[sillysaurusx]

"The email including them got lost to Meta's two-year auto-delete policy by the time I went back to look for it last year. I have a binder with a lot of them printed out, but not all of them."

RIP. If it's any consolation, it sounds like the list is at least three years old by now. Which is a long time considering that 2016 is generally regarded as the date of the deep learning revolution.

[pengaru]

> If it's any consolation, it sounds like the list is at least three years old by now.

In my experience when it comes to learning technical subjects from a position of relative total ignorance, it's the older resources that are the easiest to bootstrap knowledge from. Then you basically work your way forward through the newer texts, like an accelerated replay of a domain's progress.

I think it's kind of obvious that this would be the case when you think about it. Just like how history textbooks can't keep growing in size to give all past events an equal treatment, nor can technical references as a domain matures.

You're forced to toss out stuff deemed least relevant to today, and in technical domains that's often stuff you've just started assuming as understood by the reader... where early editions of a new space would have prioritized getting the reader up to speed in something totally novel to the world.

[moglito]

"considering that 2016 is generally regarded as the date of the deep learning revolution" --

I thought it was 2012, when AlexNet took the imagenet crown?

[sillysaurusx]

That's probably fair. But you'd be hard-pressed to find a DL stack to try out your ideas with prior to 2016, since that's when Tensorflow launched. :)

(Gosh, it's been less than a decade. Time sometimes doesn't fly, considering how much it's changed the world since then...)

[abrichr]

Theano was first released in 2007.

[sillysaurusx]

That’s actually fascinating. Were there many experiments done in it back in the 00’s?

I’m just trying to imagine the things you could do with it back then. 2007 had relatively fast gpus for the time, but certainly nothing compared to today. Yet it’d certainly be enough for MNIST training, which makes me wonder what else could be done.

[abrichr]

You can look at Yoshua Bengio's Google Scholar profile [1] and scroll down to see what they were working on around that time.

Here are some papers with many citations:

- An empirical evaluation of deep architectures on problems with many factors of variation [2]

- Extracting and composing robust features with denoising autoencoders [3]

- Scaling learning algorithms towards AI [4]

[1] https://scholar.google.com/citations?hl=en&user=kukA0LcAAAAJ...

[2] https://scholar.google.com/citations?view_op=view_citation&h...

[3] https://scholar.google.com/citations?view_op=view_citation&h...

[4] https://scholar.google.com/citations?view_op=view_citation&h...

[ladberg]

FWIW in 2016 I was at an ML team at Apple that had been shipping production neural networks on-device for a while already. At the everyone used an assortment of random tools (Theano, Torch, Caffe). I worked on an internal tool that originally started as a Theano fork but was closer to a modern-day Tensorflow XLA (and has since been axed in favor of Tensorflow for most teams).

[jebarker]

Yep, I worked on a production DL system based on Theano in ~2014

[vtantia]

Whoops, Carmack referenced the thread and tagged Ilya in it a veiled request to publish the list - https://twitter.com/ID_AA_Carmack/status/1622673143469858816

[mellosouls]

Sorry - where is that sourced from? Or are you meaning it was a personal communication to you? Or it's a joke?

[sillysaurusx]

He told me.

[mellosouls]

Thank you. I missed the clue when I checked your Twitter profile before asking. Made sense on the revisit. :)

[webmaven]

Which "he", Carmack or Sutskever?