Machine Learning Research & Interpreting Neural Networks
Вставка
- Опубліковано 18 лис 2024
- Machine learning and neural networks change how computers and humans interact, but they can be complicated to understand. In this episode of Coffee with a Googler, Laurence Moroney (@lmoroney) sits down with Christoper Olah (@ch402) from the Google Brain team to chat about Distill. Distill is a platform for interactive research, peer review, and community to make machine learning and neural networks more interpretable and accessible. Watch to get a deep dive into some of the research Distill focuses on, Lucid, Deep Dream, and more!
Visit the site to start diving into research → distill.pub
Read more about Feature Visualization → bit.ly/2JxzzBB
Read more about Interpretability → bit.ly/2LNs1HB
Watch more Coffee with a Googler → bit.ly/CoffeeWi...
Subscribe to the Google Developers channel → bit.ly/googledevs
I love his entusiasm while talking
He was great. I just popped a question, and sat back and let the goodness flow! :)
As a self-taught machine learning engineer who dropout of college this guy really inspired me.
This guy really loves what he's talking about, it's amazing to see his passion for ML and DL
Indeed! :)
Everything about this guy screams he's brilliant.
Christopher Olah is incredible. Thanks for posting this discussion!
What a wonderfully generous and effervescent intellect.
And Christopher's isn't bad either, right? :D
@@laurencemoroney655 hahahahha
Amazing person, wish we're friends.
Haha, me too! :)
It's amazing how i could use tensorflow to help me count how many "sort of" has said in the video.
You don't need tensoflow for that, simple regex on transcript would do.
So, that would be a sort-of AI? :)
After watching Christopher Olah, I'm more convinced to use ImplicitCAD.
What a great session that was. The more you can learn about how your networks are learning, the more you can optimize and innovate on them in the future.
What passion can do is amazing!
He's an amazing speaker. I'd love to work with him. Thanks for making this video
Welcome! :)
Chris, you're my new favorite person
Haha! Mine too! :)
He's like a hyper-caffeinated italian puppy! :D So great.
Fascinating project, by the way. We hear a lot about neural networks, ML and the like in abstract terms, but visualizing the thing working really helps when trying to understanding it beyond the level of 'and magic happens' :)
Absolutely! I came away from this conversation much smarter! :)
Alguien que no tenía ni idea de ésta historia ,se enriqueció de poesía. ...porque la música es una herramienta importantisima para muchas cosas. La más completa es ése "algo" que te cambia el humor ,pasas los recuerdos de misterio a real y viceversa.
LM: "...as soon as I get into mathematical notation...part of my brain begins to go to sleep...."
Thanks for saying that; reassuring...
Heads up, somebody stole all the pictures out of the frames.
Heheheh. It's getting it to wake up that's hard :)
Laurence makes a great point (apart from the fact that this is an awesome interview): Wish all the papers we publish had this mandate that there must be an online visualization tool/interface/notebooks which can be run over the web with minimal effort. Instead of the users going through 1000s of "bumps" trying to match the infrastructure to run 10 lines of code. So much for reproducibility argument. LOL. Good work both of you otherwise.
Thanks Mithun! :)
He is my spirit animal
Haha! Yes, he's my Patronus. :D
awesome interview
*THANK YOU* First time I see the Legend on TV
Why thank you! Oh wait, you mean Christopher. Well...still thank you! :)
So beautiful, unfortunately I can't bet like this guy. :(
I hope to work at Google, I love this company, (really ).
Jomman Media House I only understand Portuguese and English
E ai, Pedro. Acompanha essa comunidade de machine learning da gringa também?
Lucas Lopes Algumas vezes, é um tema bem interessante
It's a wonderful place full of wonderful people! Hope you can join someday.
What's your opinion of molten salt reactors?
Not sweet enough
Thanks for the guideline :)
its a hirarcical pattern recognizer just like it have been envioned in many science fiction movies, and is just like a holographic memory where parts of image on one level makes up bigger wholes on another.
I really loved this, I am as excited as he was now. Thank you
Does the Google Store by the HQ sell coffee mugs of Coffe with a Googler cause these video are my fav Google vids! ☕️
here is an idea. maybe this is what you guys are already doing in a sense. use one layer per target pattern and let each layer in paralell look at the same target data where each layer detact their own recognition of the same image and then use a layer higher up in the hirarchy to deceide what input layer that contain the closest match of the target image. this way, the computer chould be able to detect lots of objects in pictures with random content.
by using only one type of training data to detect objects in a picture is makes the program a liar as each result of each layers recognition must be weighted by a layer higher up in the hirarchy to deceide what is the closest match. the idea is that a program can more acurately find out the right version of reality finding the real object and not some imaginary simmilarity.
i was thinking if we could use such a method on pictures of rock formations and see if the computer was able to see the same as a human may have been able to see or better or if it see nothing.
maybe we could also do a cloud game where the neural net is trying to make out different objects in clouds or a image of noise but get different images of objects each time on the same noise pattern. this way we will know if it have real imagination.
i was also thinking what a dionsaur would have looked like in 3d if a sceleton was reconstructed into a animal with flesh on it by a neural net system using a variety of types of training data of sceletons and animals we know how looks like and let the computer reconstruct a version of what a dinonsaurs may have looked like.
another thing is to let such a software try to recontruct animals we know how looks like but by letting the computer do its try to replicate the animal the way we know it looks like. am sure that the neural net will fail at first and create a monster that looks more like an alien than the target animal. i was thinking if it could be able to reconstruct a animal in 3d unsupervised by using a variety of training data from different animals and sceletons and let the software make the closest guess to replicate the target model or image.
im thinking that if a neural net become a painter, it may not paint pictures the same way as a target image if it have to draw the copy on the screen instead of just identifying it.
i wish from this comment that someone would want to use my ideas even if its not really much different than what google is currently doing.
example google sketchup is recognizing what a person is drawing i would like to do the reverse. i would like to show the progam a drawing and let the program redraw the image as close to the original and see if its able to reproduce the original by using a variety of type of trianing data of different objects to mimic the target object. if the computer is guessing wrong and create somthing unrecognizable, it is likely that its far from detecting the objects the way we want to belive. i was thinking this could be a good way to debug how the computer recognize objects by trying to imitate the target object. let say you show the computer a triangle and then let it use a paint tool to repaint the triangle but do not let it copy it by training but let it do by what it already know base on pre-exsisting training data but in a hirarchy. this is the only way to know if the machine see what we see or if it see somthing totally different.
god his zest for his craft. Amicable
Great interview format. Very charming.
Thanks! :)
Wonderful. Thanks.
:)
amazing
Nice video
Great.
Everything is best except this.
1. Decrease the hand gesture. It makes me confusing.
2. Try not to use "You know".
3. Try to speak slowly. People who not use English as first language also watching this.
Sorry! It's really hard when we get enthusiastic to slow down. Will bear that in mind for future videos. Kamsa Hamnida! :)
@@laurencemoroney655 I don't understand his complaining. You did a phenomenal job, Laurence. You let the guy speak and didn't interrupt at any point. That's more than 99% of the interviewers that I have heard. Good work.
I sort of agree. There could be a slower way of explaining things.
But then again everyone has a different way of explaining things. And everyone likes different styles.
In general its not easy to catch up with the smart people.
Soon .... 2027 !