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: formal-paris-saclay.fr/

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

#VimTips Or if you are already editing your .vimrc (which is likely since the only reason to source it is if you modified it since vim started), then you can do


because % stands for the current file.

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

The crowdfunding of #Mobilizon has reached its first milestone in 5 days and that's really cool. I really hope it will reach the 2nd one for which Framasoft has pledged to make it federated through ActivityPub.

Nowadays instead of setting up a Doodle, most people I know set up a Framadate. Having a similar shift from Facebook events to Mobilizon would be great.



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.

Grad school etiquette Show more

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 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!

Scholar Social

Scholar Social is a microblogging platform for researchers, grad students, librarians, archivists, undergrads, academically inclined high schoolers, educators of all levels, journal editors, research assistants, professors, administrators—anyone involved in academia who is willing to engage with others respectfully. Read more ...