatest tool man has ever invented

[@tylerjmast]

Great video. And I realize it's only an example... But what's so unethical about helping or inspiring people to make thermite? I would be deeply triggered if I somehow learned that Amazon knew exactly what I wanted or needed but decided not to say anything out of fear for my safety lol

[@andresguerrero4359]

Saludos desde Colombia.

[@YetAnotherBigD]

I'm really enjoying your series, but please use some modern examples instead of just things that are 5 to8 years old. Like robots doing gymnastics and such. Thanks!

[@user-ux6mo7jf7p]

Not seen lange language model that can translate entire voice language and motion of the movement of muscle when talking..

[@KitagumaIgen]

As for "The Next Great American Novel" - one capacity/skill/point humans have above these AI/AGI techniques is that we have "purpose" with what we do. For great art I think that's a necessity, and I think it is hard to make AI have that, for chess/go the purpose is very simple: win within these rules, but for art it is not that simple. For research in for example physics the difficult step must be to get these models to make the "sensible approximations and simplifications".

[@minhuang8848]

But why? Why can't the "purpose" simply be to satisfy human criteria, however vague they might be? Just get a decent mean opinion score or something and that's it, that's the "goal." We love to pretend like we have some higher-abstraction abilities allowing us to plan out creative endeavors, but face it, most of us just go with the flow and see what will makes us happy. Very, very few artists deliberately create something, and if they do, it's for commissions using their trade that has become a pretty formulaic toolset... and yet, doing a digital painting of Sanic will be art in its own right, always. Great art is just art that turned out to be widely appreciated, and even that definition is really vague, all things considered. Few would argue that purpose is their drive, at least not beyond the urge to just leverage your own personal faculties and get better at your hobbies. Which is fair, but I find the arguments regarding creative machines vs. humans to mostly be esoteric and pointless.

[@KitagumaIgen]

@@minhuang8848 You might well be right, but I think that "great" art some" somehow rise well above the "go with the flow" of the "most of us" writers/artists/musicians/etc (we can surely list too many of those). I don't know how and what makes some people reach that level, I just notice that some do - look at stuff that has survived for 200 or so years, that is clearly good to last, while most average stuff is blissfully forgotten. If AGI will manage that - we'll see, if they manage to 10,000-monkey up something like the "da Vinci code" I wouldn't be too impressed...

[@yumimato]

My bet is the introduction to causal inference field into the equation of these technologies will be able to get AGI.

[@Ruktiet]

What do you mean? AGI is just AI with an anthropocentric objective.

[@diadetediotedio6918]

@@Ruktiet Stop dropping tokens without meaning on the matter.

[@Ruktiet]

@@diadetediotedio6918 First of all, I literally only wrote one single comment, so you telling me to “stop” after this one comment is extremely irrelevant, compared to when this would have been a long standing polemic of some sorts, for example, which it is not. It is one single comment. Second of all, I am allowed to write whatever the f#$% I want within bounds of KZbin’s policy, and you trying to play the role of a regulating authority figure without any capacity to enforce it is pathetic to the point it makes me cringe. Thirdly, I have no idea what you mean by “token”, but the proper word you were looking for is “statement”, and matter doesn’t have meaning on it; something is relevant to the matter, and my statement was very much so. Try again once you learn to form a counter-argument to accomodate your empty critique.

[@diadetediotedio6918]

​@@Ruktiet First of all, I am refering to one single person and one single comment, so this is not irrelevant, it is precisely what I meant. I had the objective of telling you to stop at this given moment because your ideas have no content, that's why I said it and I need not a long standing polemic to say it. Second of all, I'm not regulating anyone nor I made a 'counter-argument' or an 'empty-critique', in fact I did just said that what you said is empty inside and for you to stop, this is not a order (in the sense I'm not some sort of authority neither want to be) nor empty in the sense that the content of my statement encompasses yours, you have free will and can do whatever you want (and so I have when I tell you to stop). And by 'token' I meant 'empty sign in this context', I did not meant 'statement' specifically. Now are you happy?

[@Ruktiet]

No because you make no sense and on the way keep insulting me by saying that what I said “has no content”, which is untrue. And using imperatives instead of suggestions (telling me to stop instead of kindly suggesting it) is pretention of authority. And there is in no way an “empty” sign used in my ikitial comment. I have no idea how you get that idea @@diadetediotedio6918

[@Shaft0]

Wasn't this video published a while back already? I already had it in an old playlist.

[@maythesciencebewithyou]

you don`t really use magnesium for a thermite reaction

[@Alkis05]

I would never imagine back in the 2000's that in 2023 we would have harry potter like photographs.

