View profile

Mostly Harmless AI

Mostly Harmless AI
By Alejandro Piad Morffis • Issue #6 • View online
🖖 Welcome to another issue of the Mostly Harmless AI newsletter.
Machine learning is the hottest topic in AI these days. One of the biggest challenges, though, is that it requires vast amounts of data. Finding good labelled data is very hard, and using unlabelled data is difficult.
In this issue, we’ll take a look at a promising machine learning technique that sits between these two extremes and promises to bring the best of both paradigms: self-supervised learning.

🗞 What's new
Facebook recently released a blog post about their new project, Learning from Videos. One of these is SEER, a new vision model that beats state-of-the-art in several well-known benchmarks. These are all instances of the same underlying technique: self-supervised learning. What is it, and why is it hitting the news?
Facebook AI
Our project, Learning from Videos is designed to build AI that automatically learns audio, textual & visual representations from publicly available videos on Facebook. Learn how this will improve AI-powered products — starting with Reels’ recommendations:
You can read a deep dive from Yann LeCun itself, but the TL;DR is this. Self-supervised learning (SSL) is a technique for leveraging vast amounts of unlabeled data by using the data’s own structure as supervised labels. Let’s unpack that.
The most famous self-supervised models are probably modern Transformers: BERT, GPT, and company. They are trained on lots of lots of text, but, here is the key part: we don’t need labels. Instead, we take a sentence, hide some words, and train the model to predict the unknown words from the known ones. Pretty clever, right?
By forcing the model to predict the missing bits, we are implicitly making it learn the underlying structure of the data. With text, it is pretty straightforward (well, after they tell you about it…) but with images is a whole different story.
This is where Facebook’s new SEER model enters the picture. It is a major breakthrough, akin to what BERT and similar models meant for NLP a few years ago. And the best part, they open-sourced it.
📚 For learners
If you’re interested in learning more about self-supervised learning, check out this awesome Github list of resources. It has a large collection of papers and resources on self-supervised learning in several domains, from images to audio, to natural language processing. And very up-to-date.
You can also take a look at PapersWithCode section on self-supervision, which collects some of the most relevant papers with links to their Github implementations.
And finally, here’s a not-so-recent talk from Yann LeCun that you’ll enjoy on the topic.
Self-Supervised Learning
Self-Supervised Learning
🔨 Tools of the trade
Probably the best all-in-one resource for self-supervised learning in computer vision is Facebook’s VISSL project. It contains implementations of the most popular SSL models in Pytorch.
GitHub - facebookresearch/vissl:
If you prefer Tensorflow, here’s Google implementation of SimCLR, one of those state-of-the-art models.
GitHub - google-research/simclr
🍿 Recommendations
Taking a step away from the hard technical topic, today I want to recommend one of the best sci-fi series in recent times, and probably the best modern take on the issue of artificial consciousness: Westworld.
If you haven’t seen it yet, I cannot recommend it enough. The ethical and philosophical topics are deeply treated, and the drama is top-notch, including some mind-blowing storyline twisting and turning that will leave you confused at times, but always impressed.
Westworld (TV Series 2016– ) - IMDb
🎤 Word of mouth
This week’s AMA was not as crowded as usual but also packed with interesting questions. We talked about Explainable AI, how to keep up with tech, the value of getting a Tensorflow certification, why not to do a PhD, and more.
Alejandro Piad Morffis
Hey folks 🖖!

🎙️It's Saturday again, let's do another AMA round, just for fun! Ask me anything about CompSci, AI, machine learning, or anything else.

I won't promise I have the answer (I most likely don't) but maybe we can figure it out together.

Let's do this 👇!
And tying back to self-supervise learning, here’s an interesting discussion in HackerNews about the challenges and potential issues with this paradigm. As usual, take it with a grain of salt.
👥 Community
In this issue, I want to recommend you to follow these two Twitter accounts. They are still small (in Twitter numbers), but very productive, and I’ve enjoyed interacting with them both a lot this last few months: Tolani @JaiyeTikolo, and Dimas @DreamOnShadows. Give them both the chance to fill your timeline with interesting stuff all around AI and related tech.
☕ Homebrew
On my end, I’ve been working on a lot of stuff, including new scripts for podcast episodes. If you want to weigh in, I would love to get your opinion on what topics to touch on first.
Alejandro Piad Morffis
🎙️ Would you like to help me craft my next podcast episode?

Here's how:
👉 I'll suggest some topics next, one per tweet.
👉 You like the one(s) you prefer I make first...
👉 Comment with questions and suggestions of what you want to see covered in each topic.

Let's do it 😉
Finally, I’ve cooked up a small and dirty script to schedule Twitter threads from Trello, using Python. Here’s a short thread about it:
Alejandro Piad Morffis
I've been struggling for some time with Twitter scheduling.

There are great apps out there, like @FeedHive_io or @tweetastic_app, precisely for this!

But I'm just too nerd. Here's my (incredibly suboptimal and very personal) solution 👇
And here’s the script. Feel free to remix it and reuse it as you see fit. It’s 100% open-source code.
👋 That’s it for now. Please let me know what do you think of this issue, what would you like to see more or less of, and any feedback you want to share. If you liked this newsletter, consider subscribing (in case you’re not) and forwarding it to those you love. It’s 💯 free!
Did you enjoy this issue?
Alejandro Piad Morffis

A weekly newsletter on all things AI, including recent news, hot resources, and interesting conversations happening all around the Internet.

If you don't want these updates anymore, please unsubscribe here.
If you were forwarded this newsletter and you like it, you can subscribe here.
Powered by Revue