AI That Creates AI

Подписаться 769 тыс.

Просмотров 107 тыс.

50% 1

How can we build AI that creates AI? This concept actually already exists! Google uses self-creating AI as part of its AutoML service that finds the best model for customers. Genetic algorithms and neuroevolutionary strategies offer us a way to replicate the process of natural selection en silico. I'll talk about ways of using machine learning algorithms to create better machine learning algorithms in this video.
Code for this video (with coding challenge):
github.com/llSourcell/neuroev...
Der Waldi's Winning Code:
github.com/DerWaldi/youtube-v...
Paramdeep's Runner up Code:
github.com/ParmuSingh/face_swap
Please Subscribe! And like. And comment. That's what keeps me going.
Want more education? Connect with me here:
Twitter: / sirajraval
Facebook: / sirajology
instagram: / sirajraval
More learning resources:
• AI Codes its Own ‘AI C...
• AI evolves to compose ...
• Art By Artificial Inte...
• Amazing! Conversation ...
• Google's DeepMind AI J...
gekkoquant.com/2016/03/13/evol...
Join us in the Wizards Slack channel:
wizards.herokuapp.com/
And please support me on Patreon:
www.patreon.com/user?u=3191693
Signup for my newsletter for exciting updates in the field of AI:
goo.gl/FZzJ5w
Hit the Join button above to sign up to become a member of my channel for access to exclusive content! Join my AI community: chatgptschool.io/ Sign up for my AI Sports betting Bot, WagerGPT! (500 spots available):
www.wagergpt.co

Опубликовано:

9 фев 2018

Ссылка:

Скачать:

Готовим ссылку...

Добавить в:

Мой плейлист

Посмотреть позже

Комментарии : 263

@SebiMu 6 лет назад

Yeah coding challenge winner :) Thanks Siraj!

@SirajRaval 6 лет назад

thanks for the submission! Looking forward to more

@markonfilms 6 лет назад

Cloned your code and followed your instructions and it is absolutely awesome! I'm super impressed with the quality of your work. Also thank you for the comments in the code. It really helps when I try tweaks. I am looking into making this code properly scale on multiple GPUs. I see that Keras has a fairly simple way to implement this when using TensorFlow as the backend. I am having a tough time trying to figure out how to adapt this particular project to run on multiple GPUs however.

@SebiMu 6 лет назад

Nice :) I am really excited what results you can get with that computing power. My first thought would be to simply split the two decoders on a gpu each, but I have absolutely no idea of multi GPUs.

@markonfilms 6 лет назад

Der Waldi I've been playing with the code a lot and taking a crash course into Keras. I was able to get the model to technically run on all the GPUs but not with any real speed benefit. I also tried a script a found to split the batches up but it caused the tensor dementionality to go wrong in some way that wasn't apparently easy to fix. Another issue is that the training data for the faces I'm testing with right now aren't very large so there's less potential to optimize for multiple GPUs from what I've read. A single 1080 is actually fairly impressive to me in the results. So far I haven't actually finished training because I've been messing with parameters etc too much.

@JoswinConcessao 6 лет назад

Next Video: AI that creates an AI which creates an AI

@tarunkashyap8515 6 лет назад

inception AI

@JoswinConcessao 6 лет назад

tarun kashyap haha

@SideControl 6 лет назад

ACeption

@Bestname_Ever 6 лет назад

Pure Motivation ERROR: AN ACEPTIONS HAS OCCURRED, RESTARTING WINDOWS

@2LegHumanist 6 лет назад

It's not an AI that creates an AI.

@fabdenur 6 лет назад

Hi Siraj, I've complained a couple of times here in the RU-vid comments about you only including brief flashes of the results of the codes. Now it's my time to praise: in this video you included a full 20 seconds (from 8:23 to 8:43) of the neural net playing the game. It was very satisfying to get to appreciate the model in action. Congratulations and keep it up!