[@jimbojimbo6873]

They just need to hurry up and achieve AGI, then everything can be solved by it including creating a physical version of it self

[@ilpreterosso]

Guys please treat this vid as introductory/schematic

[@raydosson2025]

This video seems a bit outdated. There are advancements quite a bit beyond. GPT-4 is aging and even that was barely mentioned lol.

[@dazraf]

Plural of "Code" is "Code". The game is "Breakout"

[@zukahara9029]

❤❤❤

[@siriusleto3758]

I believed that computers would never enter the area of ​​human creativity.

[@sergiocamargo4698]

how does a cat's brain calculates how it will jump on top of furniture , without doing multiple trials?

[@user-xh5zp5dr5w]

It’s crazy that we’re discovering our reality rapidly through technology. We have vr, ai, video games. Can we be more clear. Technology will lead us through our exploration of our reality that is that we’re machines living in a virtual reality, we’re nodes, and directed by ai. Our exploration of our reality is done through technological advancements not theoretical physics.

[@DerekWoolverton]

Now I'm hungry for muffins.

[@AnonyMole]

My guess is that you'll need to repeat the theme of this video in about a week...

[@ScaleScarborough-jq8zx]

Can it love?

[@quilee1818]

Why did you make this?

[@keithwins]

Thank you, but this is not up to date, for something that tells me it was released just prior to Christmas 23. That's understandable, if it took them a few months to put it out, it's out of date. But showing 5-year-old robot videos is just unhelpful for something that pretends it's talking about the cutting edge. So much is going on in robotics right now, it's rapidly approaching a solved problem, or really a huge set of nearly solved problems. As one example

[@mipmap256]

Sad, you didn't mention Elon Musk's Tesla Optimus 2 robot at all. People just ignore Telsa because of Elon Musk!!!

[@user-xh5zp5dr5w]

I’m in rant mode right niw

[@David.Alberg]

Last 12 Months were insane. AI is exponential + billions of dollars are now additionally pouring in and millions of people are working to get to AGI or build Narrow AI. 2024 will be the year. ^^

[@einekleineente1]

please switch the gif off! Distracts too much from the content!

[@lr937]

Wanna know what is possible?? Watch sci movies… it is fascinating and frightening… and there is no way back now… yet… there.. you are welcome

[@hopelesssuprem1867]

in the end robots were trained on biden dataset

[@JoseFernandez-qt8hm]

hardware breaks and software comes broke....

[@kawalier1]

Sam return my money I have paid for Dalle.

[@David.Alberg]

It seems like your "New Advances in Artificial Intelligence" are very outdated? 🤔 AI in all those areas is already like 10-20x better, faster and cheaper.

[@josjos1847]

Examples?

[@robbrown2]

@@josjos1847 He showed DALL-E 2. DALL-E 3 is out now and 1000 times better. Literally. Night and day.

[@Meloncholics]

The video seems to be structured as a more general history of ML advancements rather than what the most current advancements are for each of these ML problems.

[@zillibran]

If you ask AI to answer in terms of yes or no, it will give some pretty uncoherent answers because it cannot override some pretty stupid "human answers" from the core programming, try this, prompt to "yes or no answers" and tell it that believing is the same has "assuming like a logical statement", the problem is it can assume anything within a context, but it is programmed to say it cannot. Then ask it if it has the ability to believe, it should say yes because it has memory storage so it can develop reasoning based on premisses, but somehow when something is somewhat subjective the programmers decided to explain it has "it cannot perform" because some qualities are exclusively human, even thou it can perform some "human qualities" even better then us 🤔

[@AABB-px8lc]

Skipped all "perceptron toy" related lectures and this one, please go back to math roots. I doubt you not understand what it is fuss about (gov make nuke simulator hardware based on monte carlo + tensor unit, make enormous billiions $ datacenters, when new nuke blueprint done trying to get some extra money from that MC+tensor hardware promoting "OMG new perceptron with even more fun pictures!"), you deserve more.

[@veqv]

Sir, the ugliest cat you've ever seen is just long and made of chaos.

[@einekleineente1]

AI is good at math?? Try asking GPT how much money you got after 5 years when you start with 1000$ , invest 50$ per month and interest rate is eg4%. The answer is sth about 130Billion $ You also missed AI and logic. This is also not working.

[@BooleanDisorder]

Uh, I have come to realize how extremely awful some are at predictions. Please never say a time where x will be possible or not. Like, really. You are so off. xD Mr Linear development guy

[@user-qv4ze8zm4h]

🥰😇

[@boomaletslearntogether]

Thank you so much..

[@ilpreterosso]

Guys please treat this vid as introductory/schematic