Julien Girard @rhapsodos@scholar.social

Why do we want to verify our deep learning algorithm? To have guarantees that our programs actually perform the way we want them to do and to identify failure cases. On a wider scale, the aim is to better align our human goals and standards (economical, ethical) with the actual behaviour of our algorithms. Algorithms are human creations and can embbed all our biases, including unwanted (racist biases among others).
Being able to verify that our algorithm is not racist is actually a big topic

I totally forgot to mention, but I helped to organize a spring school bridging formal methods and deep learning. The aim was to provide an insight on how deep learning can help existing verification techniques, and how to adapt our verification tools for deep learning. We had some really cool presentations with talented speakers, and I hope it will spark some really interesting coopérations ;)

Video recordings and materials should be available soon here: https://formal-paris-saclay.fr/

Will probably begin to use Gnu Linear Programming Kit soon, any good resources to learn it over here?

Help me I'm addicted to TIKZ

Spent the afternoon trying to hunt for VRAM leaks in my torch code, failed miserably. What a cool day!

https://qz.com/547641/theres-an-awful-cost-to-getting-a-phd-that-no-one-talks-about/

First pic of a black hole, coming from EU funded teams

https://twitter.com/UEFrance/status/1115966656189214721

And a cool video to understand what we actually see:
https://www.youtube.com/watch?v=zUyH3XhpLTo

https://www.theregister.co.uk/2019/04/02/tencent_tesla_hacking/

Let's use deep learning they said. It will be better than humans they said.

Another straightforward exemple of why AI safety must be of concern.

Hello everyone!
I'm a first year PhD student in Computer Science, formal methods.

During my thesis, I want to learn how to bring trust in #machinelearning software, and I firmly believe that all the tools we already developed in formal verification can be leveraged for that goal.

Everyone seems to be really hyped about deep learning, but that kind of software is really vulnerable. I want to discover how we can cope with that.

So, let's discuss! #introduction