@psinha6502 5 лет назад

Why I watch these videos even though I understand a little bit is because I know someday these things will start making sense to me once I am done with what I am learning nowadays....And that drives me to learn things

@inquiremore1178 6 лет назад

Hey there Siraj, I love your videos and they serve as a great inspiration for me to pursue an education in AI. One thing I'd like to mention though is I notice that your video has a fair bit of background noise. To avoid this, you can download Audacity (free software), which I use for the videos I've made, and use the noise reduction feature improve your sound quality even if your recording setup isn't perfect. Best of luck!

@moulikchaturvedi5407 6 лет назад

Hey Siraj ! First off, I'm a big fan and love your work ! And what if we create a model which stores it's data collected on a graph and use random mutations to create a few offsprings of the model and choose the best model of them and repeat the process. That will be a genetic algorithm, right ? Will it be the right approach ?

@arthurvanderwal 5 лет назад

Was expecting you'd sign off with 'but now I got to reproduce...' - awesome video as usual!

@cientificamente3628 6 лет назад

This comment is just to thank you for the great content you produced. I think, in a way, you're doing for the current generation of programmers what the creators of "The World of Beakman" did to my generation: they inspired kids and teens, showing that science can be diverted. Once again, congratulations on your great work!

@tristant9686 6 лет назад

I like the small summary at the end of your videos.

@cod-newbie9166 5 лет назад

I love your content, I just want to know if it would be ideal to use tensorflow to start ML without knowing how to code the neural network from Scratch?

@dhrumilbarot1431 6 лет назад

I just saw Josh tanenbaum's agi mit lecture and was thinking about it.....there you go lucky me!!!

@saveyourlifes7867 6 лет назад

So, if I got it right, autoML is composed of two AI, the parent, and the child. The goal of the child is to have the lowest error rate on a given dataset. The role of the parent is to find the best architecture for the child so that its error rate is the lowest possible. But what do you exactly mean by "architecture"? Does it have the capability to add convolutions, LSTM... or does it only modify hyperparameters? How is it better than grid search?

@KennethSaul 6 лет назад

Every once in a while I run across a video that reminds me how uneducated I am. Thanks Siraj. Not to say that the presentation was bad or anything.

@_taylor_v 6 лет назад

Is there a benefit to not using crossover of multiple "fit" models? I guess there are many different ways to implement genetic algorithms, but I've never seen one rely so heavily on only one fit member of the population and not at least a few for crossover. Interesting stuff! Nice video!

@RadhikaMeera 6 лет назад

Hi Siraj, what I should the difference between AI, machine learning , deep learning . Do you have a video just like the one you had for machine learning in 3 months for AI

@BlackXxScopez 6 лет назад

Would it be beneficial to keep some momentum when performing gradient descent to avoid local minimums?

@greenserpent7712 6 лет назад

I was waiting for you to cover this! Nice!

@TheZappingChannel 3 года назад

Wait, so the thing we optimize with GA are the hyperparameters or the trainable weights themselves?

@empretecasociados8090 6 лет назад

y creeme que aprecio tus aporte y tus contenidos, pero también algunas veces también entiendo a tus detractores. Ellos son mucho más conservadores, creo que eso es porque están mas focalizados en el hacer hoy. contrario a ti que estas mas concentrado en comunicar en una visión "Janusvian" del problema (ver el pasado, presente y el futuro de la AI). Disfruto mucho tu forma de comunicar, es muy agradable e intrépida. Espero mas videos Gracias Siraj

@iprototypei6416 6 лет назад

Hi Siraj, which machine learning framework can you consider ?

@roguedrones 6 лет назад

Siraj Raval, hardcore Ai genius.

@Grepoan 6 лет назад

I'm trying to find the 'winners' for the code competition in this video, so that I can learn some alternative techniques. But I can't find the 'next' video, because I did not load this from a playlist. Could you link to the 'next' video where you have the top two codes for this functionality? Thank you

@primalcam812 6 лет назад

Yo, Siraj hittin the gym! LOL Nice man

@PrafullaPallal 6 лет назад

Good work. Do keep posting.

