Guest speaker series on Graphical Models and Neural Networks

Deep Neural Networks for Optimization of Graph Structure

Optimization of a graph-structured solution appears ubiquitously in many important applications (e.g., drug discovery and program synthesis). This talk will introduce how one can use machine learning to solve such optimization problems, covering both classic and recently emerging approaches. Especially, I will introduce my works based on the question of how the recently developed deep neural network (DNN)-based solvers can be compared or combined with the classic solvers. The first part will be about the maximum independent set problem; I will introduce a scalable structured prediction framework that bridges the performance gap between DNN-based and classic solvers. The next part will be about the optimization of molecular graph structures for drug discovery. I will demonstrate a training scheme for DNNs based on the guidance of classic genetic algorithms, significantly improving the state-of-the-art results.