Tweet

This Thread may be Removed Anytime!

Twitter may remove this content at anytime! Save it as PDF for later use!

More from @fchollet

François Chollet

@fchollet

14 Feb

There's a pretty strong relationship between one's self-image as a dispassionate rational thinker and the degree to which one is susceptible to fall for utterly irrational beliefs that are presented with some sort of scientific veneer

The belief in recursive intelligence explosion is a good example: only someone who thinks of themselves as a very-high-IQ hyper-rationalist could be susceptible to buy into it

If you want to fool a nerd, make long, complex, overly abstract arguments, free from the shackles of reality. Throw equations in there. Use physics analogies. Maybe a few greek words

Read 5 tweets

François Chollet

@fchollet

14 Feb

An event that only happens once can have a probability (before it happens): this probability represents the uncertainty present in your model of why that event may happen. It's really a property of your model of reality, not a property of the event itself.

Of course, if the event has never happened before, that implies that your model of how it happens has never been validated in practice. You can model the uncertainty present in what you know you don't know, but you'll miss what you don't know you don't know.

But that doesn't mean your model is worthless. Surely we all have the experience of writing a large piece of code and having it work on first try.

Read 4 tweets

François Chollet

@fchollet

16 Jan

2020 was definitely a step backwards. If you're wondering how great civilizations can end up collapsing: they just have many 2020s in a row over several decades, with exponentially compounding cascade effects at each new development.

Factors of decline are multiplicative. E.g. cultural & educational deterioration leads to an incompetent government. An incompetent government makes a pandemic much worse. A bad pandemic accelerates institutional decline

For the record, I don't think civilization will collapse in the near future (within the next 400 years). Not even as a consequence of catastrophic climate change over the next two centuries. But we will go through some pretty rough patches

Read 6 tweets

François Chollet

@fchollet

6 Jan

DALL-E is the kind of application that you'd expect deep learning to be able to pull off in theory (people have been building various early prototype of text-guided image generation since 2015) that becomes really magical when done at a crazy scale.

As usual with deep learning, scaling up is paying off.

https://twitter.com/fchollet/status/1294413865610956800

In the future, we'll have applications that generate photorealistic movies from a script, or new video games from a description. It's only a matter of years at this point.

https://twitter.com/fchollet/status/1294413865610956800

Read 5 tweets

François Chollet

@fchollet

5 Jan

Here's an overview of key adoption metrics for deep learning frameworks over 2020: downloads, developer surveys, job posts, scientific publications, Colab usage, Kaggle notebooks usage, GitHub data.

TensorFlow/Keras = #1 deep learning solution.

Note that we benchmark adoption vs Facebook's PyTorch because it is the only TF alternative that registers on the scale. Another option would have been sklearn, which has massive adoption, but it isn't really a TF alternative. In the future, I hope we can add JAX.

TensorFlow has seen 115M downloads in 2020, which nearly doubles its lifetime downloads. Note that this does *not* include downloads for all TF-adjacent packages, like tf-nightly, the old tensorflow-gpu, etc.

Read 13 tweets

François Chollet

@fchollet

4 Jan

Here's a word-level text generation example with LSTM, starting from raw text files, in less than 50 lines of Keras & TensorFlow. colab.research.google.com/drive/1B9yLXcJ…

Of course, I should point out it's not 50 lines because Keras has some kind of built-in solution for text generation (it doesn't). It's 50 lines because Keras makes it easy to implement anything. It only uses generic features.

It uses a utility to read text files, a text vectorization layer (useful for any NLP), the LSTM layer and the functional API, the callbacks infrastructure, and the default training loop.

Read 4 tweets

Share this page!

François Chollet

Try unrolling a thread yourself!

More from @fchollet

François Chollet

François Chollet

François Chollet

François Chollet

François Chollet

François Chollet

Did Thread Reader help you today?

Like this author's thread?