Lure is a theatrical experiment around the story of Parzival (by Wolfram von Eschenbach), putting the emphasis on the interaction between "AI" and humans. During the live performance, a drummer jams with an AI machine resulting in a musical dialogue. We deployed a tailored flavor of VAEs (Variational Auto-Encoder) to allow the drummer to steer the musical generations towards particular directions during the dialogue. We also displayed a 3D-reduced version of the inferred latent-representation so that the drummer could follow the trajectory he and the AI were taking.
etami is a consortium that gathers major actors from industry and academia with the aim of creating an actionable framework for ethical AI. The focus is heavily put on good and standardized practices as well as on proper lifecycle models for AI-based systems. Algorithmic auditing, whether internal or conducted by a third-party, also plays an important role in creating trustworthiness towards AI and AI institutions.
RAMP is a method and a platform to collectively solve Machine Learning (ML) problems. At first sight, it seems like a classical ML competition, with its more-or-less clean data and scores to optimize, however, it was thought from the beginning as a collaboration tool: the participants submit their whole model (and not simply the predictions), hence they are encouraged to reuse each other's code. Empirically, RAMPs have proved to drastically increase the prediction performance in a single day while remaining an excellent training tool for the participants.
Can we get a fast, sparse predictor without compromising its accuracy? MDDAG (Markov Decision Directed Acyclic Graph) answers this question positively. By casting the inference as a Reinforcement Learning problem wherein the agent selects which features to evaluate on the fly, MDDAG allows a tailored processing for each instance during test-time. Consequently, fast predictions are made for "easy" data, while more complex examples undergo a longer treatment.
Empowerment is an elegant way to formalize how much influence an agent exerts on its environment. It is thus an interesting cost function for intrinsic motivation in Reinforcement Learning. Formally, the empowerment of an agent in a given state is the maximum mutual information (channel capacity) between its actions and its possible future states. Intuitively, a policy that maximizes empowerment will lead the agent to states with maximum "preparedness", ie., where the number of future options is maximal.
ECHOPEN is an open and collaborative project and community, led by a multidisciplinary core of experts and senior professionals with the aim of designing a functional low-cost (affordable) and open source echo-stethoscope (ultrasound probe) connected to a smartphone, allowing the radical transformation of diagnostic orientation in hospitals, general medicine and medically underserved areas. This initiate is aimed for health professionals from southern and northern countries.