@6lack5ushi 6 лет назад

Great Video!!! Where do you see a mechanical engineers role in Machine Learning. Should we stick to what we know or try to branch into machine learning as well? A bit of it is taught to us. All the math you go through has been taught to us, if anything its easier lol. But should we invest time into learning how to implement models. I do because i enjoy it but should all mechs put some work into it? KR Ben

@Dude0413 6 лет назад

I love to hear all these “THE END IS NEAR” people when we aren’t too close to AGI

@luck3949 6 лет назад

Dude0413 My estimate to an AGI is 5-20 years, which is quite near.

@SirajRaval 6 лет назад

Exactly, the comments are hilarious though

@AkashMishra23 6 лет назад

Siraj Raval I, For One, Welcome our new AGI Overlords

@orion411 6 лет назад

Never heard of Sirisys?

@RyanScottForReal 6 лет назад

We don't need AGI to be threatened by AI. The AI just needs to learn that we re often the limiting factor in many things we want to improve, and it decides we need to be removed. Furthermore an AI doesn't need to be as smart as us to trick us. That is super easy to do.

@petrnejedly6592 6 лет назад

Which genetic algorithm is the best for evolving NN?

@tsgbalakarthik Год назад

I love this video and you and the concept neuro evolution

@redsusoverparadise2700 5 лет назад

Looked for this video because Justin Y. commented on "AI learns how to play Flappy Bird"

@rabiielbeji8541 6 лет назад

AI will create the best artificial doctors, developers, designers, artists, ... AI will create the new history of all logical sciences. Maybe, I wouldn't be present at this period of progress but I'm sure it'll do it!

@rabiielbeji8541 6 лет назад

I know that current machines learning techniques and deep learning models still far to simulate all human activities. But we can improve the AI fields by inventing new technics inspired from neuroscience that consume fewer data and power and provide better accuracy

@rabiielbeji8541 6 лет назад

of course *Simon*, you are right about the bad uses of the AI technologies like deepfake and it's a common problem for all technologies. I don't say it's easy, but just I have this dream how these technologies will facilitate our life in the future and current AI technics was just the first attempt. I respect your point of view my friend :-) ;-)

@Belofsky1 6 лет назад

No it won’t.

@Belofsky1 6 лет назад

Just splash a cup of water on it, problem solved. LOL

@0EEVV0 6 лет назад

Simon WoodburyForget took 20 years to get decent NN, took us a lot of years of evolution.

@Mukund568 6 лет назад

Good stuff!

@forthtemple 6 лет назад

If you were to use AutoML for image classification would it be using purely genetic algorithms or is it a combination of genetic algorithms and classic CNN gradient descent?

@ozuwara1861 6 лет назад

Siraj, if you touched neuroevolution please make a video on these "safe mutations" presented in Uber's paper

@reyariass 6 лет назад

How about an ai that builds the best looking and most fuel efficient rocket that can get to the mun in Kerbal Space Program?

@DiamondTurtleGamer 6 лет назад

I was waiting till this was going to happen

@JazevoAudiosurf 6 лет назад

good vid. i think this is one of the more interesting future topics if we want to build a human like brain one day

@ayushpatel5033 6 лет назад

Is this method more computation expensive than gradient descent?

@zanderceo91 6 лет назад

Did that music snippet say- drop it like a bot lol? Glad to see you appreciate rap culture. I dig it. Good content as usual.

@btouw8558 6 лет назад

i am new to phyton, but you inspire me start relearning programming, what ide is free and most practical, icould not find in the playlists

@cashewfilms5372 4 года назад

"hello world! it's sirage!" me who had hell figuring out the basic console commands: aight ima head out

@sumitaggarwal9477 6 лет назад

I loved this video..........but it would be really awesome if you could add a requirement.txt file to your repo.....will make it easier to run the code

@SirajRaval 6 лет назад

Yo yo I will get on this good to see u

@subramanyam2699 5 лет назад

This is insanely awesome..

