This is my personal (and rarely updated) Blog.
For research, check-out my (more frequently updated) Google Scholar profile.
I started my PhD in October 2015. On the more fundamental side, I study the robustness of biological (or said to be) networks, such as Neural Networks (BDA2016, IPDPS2017, SRDS2017) and bio-molecular metabolic networks (upcoming paper with Andrei Kucharavy and prof. Rong Li from the John Hopkins University Hospital).
On the more practical side, I study the robustness of distributed learning when some of the agents are malicious attackers (trolls poisoning YouTube to make it recommend an anti-vaccine video to people who search for educational content, hackers putting corrupted data in a Hospital to make it mis-diagnose using AI, Learning errors due to bugs, latency etc…). Trying to naively make the reasoning on faulty neurons coarse-grained, my PhD work led to the first mathematically provable Byzantine fault tolerant versions of Stochastic Gradient Descent (SGD is the workhorse of most of today’s industrially deployed machine learning). We did so in the synchronous setting (NIPS2017a) and then on the asynchronous setting (ICML2018a), this has also led to question (ICML2018b) wether convergence is a guarantee for the safety of SGD in the presence of attackers (spoiler: it is not).
Some of these works are patented by the Swiss Federal Institute of Technology in Lausanne (EPFL), my current employer.
I also try to enjoy research with my colleagues, for example, after showing them this video of the Sardine Run, commented by the prophetic voice of David Attenborough, we thought about proving that “animals” (in an extremely simplified version…), like fish in the presence of a predator or human protestors demonstrating in a dictatorship and fleeing brutal police, can learn to gather successfully without coordination or central authority. We proved that they could do so thanks to leaderless reinforcement learning and without the need for a “creationist” imperative rule (if this do that, else do that, etc), as usually assumed in biological algorithmic descriptions of the gathering phenomenon (implementation of the idea by Vladislav Tempez). And while trying to take the idea of multi-agent reinforcement learning (RL) more seriously, we stumbled, thanks to Dr A. Maurer, on the safe interruptibility question (the Big Red Button and scary science-fiction things) raised by L.Orseau (Google Deepmind) and S.Armstrong (Oxford’s Future of Humanity Institute) and tried to port it the multi-agent case. It was not without the mathematical heavy-lifting of an outstanding Master student, Hadrien Hendrikx, that we could have made something out of it (NIPS2017b). In fact, this turned out not to be a sci-fi question that you can just solve by “cutting electricity on AI”; For instance, imagine that you want to slow down your self-driving car, but since it was trained with RL, it would prefer you rather not do so, the sci-fi version would be “the car hacks the electronic break and switch it off”, of course you could solve this by having an extra break, purely mechanical and not controlled by the electronics of the car; good, now imagine that the car has also access to your music playlist, and due to RL, it has already discovered which tracks make you enjoy speed, then it would just play that kind of tracks when it feels that you want to break.
The Interruptibility question actually fits perfectly in the broad effort of building safe distributed learning for safe artificial intelligence, initiated with the Byzantine fault tolerance question described above. Here is a short interview (French) with the Swiss National Radio (RTS) where I was asked to briefly explain both problems.
Before my PhD, I co-founded Wandida.Com with prof. R. Guerraoui and the help of a small grant from Google Zurich and built its first few hundreds mini-lessons. Wandida is now an EPFL online library led by friend, schoolmate and Game-Theorist Dr Lê Nguyen Hoang, with +500 mini-lessons spanning domains from Physics (mostly in French) to Computer Science, Game Theory and Machine Learning (English). The Wandida period was an intense period of learning from outstanding scientists who helped/contributed, such as the French Academy of Science member Gerard Berry; the occasion to discover research communities such as the the Association for Learning Technology, for whom I gave a talk at their 2015 Annual Meeting at the University of Manchester, the less academic eLearning-Africa community (Addis Ababa 2015, Kampala 2014) and some actors of the eLearning community in France.
Before Wandida, I graduated from the Ecole Polytechnique (L’X) and EPFL in 2012 and I briefly worked as an Engineer in solar cells measurements. On the side, I did research on some more fundamental condensed matter Physics of solar cells that I enjoyed thinking about, now published in the Applied Physics Letter.
Besides science, I am passionate about how the Web stimulates (or hampers) social change, with some friends, I co-founded the Web-Media Mamfakinch.com which was recognised by the 2012 breaking borders award from Google and Global Voices. When time allows, I like to write opinion pieces, such as this one for Le Monde (in French) or this short list of contributions on Medias24 (mostly in French but some Darija and Arabic when they had a website for that).
I was born in Morocco and studied there until the equivalent of a university Bachelors in Mathematics and Physics.