loved this video, especially all the illustrations! 5:00 i know it's a simplification, but if we can turn the knob to make things better, then why don't we automatically get super-human performance on all tasks?
Thanks Scott! For anyone who's not Scott who's reading this, Scott is my good friend who went to Berkeley, also studied AI, and was in the same AI research lab as me! The question he's asking makes a really good point: AI systems don't become perfect just by following this error-reducing method because 1) our algorithms are imperfect and 2) our data is imperfect. re: 1) our error function, or "loss," is often bumpy and full of hills and valleys. Sometimes it has deep wells, where algorithms get stuck because the error is higher all around them. But these deep wells can prevent the algorithms from finding even lower error states somewhere else in the landscape. We call those local minima. re: 2) our error function is always just an estimate of the true error function -- it's the average error that we measure over the data we've collected. As our dataset gets larger, our estimate gets closer to the real, unknowable error function -- but it can never be perfect. Scott, curious to hear what you would add to this!
@@maximejkb yeah that's pretty much what i had in mind! just thought it would be good to point out the caveat that we can't solve everything with optimization/learning out of the box -- and that "solving" in the optimization often isn't the same as "solving" a real world problem. 1 and 2 are both big areas of work, and ppl go to great lengths to make sure the data you learn over is representative of the real world. so that might mean making sure your cat/dog classifier isn't learning off of 99% dog images, since then it could get 99% accuracy by saying "dog" every time.
Everything about this video was so well done! 👏 Thanks so much for making it. You explained the foundational math so eloquently. How does this video not have millions of views? The RUclips algorithm is clearly stuck at a local minimum. 😅
This video deserves to have a lot more views. Your explanation is clear, straightforward, organised, and brilliant. I look forward to your future uploads man.
Great video, man! I think I remember a good example for unsupervised learning from an online caltech lecture - coins that you push in to the coffee machines will have small differences in size or mass, because of dirt, but they will cluster closely, and the a.i., in theory, will recognize these clusters without having any labels.
you described A.I. very well. i think this is the comprehensive roadmap of A.I. i ever seen . i contacted you in your instagram if you want to reply on it
excellent video. you are very good at explaining things in a way i can understand. the drawings are cool and helpful. i look forward to seeing more videos from you.
Really liked the content of the video. I'm also interested in how you made the illustrations. What graphic tablet are you using? I tried to spot the model from the video by zomming in but it's quite blurry. I do illustration for teaching/learning purposes and this tool seems quite neat. Thanks
Great vid! Entering computer science program in a few months:) Thank you for the overview! PS: loved the illustrations. What program do you use? Ever considered becoming an artist?:)
![The moment we stopped understanding AI [AlexNet]](http://i.ytimg.com/vi/UZDiGooFs54/mqdefault.jpg)
![The moment we stopped understanding AI [AlexNet]](/img/tr.png)
loved this video, especially all the illustrations!
5:00 i know it's a simplification, but if we can turn the knob to make things better, then why don't we automatically get super-human performance on all tasks?
Thanks Scott! For anyone who's not Scott who's reading this, Scott is my good friend who went to Berkeley, also studied AI, and was in the same AI research lab as me! The question he's asking makes a really good point: AI systems don't become perfect just by following this error-reducing method because 1) our algorithms are imperfect and 2) our data is imperfect.
re: 1) our error function, or "loss," is often bumpy and full of hills and valleys. Sometimes it has deep wells, where algorithms get stuck because the error is higher all around them. But these deep wells can prevent the algorithms from finding even lower error states somewhere else in the landscape. We call those local minima.
re: 2) our error function is always just an estimate of the true error function -- it's the average error that we measure over the data we've collected. As our dataset gets larger, our estimate gets closer to the real, unknowable error function -- but it can never be perfect.
Scott, curious to hear what you would add to this!
@@maximejkb yeah that's pretty much what i had in mind! just thought it would be good to point out the caveat that we can't solve everything with optimization/learning out of the box -- and that "solving" in the optimization often isn't the same as "solving" a real world problem. 1 and 2 are both big areas of work, and ppl go to great lengths to make sure the data you learn over is representative of the real world. so that might mean making sure your cat/dog classifier isn't learning off of 99% dog images, since then it could get 99% accuracy by saying "dog" every time.
As an AI Researcher who watches a lot of RUclips videos about AI, I must say I’m highly impressed by the clarity and quality of this video.
Let’s go! Appreciate that!
Everything about this video was so well done! 👏 Thanks so much for making it. You explained the foundational math so eloquently.
How does this video not have millions of views? The RUclips algorithm is clearly stuck at a local minimum. 😅
highly underrated video
As an ai engineer, this video is way better than i expected. Kudos
Thanks for watching and glad you like it!
Thanks everyone for watching - hope this was useful! Let me know what you think in the comments!
@maximejkb really liking this channel! Where you at max ?
This video deserves to have a lot more views. Your explanation is clear, straightforward, organised, and brilliant. I look forward to your future uploads man.
Great video, man! I think I remember a good example for unsupervised learning from an online caltech lecture - coins that you push in to the coffee machines will have small differences in size or mass, because of dirt, but they will cluster closely, and the a.i., in theory, will recognize these clusters without having any labels.
Thanks for watching! That's an interesting analogy, I'll have to find that lecture
you described A.I. very well. i think this is the comprehensive roadmap of A.I. i ever seen . i contacted you in your instagram if you want to reply on it
Thanks for watching -- tried to make it as exhaustive as possible!
Amazing vid man! I do appreciate the way you explain and those visuals are perfectly fit for your style. Keep going! You got my sub👍
Thanks so much! Glad the visuals helped :)
High quality content.
Hope you get more views
Appreciate that, thanks for watching!
Awesome video! I think i would include predictive modeling / forecasting (banking) as a method or application.
Great high quality video ! You deserve more views.
Thanks!! Appreciate you for watching!
This video is amazing.
Accurate content, Clear explanation, enjoyable aesthetic (animations and you’ve got a good voice bro)
Thanks for sharing! 🫡
This needs to go viral
Cool, really clear and straight to the point!
Would it be possible to make the "infographic" available to download Maxime ? Excellent video too!
great job. very articulate
Thanks so much!
excellent video. you are very good at explaining things in a way i can understand. the drawings are cool and helpful. i look forward to seeing more videos from you.
we are waiting for you after that week man. Your week is passing quite slowly it seems
Extremely well made video! Do you have any plans on releasing the notes or map?
Really liked the content of the video. I'm also interested in how you made the illustrations. What graphic tablet are you using? I tried to spot the model from the video by zomming in but it's quite blurry. I do illustration for teaching/learning purposes and this tool seems quite neat.
Thanks
Amazing explanation!
Great vid! Entering computer science program in a few months:) Thank you for the overview!
PS: loved the illustrations. What program do you use? Ever considered becoming an artist?:)
this is an amazing vid!! can't believe it's only 2k views!
Awesome job bud! Thanks :)
Great video and introduction to AI. Highly intuitive. Did AI help create it? If so, Well Done!
I appreciate that
thanks for this
Great video :)
Really glad you enjoyed it!
Thanks a lot, it was amazing👏
Great Video!
awesome !!!!
This video is insanely good! You’re doing God’s work here, thank you so much 😊
Thanks for watching!
🤩🤩🤩
Yo Max. I’m just getting into this space. I tagged you on socials let’s network