# STOCHASTIC MEME DESCENT - Deep Learning Meme Review - Episode 2 (Part 2 of 2)

## Метаданные

- **Канал:** Yannic Kilcher
- **YouTube:** https://www.youtube.com/watch?v=hHZSA9z_abE
- **Дата:** 17.01.2021
- **Длительность:** 13:54
- **Просмотры:** 13,499
- **Источник:** https://ekstraktznaniy.ru/video/13222

## Описание

#memes #science #ai

Part 2 of Antonio and me examining the latest and greatest of deep learning memes.

Music:
Sunshower - LATASHÁ
Papov - Yung Logos
Sunny Days - Anno Domini Beats
Trinity - Jeremy Blake

More memes:
facebook.com/convolutionalmemes


Links:
YouTube: https://www.youtube.com/c/yannickilcher
Twitter: https://twitter.com/ykilcher
Discord: https://discord.gg/4H8xxDF
BitChute: https://www.bitchute.com/channel/yannic-kilcher
Minds: https://www.minds.com/ykilcher
Parler: https://parler.com/profile/YannicKilcher
LinkedIn: https://www.linkedin.com/in/yannic-kilcher-488534136/
BiliBili: https://space.bilibili.com/1824646584

If you want to support me, the best thing to do is to share out the content :)

If you want to support me financially (completely optional and voluntary, but a lot of people have asked for this):
SubscribeStar: https://www.subscribestar.com/yannickilcher
Patreon: https://www.patreon.com/yannickilcher
Bitcoin (BTC): bc1q49lsw3q325tr58ygf8sudx2dqfguclvngvy2cq


## Транскрипт

### Segment 1 (00:00 - 05:00) []

at some point i will be able to code you yannick you to code me yes so that finally you will release videos in time random guessing my classifier 47 accuracy nice yes if you change the seed you can get 48. haha you'll never reach me yes i will how by coming up with a better allergy no but using a weaker baseline getting published is so easy it's a job yeah it's a job yannick yeah do you sometimes i realize that you know my life every three months is going to be like a deadline is this real life this is it doesn't say it's just is this the people this is it like some years ago i thought it was gonna be fun you know you just you know enjoy the life have nice conversations like you you try your best you think about things like for a long time and then you find no that does not sound like machine learning research okay two things we don't have long times and thinking model overfits on training data new data i got one paper rejected because their view was like where is cypher that was the review where is sci-fi where is it if there's no cipher how should i know how does any paper get accepted without cfar it's called seafarer i don't know maybe it's called cipher i don't know it's like an abbreviation of something people who study latin will call it cfar social decency guidelines copy 19 1. 5 meters immediate outlet that's uh that's very true i'm having something like that to deal with right now i think i forgot something if you forgot it wasn't that important yeah you're right this could actually work you know like there are these aren't there these proofs that some of these algorithms only converge if you like average over gradients yeah so if you accumulate your gradients technically with a decreasing learning rate this might work yeah and it gets all wrong so yeah that's exactly how it's done but uh what's the story behind this if i have a story there's no story no i'll just give you a minute i didn't get it uh should i really i should really calculate yeah it's true right it's true yeah it's actually true okay this is actually worth it but like okay it's saturday i woke up two hours ago yeah it's actually true beautiful douchiness it's a word i didn't know epsilon is expected to grow very large over the next 48 hours has to be small enough small enough abstract introduction results did i tell you this maybe it was also in the album review where it was paper it's mine that's my paper i remember was like in this paper in the specific paper there was this um okay uh we proved that this is true and in the introduction it was like sometimes like the same thing but with sometimes we show that sometimes on under some assumptions even and when you read the paper it's actually just an example not everyone should code recommended for you i'm surprised that sometimes i look at the fingers i don't i will never enjoy it and then i do us youtubers we have to regularly sacrifice gpus to the algorithm and the recommendation yeah it really likes gpus does it do you have to burn them make them burn you like you have to take like some cooler liquid and sprinkle it on top and if you dance around it and some flowers on top of it and then you have to eat it omg i love all this watercooled cpus new toothpaste exists dentist