@hitman779 6 лет назад

Ok I still don't get it. Why use neuroevolutionary algorithms where you can use hyper-parameter tuning? CAn someone explain?

@AnonymousAccount514 6 лет назад

This channel is so awesome

@larryteslaspacexboringlawr739 6 лет назад

thank you for ai recursion video

@jtekmmx 6 лет назад

Check out this example ! ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-qv6UVOQ0F44.html Neuroevolution learns to play Super Mario World

@Kaixo 6 лет назад

I used Evolution to trail a neural network to play Snake, it's way better than when I did it using gradient descent... And It's easier to code.

@Euquila 6 лет назад

Man, you are one smart dude!

@stavrospapadopoulos115 6 лет назад

Why dont you give reference to the papers mentioned in the video? (Write them in the description)

@oscarmulin114 6 лет назад

Gotta note that, genetic algorithms are a lot slower than backprop, especially when you use any significant number of neurons (around >100). As an another note, trying to reproduce neural nets with varying topologies usually does not give an offspring that's better than its parents, so topology numbers have to be fixed or else optimization space is too big. (But, maybe with modern hardware it will be fast enough, who knows?)

@w.maximilliandejohnsonbour725 6 лет назад

Interesting...!!!!!.

@ShaikAsad 6 лет назад

I was waiting for this video! Thanks! Why don't you rap anymore.

@SirajRaval 6 лет назад

rap is coming more aggressive than ever

@ShaikAsad 6 лет назад

Can you make a video on reading source code from github and completely solving a machine learning problem from start to finish?

@salmanfaroz2759 6 лет назад

@Maisonier 5 лет назад

It's possible to create a kernel with AI?

@lauriejmusic 6 лет назад

9:02 - so gradient descent isn't computed for any of the population ? same for backprop ?

@vickyredwon 6 лет назад

Siraj that creates siraj

@eav300M 6 лет назад

We love Siraj ❤👏👏💪💪✌

@SirajRaval 6 лет назад

Love u Edgar

@igorboldysh2749 6 лет назад

why do you have so many problems(green areas at the bottom of the screen) with picture in this video? Thnx 4 videos, btw

@artemkovera5500 6 лет назад

I just was thinking about GAs for building AGI.

@DariusAlexanderMusic 6 лет назад

Paramdeep Singh: Always the bridesmaid, never the bride.

@TheStarCycle 6 лет назад

Great video, but I believe NEAT might actually take 2 parents and generate offspring from them, as opposed to selecting just one parent, but I could be wrong.

@Abdega 6 лет назад

9:30 He has a perfect opportunity to have his username be “ParamDeep Learning”

@poxyu_was_here 6 лет назад

very nice, very nice indeed

@camdenparsons5114 6 лет назад

hey siraj, lets see some more on techniques to learn more with less data. also if you haven't seen 'MIT 6.S099: Artificial General Intelligence' (im sure you have) you need to check that shit out.

@Aer0xander 6 лет назад

When singularity?

@thunderzone1152 6 лет назад

you should make an AI that can make a video about Deep learning and submit it

@simpleman5098 6 лет назад

Can we build a shitcoin that creates a shitcoin?

@markonfilms 6 лет назад

Simple Man that generates a supershitecoin

@Belofsky1 6 лет назад

Yeah it’s called “SirajCoin”

@bauwndule 6 лет назад

Siraj Raval is God!

@samzhao638 3 года назад

A.I. gives us an illusion of creativity. It have no idea when given a blank paper to come up with something. Keep in mind it is still a tool.

@tylerknight99 6 лет назад

"The core problem is finding the right hyperparameters settings that result in the lowest loss or the highest reward. The traditional way of doing this is with gradient descent which is called back propagation."

@SirajRaval 6 лет назад

Meant to say parameters

@watcher4075 6 лет назад

I remember when Stewie from family Guy created an AI and that AI created another one.

