AI in 2020

Almost exactly 4 years ago I decided to dedicate my life to helping educate the world on Artificial Intelligence. There were hardly any resources designed for absolute beginners and the field was dominated by PhDs. In 2020, thanks to the extraordinary contributions of everyone in this community, all that has changed. It’s easier than ever before to enter into this field, even without an IT background. We’ve seen brave entrepreneurs figure out how to deploy this technology to save lives (medical imaging, automated diagnosis) and accelerate Science (AlphaFold). We’ve seen algorithmic advances (deepfakes) and ethical controversies (automated surveillance) that shocked the world. The AI field is now a global, cross-cultural movement that's not limited to academics alone. And that’s something all of us should be proud of, we’re all apart of this. I’ve packed a lot into this episode! I’ll give my annual lists of the best ML language and libraries to learn this year, how to learn ML in 2020, as well as 8 predictions about where this field is headed. I had a lot of fun making this, so I hope you enjoy it!

Are you a total beginner to machine learning? Watch this:

Learn Clojure:

Learn Rust:

Swift for Tensorflow:

Learn Python:

Learn PyTorch:

Papers with Code:

NeurIPS papers:

Neural Tangents:

XAI:

Quantization:

Waymo Open Dataset:

Graph Network example:

PyTorch Glow:

NeuroSymbolic AI:

MLFlow:

Graffitist:

Deoldify:

StyleGAN2:

Milasen:

Credits:
DeepMind
3Blue1Brown
J Cole

Honorable mentions:
My parents (for conceiving me)
Society (of which I am a product of)

Join us in the Wizards Slack channel:

Hit the Join button above to sign up to become a member of my channel for access to exclusive live streams!

Signup for my newsletter for exciting updates in the field of AI:

And please support me on Patreon:

[luischinchilla-garcia]

There’s a few topics here that can be seen as misleading. The biggest is PyTorch vs TF.

These two are very dependent on situation (model production, research, etc...). It also ignores that despite the increase in papers with PyTorch, the quantity comes from labs who use PyTorch publishing more, rather than PyTorch being widely implemented.

TF is heavily integrated with Google Cloud Storage, and leaving that out on a video that mentions the emphasis on the end to end lifecycle is very odd.

[impolitevegan]

I understand AI in medical applications will have a huge positive impact, but I see infinite positive impact when applied to education. Imagine everyone is educated in philosophy as Peter Singer and Derek Parfit, in physics as Feynman, in astronomy as Carl Sagan, in biology as Marie Curie and Jane Goodall, etc. We can prevent absurd amount of suffering and increase the well being of all sentience known to us.

[lanreuzamere]

Yeaahh! Now the Siraj I once knew is back with that "Energy" & "Motivating others"

[rasmustoivanen]

And once you share your project REMEMBER to give credits to the author of the original paper/project/code ;)

[zakuro]

I have no clue about how to code an ai, but I have thought: Maybe you could built a fully automatated laboratory, controlled by an AI. It could experiment with genome sequences to possibly create completely new materials or drugs. So, that's what I wanted to contribute to medicational reasearch.

[bsaiashish]

haha..that Siraj guy is never coming back..lol
nice to see you back...

[Sakura-zurz]

We have a moral imperative to keep advancing these technologies. But we do need to keep the dangers in mind

[makers_lab]

Learning by doing is excellent advice Siraj. At my company it counts significantly if an undergrad applying for an industrial placement position is working on projects in their spare time as it shows motivation beyond their course, suggests a passion for the industry and almost certainly puts them well ahead of their peers in terms of ability, almost all of whom will not have their own projects. Pursuing a degree can be beneficial for the broader base of background knowledge it can give and the chance to be exposed to avenues that someone might not have considered but that prove interesting (explore vs exploit in a sense), but it isn't essential and they rarely prepare someone well, if at all, for the real world; having actual experience of working one's own projects can easily trump merely having a degree and no experience.

[moka]

PyTorch is good especially to deal with RNNs, seq2seq and weights sharing.
But what are your thoughts about Tensorflow Eager?

[arijitsen]

Really I'm impressed after bearing all sorts of negative criticism, you are still continuing and trying to move forward. Really respect your unshakable attitude and best of luck with your future. Hope, you read this comment and do not repeat your mistakes again. :)

[threeMetreJim]

AI explainability: I think the problem is that current designs try to do everything in one go, giving an answer from an input with no way to reliably see what different stages do. Maybe breaking down a neural net into smaller units that each have a specific task, where it is known what that task is, is the way to go. The easiest thing I can think of to show this is how do people recognise a face? At a minimum I'd guess at 2 eyes, a nose and mouth. So the first stage has to recognise the individual elements. If you scattered those elements randomly around an image then it's likely that wouldn't be a face, so the next stage has to learn what relative positions of the elements are required for it to be an actual face. Each time you know what each piece of the network is doing, so you can understand how a decision was made. The example is a little simple, but that's how I can imagine it to work. I've always considered this to be how a person does the recognition thing, an 'association engine', patterns become associated with others, and when the associations are combined give you the answers you need. Lots of simpler operations, rather than one huge one that becomes a mystery. The only problem is that data needs to be broken down into it's component parts first, for the system to learn. It may be possible by making the data more detailed (multiple labels), and then monitoring what each part of the network does when presented with data that just has a single detail varied.

[kevinfortier]

The man did it. He is back from the dead. Good for you Siraj! People make mistakes, even giant ones like plagarizing the works of others, but life is long and while you are still breathing you deserve to get back up.

Your content is awesome for what it is. You are not an expert engineer, you are an expert communicator.

[slickliverpool]

I must admit the beginning of the video is hilarious!! lol

[hleet]

Your videos are getting better, sources are quoted/cited. And most of all, I can understand what you are saying because you master your subject. Keep going. I'm re-subscribing :)

[nonig]

It takes tonnes of courage to get back after admitting the mistakes and continuing with what you started. All the best! Keep doing what you are doing! :)

[rameshneupane]

I am not good at digesting critical comments! You did it and coming back stronger !! I love that !!

[Sakura-zurz]

Automated vehicles promise to increase driver safety??
Cybersecurity is a concern for everyone, from individuals to companies to national government.
Computer algorithms are helping the police predict where and when crimes will take place.
The spread of helpful robots in everyday life is accompanied by a fear that such technology may one day be adapted to build killer robots.

[paul_pierre]

Aaaand we're back. Disregard the haters Siraj. Right your wrongs and keep that energy up bc it will outlast and outshine your critics and make us all better than before we subscribed. ❤️

[rameshneupane]

Hay ! I love your videos as always !! And I see you coming stronger from your last public story!!! Keep doing great job !!

[forfreedomssake]

Nice to see you again Siraj :) I love your channel <3