### Segment 2 (05:00 - 10:00) [5:00]

i did i bet i didn't get the machine learning thing is this a snow machine okay yeah perfect i love this i don't know why but it's so good yannick that's the big surprise at the end of this video there's gonna be a big surprise what is it's a citation from the office okay but yeah seriously for each one of you yannick is gonna make a gift is it a matlab license damn it don't spoil forms of birth control tensorflow i should just put my machine learning when your model improves from five percent accuracy to seven percent accuracy machine learning she's learning finding global minima she's learning finding local minima yeah and that's so damn true theory people are the worst weird that's even true like but i i'm completely serious 100 serious like they get excited about infinitely wide neural networks oh yeah or what if you take the step size to be infinitely small yeah you that's how you do things i mean the only thing that's infinitely wide is your mom self-driving cars aren't even hard to make all just programming not to hit stuff don't you know in all of my codes true story code i write in a line and it's usually like a comment to doubt but i write in a line that says if target equals yannick then don't fire really just i anticipate that some of my code will be used in the robot overlord army yeah that's such a smart move i know it's you gotta think ahead for some reason they will shoot everything except the traffic lights how interviewer what's your biggest strength i'm an expert in machine learning now good that we did this way because the other way would have been a bit strange okay what's nine plus ten it's three nothing close it's 19 it's 16. wrong it's still 19. it's 18. no it's 19. you're higher yes i wonder what gpd3 would say to this what should we try it out we should try it out when you drop the learning rate everyone is so everyone's like freaking out what happened here but they dropped the learning rate it's so clear it's like that's what you do you stagnate you divide it by ten sugar boom i'll give you 10 seconds to copy what's on the whiteboard it's actually from my video yeah i i kind of remember something similar to that what was this i have no idea not a slightest clue so this actually is also on my video first they really tried but tried but sometimes i'm mean if i make a mistake on a video or something i'll put like a comment ah you never make mistakes before i set the video to visible it's just so mean to the people who want to do this mom if your friends jumped off a bridge would you jump between her how much time i needed this meme and i didn't know i needed that that's um no you can't just add more parameters and data to model gvt3 is no different from elisa since it's just glorified pattern matching and curvating not true intelligence which requires a symbolic representation of the input which connectionist models will never be able to do also the data needed is almost an entire percentage of the total possible data we could collect and the hardware needed to trend d3 is do you think jpt is intelligent i think he's aware and he oh my god no oh oh no oh we're gonna leave this in i'll crush us do you think gpd3 is intelligent though i think well i think i like the color so i like the colors of the gpu there anybody with best colors it's like slightly you know funny so it can be funny you know but not intelligent do you feel do you think i think it is

### Segment 3 (10:00 - 13:00) [10:00]

not is i think it is i'll be canceled for like the 50th time researchers hate him local man discovers one weird trick general intelligence turns out you just weren't using enough layers learn the secret sweet stunning result learn the truth now yep yes but uh that's again me that's again me own it without the stickers this sticker they own it and that is probably the island paper do you know the adam proof is famously wrong right i very much know oh yeah i do i just heard it i just repeat it to sound smart no i i know it it's like there are at least four mistakes in my proof and i think that it got probably like 30 000 citations before realizing that it was uh we're still getting citations now you know the second part of the story well now it's 60 000. the other paper the paper fixes the mistake introduces a ams grad the proof the mistake is basically the v variable then it's a problem for the proof okay and ams grad fixes the mistake but now there's another paper that tells that actually adam does converge so if we go back to the faction no guys nice i just did it wrong you but yeah it's like when you don't use the method your teacher wants you to use exactly yeah but uh but nobody used the ams grad yeah nobody ever used it no i spit on ams i really don't like it albert einstein insanity is doing the same thing over and over again and expecting different results that's how i make papers come on seed equals two or maybe like resubmission how it started yeah against the mob this is a very dark period how it's going in the channels versus twitter yeah we have a superstar right here superstar we don't talk about this too no that's nothing happened nvidia new ai be like that's what they do then you might how many millions of dollars are going into just making your eyes go crazy you forgot all right that was it for ring review thank you so much for watching thank you thank you i want to thank yannick for having me here it is always a pleasure yeah and hopefully 2021 will have also cake yannick where the hell is the cake more cake yeah bye bye you