@amitkhopade9372 5 лет назад

Create video on Facebook Stopped AI project, include why they stopped ,what was that ..?

@kevinowenburress2435 6 лет назад

So I need to make neural networks that make neural networks that make interacting neural networks. What's new?

@star277hk 6 лет назад

your arm muscle looks bigger than before

@buenaventuralosgrandes9266 6 лет назад

he probably doing weightlifting

@Mzulfreaky 6 лет назад

He made an algorithm to properly get them gains

@huseyngadirov7658 6 лет назад

He def uses a GAN

@OktoPutsch 6 лет назад

It's a Washing Machine Learning problem. Shrinking in every way.

@herbivorouscyborg2398 4 года назад

I'm struggling to understand how the "reproduction" step is supposed to work. How do you produce offspring which benefit from the parent's good performance while also mutating it to perform differently?

@Y2Kvids 6 лет назад

So AI tests AI algorithsms ?

@zakariyamohamed6388 3 года назад

Is there any possible way that we can get A.i to teach human? I think it is possible 🤔

@CodeDham 4 года назад

Nice video #codedham

@andrewkiminhwan 6 лет назад

Never Stop Learning! I'm talking to you Ais

@sunegocioexitoso 6 лет назад

Plot twist: Siraj it’s an AI

@Luke7389 6 лет назад

So if its like a ball rolling down it means there is gravity in an algorithm and AI finds the best way through it ?!?!

@diegoantoniorosariopalomin4977 6 лет назад

I'm going to say it again , strong ai will be achieved by distributed neuroevolution

@artemkovera5500 6 лет назад

Maybe. In fact, the biological brain is a product of evolution.

@luck3949 6 лет назад

Diego Antonio Rosario Palomino I think program synthesis techniques will achieve AGI sooner then any variations of deep learning

@descai10 6 лет назад

Now make an AI that recreates itself but more efficient. Recursively.

@PhilippGrafendorfer 6 лет назад

bizeps is coming!!!

@Partytron 6 лет назад

neat

@cryptotrading2420 6 лет назад

why is your front part of hair partly colored?

@xrayer4412 6 лет назад

My mind is blown

@madjoemak 6 лет назад

I wish I was the one who designed neural networks... Sadly I'm not 😣

@dzhiurgis 6 лет назад

Salesforce had this for like 2 years or something

@planktonfun1 6 лет назад

finally neat

@octaviusdelmonte9019 4 года назад

Also known as: once thought "invaluable" programmers now made obsolete.

@2LegHumanist 6 лет назад

This is just parameter tuning. Why are you hyping it?

@kalebrand 6 лет назад

Because it's not just tuning parameters of a fixed architecture, it's also trying different network architectures and choosing the ones that perform best

@2LegHumanist 6 лет назад

kalebrand Where did I use the term "fixed architecture"? The architecture itself is made up of a set of hyperparameters. It is not changing the underlying code or algorithm. It is not something that can lead to the next backpropagation or a new approach to neural network based predictive analytics. What it does is save a data scientist a bit of time and produces models with a couple of percent better accuracy than what humans have been able to do. It is a great achievement, but to call it AI creating AI is pure hype and clickbait.

@2LegHumanist 6 лет назад

...and the fact that people commenting here equate this to kurzweilian self-improving AI just proves how irresponsibly misleading this turn of phrase is.

@kalebrand 6 лет назад

That's a fair assessment. News and media surrounding AI/ML is almost always really misleading. All we have is algorithms that can minimize loss on a single objective, and that's all we're gonna have for quite some time. AI is just glorified linear algebra, and people are freaking out doomsday scenarios because they're being so misled.

@NoahHornberger 6 лет назад

I think its because A is the first letter of the alphabet and I is the word we use to talk about ourselves. So it seems like something extra special just because of the letters and their over-interpreted meaning. If it was called SK or GX no one would be afraid. AI sounds like a creature with more sentience, but in truth there is no brain, nerves or even a center of awareness.