Jason's Machine Learning 101 | Hacker News Reader

Jason's Machine Learning 101

docs.google.com

332 pointsantgoldbloom24 comments

This seems a bit 100 than 101 but if anyone's looking for something more next-level, check out this ML Crash Course by Google (featuring Peter Norvig and many other Google engineers) https://developers.google.com/machine-learning/crash-course/...

Indeed, I highly recommend the MLCC by Google if you want to go deeper technically speaking and more hands on. My deck was aimed at complete beginners (even those with no / limited math background). Having taken courses like the MLCC myself way back when I found that there were some gaps in my own knowledge before I could appreciate that in its entirety which is in part how my deck came about to being.

Nice job, man. By the way, how do you create those animated gifs for explaining?

The GIFs for the product demos such as the ones I have for Soli etc can be made by dragging an MP4 into Photoshop and then exporting for web as GIF after resizing. You can then drag those into Google slides and resize as needed.

Warning, the course uses Python 2.

Here's the discussion on HN when it was posted: https://news.ycombinator.com/item?id=16493489

I still have to agree with minimaxir's comment, but I'm nowhere near as pivotal/experienced/well-versed in ML as the parent comment.

Minor correction: the Boston Dynamics robots don't use machine learning. This is a common misconception, but almost all robots today still use hardwired control laws. We don't really know how to make robots that can teach themselves how to do things yet.

Sure- Boston dynamics has bleeding edge controls.

However, I would be really surprised if they didn’t use machine learning! Their robots have strong perception systems. How do they accomplish that without machine learning?

See this Quora question by a research engineer at Google Brain:

https://www.quora.com/What-kind-of-learning-algorithms-are-u...

It doesn't look like they are using a classifier to do object recognition, although I confess I've never heard of the "sequential composition of cost funnels" the post is describing, neither do I claim to have understood it at any depth just from that post.

In any case, it does look like most of their AI (if you would even call it that) is hand-crafted. I understand that this is the done thing in robotics, in general.

Note also the various announcements by prominent deep learning groups, like DeepMind and OpenAI, about teaching robots, or robot hands, to manipulate various objects of limited shapes and forms. If deep learning and deep reinforcement learning was particularly successful in training robots to interact with real-world environments, you can bet you'd see a lot more announcements advertising this, with titles like "We taught a robot to peel potatoes using deep learning" etc.

It would be interesting to see if other machine learning techniques are often used with robotics. I am aware of one paper [1] that uses Inductive Logic Programming for robot vision, but robotics is really not my field so I'm probably missing lots of other work.

__________________

[1] Meta-Interpretive Learning from noisy images

https://www.doc.ic.ac.uk/~shm/Papers/logvismlj.pdf

Full disclosure: one of the authors is my PhD advisor

Control algos don't use machine learning yet for the most part, but base perception most certainly involves convnets.

Reinforcement Learning is Machine Learning.

Sharing because this is a nice intro deck that new people on our team (Kaggle) have found useful for coming up to speed on ML

Thanks for sharing! Kaggle is a treasure!

I see this is referred to as a deck, but I don't know what that means. What is the specific meaning of deck in this context? Where do I see more decks like this?

Reference is to slide deck. A deck of slides. This presentation was created using Google Slides (slides.google.com). I am sure others exist but its up to the creator to make them public / shared for others to see.

This is a god damn treasure trove of info for people new to ML. Thank you!

isn't everyone tired of these beginner crash courses in ML?

Clearly there’s still demand. Maybe there’s a need for something better.

Can't download or print (chrome/safari). Working for anyone else?

This was on purpose. This is a "living deck" so will be updated from time to time and didn't want to maintain several copies. If you have the Google Drive app I believe you can download for offline viewing and then it will sync automatically when connection.

astonishingly devoid of references, and it ignores multiple decades of computer science

That seems fine for a 101 course. If students are compelled to continue, they will find all sorts of breadcrumbs from 102 onwards. Or do I just have low standards?

For a 101 course I would expect some history from the 20th Century.

If it is just a product catalog for Google APIs then buyer beware.

"Please don't post shallow dismissals, especially of other people's work. A good critical comment teaches us something."

https://news.ycombinator.com/newsguidelines.html

You're one to talk.

Reply on news.ycombinator.com