Nathanaël Fijalkow (CNRS)
Title: Machine Learning Guided Program Synthesis
Abstract: In this talk I'll describe some general solution framework for program synthesis, which is the art of getting a piece of code from its specification rather than the sweat of its programmers. It uses machine learning to make predictions, which raises a lot of (theoretical) questions. I'll discuss using SAT and SMT solvers for data generation and probabilistic context-free grammars for program search.
Anna Lukina (Institute of Science and Technology Austria)
Title: Reliable Learned Controllers
Abstract: While machine learning undoubtedly offers a powerful toolset for system designers, relying on learned models may lead to catastrophic system failure in deployment. In the real world, especially with a human in the loop, a system controller obtained via machine learning can encounter novel, not seen in training, situations. Most learned models will in this case be forced to make ill-informed or even wrong decisions without alerting the human in the loop, possibly leading to severe consequences. I argue that, in contrast to the state-of-the-art, we must iteratively monitor the inner behavior of the model at runtime and actively adapt to changing conditions to give timely warnings about suspicious events, in addition to model verification. Guaranteeing reliable control of and interaction with automated safety-critical systems, which have become omnipresent in our daily lives, is of great importance, but requires outside-the-box thinking.
In this talk, I will describe a few technical insights into the methods from my recent publications. With this, I hope to create a lively discussion on the current and future challenges of learned systems.