Unsupervised Cross-lingual Representation Learning at Scale https://twitter.com/alex_conneau/status/1192490719031656448?s=19

Hello, friends 🎉

After 2 months of sleepless nights and battles with gradients, we're happy to share with you our new online course about Machine Learning - “Object Detection with PyTorch”
https://learnml.today

You may ask why we did it if there are so many similar courses? - And that’s a good question 💛

💎 The reason zero - we like ML and we like to teach people, cause that's one of the greatest things that u can give back to the community - knowledge that was gained by hard work and a huge amount of time spent. So the next generation can go farther from where we stopped.

💎 The first reason - creating courses on some small area/problem/task to transfer knowledge to other specialists as fast as possible and without bullshit.
We are not professors from the university, we are practical guys who are paying their bills by doing researches and solutions in Machine Learning and AI areas. And as professionals, we often have a need to get new knowledge with deep understanding of problematic of tasks asap. But all we can find on the market is online courses about simple stuff that good for newcomers but not for us. And that’s a big problem because getting knowledge from papers is time-consuming.

💎 The second reason - a huge gap between studying materials and real tasks. I think many of you had this feeling in the past when the teacher describe you something like 2+2 = 4, and then in real life, you get task to calculate the trajectory of a spaceship to Mars.
So we want to build courses from developers for developers, with giving real practical knowledge without gaps, so students can be ready for real-life after the course ends.

💎 The course will lead you from the basics to the latest state-of-the-art solution and will consist of 7 lessons:
⭐ PyTorch Basics
⭐ Single Object Localisation
⭐ Single Shot Networks / Yolo
⭐ Single Shot Networks / SSD
⭐ Region Proposal Networks / Fast R-CNN
⭐ Region Proposal Networks / Mask R-CNN
⭐ Bonus Material

💎 Requirements: Python, Base Math, ML Basics (CNN networks, Dense Networks)

💎 Time: 4+ weeks (2 lessons/week but we will look by student progress)

💎 Interaction with students: GitHub Issues

💎 Environment(free): Google Colab, Ram: 12GB, Disk: 350GB, GPU: Nvidia T4 16GB
💎 Start date: 18 November 2019

Also we have small demo tutorial for you: https://bit.ly/traffic-counting-with-opencv

You can subscribe on course on our page https://learnml.today

Interesting hardware news from Intel https://www.engadget.com/2019/11/12/intel-nervana-chips-for-ai-in-cloud/

Create 3D scene from 1-2 images
https://arxiv.org/abs/1911.04554

How make smaller and faster network by knowledge got from big one?
https://blog.floydhub.com/knowledge-distillation/

DeepFake that we deserved

South Korea has created an entire city for testing self-driving cars with 35 kinds of road test facilities. Here is the video from BuzzFeed news

Here is new paper https://arxiv.org/abs/1911.08541 on debluring images. if you look closer you will notice that all blur was done almost with linear motion. Such blur can be removed without NN at all.
Canon already use this method in their cameras for software stabilization.
For example here paper on restoring licence plate after motion blur without NN https://iopscience.iop.org/article/10.1088/1757-899X/263/4/042024/pdf
Or here with deblur kernel estimation https://www.cse.cuhk.edu.hk/~leojia/projects/robust_deblur/

​​Hey friends, I want to share a great job opportunity with you.
Grammarly Kyiv office looking for ML/NLP engineers who are willing to build an awesome product by implementing their innermost fantasies.
https://bit.ly/grammarly-job

We just created an additional channel for our community on Discord. Please join if you are using it.
https://discord.gg/ZdxKbf

If you were searching the way how you can support us to make our community stronger and build more educational content, we have the answer now!

We just launched our page on Patreon so everyone now can become a supporter and help us to educate people :)
https://www.patreon.com/mlworld

Creating some interesting effect from one photo
https://github.com/sniklaus/3d-ken-burns

Nvidia outsourced cool library Kaolin for 3d research https://github.com/NVIDIAGameWorks/kaolin

Just found interesting thing torch.argmax(0,0,0) will return 2 instead logically 0 (torch.max also).
So be aware of this when you will be using it.