|Topic:||Self-supervised learning for geometric constructions from images|
|Announce as:||Diplomová práce, Bakalářská práce, Semestrální projekt|
|Description:||We consider the problem of geometric constructions in Euclidea (an online game where the task is to find a sequence of construction steps leading from an initial configuration of objects to a given target configuration). This project is a follow-up to our paper from CICM2021 conference where we utilize the Mask R-CNN visual recognition neural architecture together with tree-based search in a supervised setting. The novelty is the direct use of image data as the input, which is in contrast to previous methods that rely on a pre-specified symbolic representation of the problem. The goal of this project is to extend the work to learn to solve geometric constructions in a self-supervised manner, i.e. without access to step-by-step solutions of the target geometric construction problems, which is a major learning challenge. The project will also involve evaluation of the complexity of construction tasks, developing suitable symbolic and visual representations of the construction problems, or automatic theorem proving. For a detailed description please see:
|Bibliography:||Macke, J Sedlar, M Olsak, J Urban, J Sivic. Learning to solve geometric construction problems from images. Conference on Intelligent Computer Mathematics (CICM), 2021.
Euclidea - Geometric Constructions Game with Straightedge and Compass.