École Nationale Supérieure de
Techniques Avancées

Abstract

Deep learning (DL) has revolutionized many tasks in recent years, ranging from video processing to machine translation to audio recognition. Data of these problems are usually presented in Euclidean space. This work describes alternative DL models that are adapted to data are generated from non- Euclidean domains. These data are represented as graphs that are derived from topography, where the deepest parts of a cortical sulci are nodes of the graph. In this project, we implemented GraphSage, DCNN, GAT to solve the problem of binary node classification. In a number of experiments on real dataset, we demonstrate that selected methods outperform proposed trivial model. All models have been implemented in Python.