3IA PhD/Postdoc Seminar #6

Towards safe deep semi-supervised learning
Abstract: Semi supervised learning (SSL) provides an effective means of leveraging unlabeled data to improve a model’s performance. Even if the domain received a considerable amount of attention in the past years, proposed methods lay on assumptions difficult to test in practice and do not come with theortical guarantees on the safeness of using the method. We popose a modification of the SSL framework applicable in various scenarios and provide theoretical guarantees on the safeness of the method even without stong assumptions on the data distribution. We test the safe method proposed to Pseudo-label and reach similar performances on MNIST, we also show the method does not degrade the model's performance in ill situations.

Hybrid combinatorial optimization and machine learning algorithms for energy-efficient water network
Abstract: Pump scheduling is the decision-making problem of planning the pump operations in drinking water networks to minimize the energy cost and satisfy the demand over the day ahead. Along with the difficulty derived from selecting the pump status from binary variables on each time step, Non-convexities arise from the static head-flow relations. The resulting non-convex Mixed-Integer Non-Linear Program (MINLP) is intractable in practice for standard mathematical programming tools. To address this issue, we developed a branch-and-cut algorithm based on a polyhedral relaxation of the non-convex constraints. Tightening this polyhedral relaxation may be computationally expensive, but it is a key factor of the efficiency of the overall algorithms. Therefore, we are investigating a smart relaxation aided by machine learning on the non-convex MINLP formulation.

Open discussion on the two contributions