Hopefully Not Boring Weekly AI Stuff #2
Everything interesting about AI I've found on the internet this week. I know I am late and it is nearly Tuesday, sorry. But my blog has "irregular" in the name. Anyways, hope you will enjoy it!
This week:
Nice ML YouTube channel
Extremely giant model from The Beijing Academy of Artificial Intelligence
Google Applied ML summit
Kevin Spacey as Barack Obama
Unreal engine trick
"Institut Lumière" vs AI video enhancement
Albumentations 1.0 is released (best augmentations in python)
DVC studio release
Golden words from Yann LeCun
P.S. May be interesting to watch this week:
PyTorch Community Voices | TorchIO | Fernando Pérez-García
Join us for an interview with star PyTorch community member Fernando Pérez-García as we learn about and discuss TorchIO, a medical image preprocessing and augmentation toolkit for deep learnings written in PyTorch.
To the details.
ML channel for those who want good quality content
Recently, someone, someone on Reddit pointed at a nice 3Blue1Brown-styled ML YouTube channel. I had a look and liked it as well. Have look here and also watch this video about transformers.
1.75 trillion(!) parameters model from The Beijing Academy of Artificial Intelligence
The name of the model is Wudao. So, according to this article, Wudao can do, hmmm, anything you like.
the institution demonstrated Wudao performing tasks such as natural language processing, text generation, image recognition, image generation, etc.
What is also interesting, is that Wudao is funded by China’s Ministry of Science and Technology + Beijing’s municipal government.
And, worth mentioning, no real research article published (yet) and no code available anywhere. Only this FastMoE package Github repo that BAAI researchers developed to train Wudao.
Google Applied ML summit
On the 10th of June Google will hold this summit online. Registration is now open.
Some presentations names:
The Next Five Years in ML
Live Q&A: The Next 5 Years of Reinforcement Learning at Spotify
How To Build an AutoML Model with UI and the SDK
Deploying and Explaining Your Model
Scalable ML Deployment Using PyTorch and Kubeflow Pipelines
How Google Does ML
and more
I am already registered. And you?
Audio-driven Neural Rendering of Portrait Videos
Recently this thread was upvoted on Reddit.
It has a video in it, where Barack Obama on the left side is, sort of, saying what Kevin Spacey says on the right side. Well, as for me, this is not even close. Sorry :(
Below is not a video, but screeshot. Watch video on reddit thread.
Unreal engine trick
Just cool and funny images from a combination of VQGAN, CLIP, and unreal engine.
Find more images in this Twitter thread.
The "Institut Lumière" tries to delete from YouTube the AI enhanced video "Arrival of a Train at La Ciotat" made by AI enthusiast Denis Shiryaev
I recently made a long read about it here. Additionally, Denis shared his post on YouTube. Read at least of them to be aware of what could happen to AI enthusiasts all over the world.
I hope one day the video will come back, and won’t suffer from corporate madness.
Now a bit of technicall stuff.
Albumentations 1.0 is released
I use this package for a long time. Imho, it is best augmentations in python.
Have a look at their overview and docs. Full release notes are here.
Oh, they recently developed online demo. (I think it is build on streamlit)
DVC studio release
DVC stands for Data Version Control - good tool for MLops. You can not only track you experiments and models, but also version control everything, including - surprise - datasets. Their disatvantage up untill now was lack of UI. And now they cover this area as well.
Have a look and their blog-post: Introducing DVC Studio
Finall one
Description of ML specialists from Yann LeCun
I will just leave it here