Adnan Darwiche from the University of California will present the South African Centre for Artificial Intelligence seminar with a talk entitled, "Reasoning about the behavior of AI systems"
Adnan will discuss the compilation of some common machine learning systems into symbolic and tractable representations that precisely capture their input-output behavior. This includes classifiers based on neural networks, Bayesian networks and random forests. I will show how the compiled symbolic representations can be used to explain and verify system behavior, including bias, in addition to quantifying system robustness. I will also discuss a new class of tractable, machine learning models: Testing Arithmetic Circuits (TACs), which are as expressive as neural networks. The structure of TACs can be synthesized from domain knowledge and their parameters can be learned from labeled data using gradient descent. However, one can reason about the behavior of TACs and provide some guarantees that are invariant to how TACs are trained from labeled data..
Adnan Darwiche is a professor and former chairman of the Computer Science department of UCLA. He directs the Automated Reasoning Group , which focuses on symbolic and probabilistic reasong and their applications, including machien learning. Professor Darwiche is an AAAI and ACM Fellow. He is a former editor-in-chief of the Journal of Artificial Intelligence Research (HAIR) and author of "Modelling and REasong with Bayesian Networks", by Cambridge University Press.