Deep Learning for Computer Vision (Andrej Karpathy, OpenAI)

[@jiachenglee2554]

slides: docs.google.com/presentation/d/1Q1CmVVnjVJM_9CDk3B8Y6MWCavZOtiKmOLQ0XB7s9Vg/edit#slide=id.p

[@TheFireHacker]

David Lee Thanks for sharing deck :)

[@sidthegoodboy]

Thanks! :)

[@incendioraven4269]

= = I have a feeling that I accelerate the video.....since he speaks so fast

[@macevsculture]

@@Penaming 1.25 speed is even better

[@tacitlytesla7912]

He just speaks very staccato, I think he speaks normal speed

[@grlldfsh123]

I love listening to things I cannot understand

[@berargumen2390]

Keep listening, you will eventually understand

[@alvinphantomhive3794]

makes u feel smart uh? cause i felt it that way too 😂

[@colinmaharaj]

Don't say that, give it time.

[@anuraghivare5875]

Same here.....😂😂🤦‍♂️

[@HemangJoshi]

For real bro . For real... 🤣🤣🤣🤣

[@rishabhpatra4950]

Anybody just love the way Karpathy speaks? Makes me want to listen to him with intent

[@MohammedYounis-w6h]

same, he just gets to the meat of the stuff right away

[@colinmaharaj]

Interesting you'd say that, there is so much to learn even from asian schools like in India for example, but some accents can be harse when you watch the videos. Whats interesting about what you said is I think Andrew Ng also speaks eloquently and his lectures are worthy of a listen, and its the same topic area.

[@endgamefond]

Agreed. Idk if it is the nervous pause or else but he is so interesting when he explains stuff.

[@marciovenanciobatista]

Hi from Brazil!! Thank you for sharing this. It is a really good material for researchers and who is initiating in this area.

[@PaAGadirajuSanjayVarma]

Thank you lex sir for posting this

[@gue2212]

Hey Lex, can´t thank you enough for splitting up the day-long streams! Much easier to consume -- as I wanted to download to enjoy it on mobile! I came across Andrej Karpathy´s Deep Learning for Computer Vision yesterday. I´ve been trying to really understand CNNs and the Deep Learning paradigms technically for some time now and sucked up everything I could since July. Andrej´s lecture is the very best I found to get up to speed on most important details in the least possible time. Cheers G.

[@MawaMaverick]

Yest i watched it 2 days before i watched this one as well. Andrej Karpathy's lecture is absolutely amazing. He covers so much in such a short time with such detail.

[@livesportshub]

This People, They did not understand anything but they liked him.

[@karanpatel1419]

How did this went under my radar for 7+ years

[@sanchitsinghiitg]

omg! He has done a great job in explaining this entirely new field of research in an hour!

[@TheFireHacker]

Great video. I came here after completing 4 video sessions by Lex (MIT 6:S094). Thanks for compling these videos.

[@bingeltube]

Very recommendable! Good overview!

[@russianmadman1]

This is a great talk, really helped with my university project

[@maged_helmy]

Does someone have 2016, 2017, 2018, and 2019 winning architectures list?

[@gibrains]

This why youtube has playback speed settings.

[@rui569]

🤓

[@ridhakiller]

Amazing

[@gue2212]

Extra comment, so the important link is visible without expanding: I´m looking forward to dig into the extended version, namely the CS231n Winter 2016 Lecture series referenced in the video. (at kzbin.info/www/bejne/nV6zp4uLn6lpp8k )

[@thesezeshacks9163]

Ok I disagree with how he answered the woman's question. There is a way to choose a MINIMAL amount of layers, based on the complexity of the data. It depends on convexity vs non-convexity in the data space. non-convexity requires a minimum of 2 hidden layers to represent. This is why before you dig into deep learning and all the modern stuff a good basis in Universal Approximation Theory and NN's as Universal Approximators is critical!

[@offchan]

27:02 How are the gradients computed? 1) Gradients of the loss with respect to the activation 2) Gradients of the mean/sum of activations with respect to the input image

[@maosun7474]

I think it compute the lost in the last layer and do the back propagation to compute the gradient of each layer so as to optimize.

[@offchan]

That is choice number 1.

[@1Kapachow1]

Great lecture - thanks! :)

[@maqboolfida786]

The first terminators on the planet would sound exactly like Andrej Karpathy!... love to see him say 'I'll be back' :)

[@QuintinMassey]

So amazing...

[@xDMrGarrison]

that deepvis demo was kinda magical and mindblowing..

[@aydtsang]

does anyone know why lex hasn't had karpathy on the podcast yet?

[@MahJohn]

I ask myself this question as well. Just wondering!

[@Splish_Splash]

so.. know he had

[@ahmadkarim8632]

can anyone load the slides of this presentation

[@dciug]

At 1:19:40. Isn't that what Google is trying to do right now?

[@disectormusic]

he is one of the wingmakers

[@al7ay]

recommend to watch it on x0.75 ))

[@matteo7moh]

cs231n

[@lemyul]

thanks frid

[@jiansenxmu]

Pro-Tip: 0.75x speed. You're welcome :)

[@bayoudhkhaled1034]

Good explanation

[@ae_acar]

Cool, thanks

[@sournest]

Cell phone distributed training is a radical idea.🤔🤔🤔

[@pratik245]

Great

[@pkr619]

the fun starts @28:14

[@alexanderskusnov5119]

Low resolution.

[@aaaa8766]

slides

[@kodamusic3327]

даня кода