Thread by @dh7net

#NeurIPS2018 is over! Time to wrap up!

In this thread I'll share what I found the most interesting in the field of ML and creativity.

Everything below worth your time IMO!

First day, the talk by @viegasf & @wattenberg about DataVisualisation was excellent: theoretical and practical.
static.googleusercontent.com/media/research…

Fernanda Viega co-lead the People + AI Research team for Google Cambridge. She makes Art & Design Projects too. More here: fernandaviegas.com

About Dataviz, this #seq2qeq demo show what can be the debugger of the 2.0 software stack: A tools that show how your inference are calculated, and can help you to understand your model, or spot bad data in your training set: seq2seq-vis.io

Another amazing demo: #GANpaint from @henddkn, David Bau & @junyanz89 & Antonio Torralba, give an idea about what can be Phoshop in the future: bit.ly/ganpaint MUST TRY :)

Another amazing demo from @NVIDIA: the rendering of a video game done in real time by a neural network: using #vid2vid

A word about diversity: the women poster session Sunday was great. So many amazing poster.

I wish I took a picture of this black women dressed in a traditional Ghana dress, explaining her badass ML poster.

The full session was a stereotype killer.

The next day, the invited talk from @EdFelten prof @Princeton, is a must if you want to understand politics as a scientist.

Leon Bottou & @obousquet received the "test of time" Award. For this paper:
papers.nips.cc/paper/3323-the…
They studied large scale training, way before it was a thing.

WARNING: Best talk: All the cells of our bodies (not only neurons) communicates using electricity. By hacking this network we could reconstruct damaged organ. @drmichaellevin deliver a perfect flow and made the prof that #comicSans is cool again.
facebook.com/nipsfoundation…
MUST SEE

Joelle Pineau talked about reproducibility in reinforcement learning. She gave a a checklist if you want to summit to #NeurIPS2019 and contribute to solve the reproducibility crisis.

David Ha (@hardmaru) presented on the main stage "World Models", a paper he wrote with Jürgen Schmidhuber. I'm a big fan of this paper arxiv.org/abs/1803.10122

Kunle Olukotun explained that despite the end of the moore law, new architecture will be created for ANN. They will be more efficient (more compute power per Watt). So don't worry, the huge models today will look small tomorow.

Friday morning, during the workshop about visual intelligence, Antonio Torralba made an amazing talk about learning to see and hear. Unfortunately I can't find it online, but he and his team seems to be ahead in this field. I will follow his coming work. 1/3

Antonio Torralba presented an algorithm able to learn end to end the relationship between words (from audio signals) and objects (from pixels). It's fascinating this approach works so well with a relatively small dataset (380k image with an audio description). 2/3

Here an older talk from Antonio Torralba to give you an idea:
But the one during #NeurIPS2018 was even better :-) 3/3

Saturday was the ML + creativity workshop organized by @elluba. This was the main reason I attend to #NeurIPS2018. The concentration of talented people in the field of ML + Art was astonishing. 1/3

#neurips4creativity opening was "Creativity is intelligence" from Kenneth O. Stanley.

He explained why serendipity is the best way to do a search.
This is something that resonate so much to me!

infoq.com/presentations/…
A MUST SEE 2/3

I can't summarize everything there (too many great people, too many great work) but click on #neurips4creativity to learn more. 3/3

Unfortunately @quasimondo & @kcimc were not present but where quoted a lot during #neurips4creativity but also during the whole #NeurIPS2018

Of course this thread capture only a little fraction of #NeurIPS2018 seen thru the lenses of a ML + Creativity enthusiast.
There was so many sessions, so many poster, that it's impossible to tell everything there.

I won't mention names here, but the best part of #NeurIPS2018 was all the informal chat, lunch, dinner, drinks that allows me to see all my ML & Creativity friends, and meets new one!

To conclude: I went at #NeurIPS2018 with 1 question: are progress in this field continue to accelerate or start plateauing? 1/4

It seems there is no new breakthru since the realisation that gradient descent at scale can train huge neural networks with huge database. This way of training networks will continue to delivers for a while.

In that sense we can feel the inflection point of the S curve.

2/4

But because the field is so active, with so many talented people, it's also possible that there is new breakthru that accelerate the field even more and allow it to reach a whole new level in the coming years.

3/4

For the industry this is another story, the number of applications, possible from the research already done today, is HUGE. It will take at least a decade to see them all.

4/4

That's it for me. Looking forward #NeurIPS2018

Meanwhile, please reply with links to all the things I miss.