Graph representation learning for multimodal data - challenges and
By Maysam Behmanesh

Maysam Behmanesh will give a talk on Graph representation learning for multimodal data - challenges and innovative methods

Abstract

In this talk, a multimodal graph wavelet convolutional network (M-GWCN) will be introduced as a geometry-aware data analysis approach for multimodal data. M-GWCN is defined in a practical scenario involving heterogeneous modalities, without relying on any prior knowledge indicating correspondences between modalities. It simultaneously captures intra-modality representation by applying multiscale graph wavelet convolution, and cross-modality representation by learning permutations that encode correlations among various modalities. Following that, two innovative graph representation methods, namely Time Derivative Graph Diffusion (TIDE) and Smoothed Graph Contrastive Learning, will be presented. These methods are designed to address structural limitations in geometric data analysis for both unimodal and multimodal datasets.

Biography

Maysam Behmanesh is a Postdoctoral Researcher in the GeomeriX group at the LIX research laboratory of École Polytechnique IP-Paris, a position he has held since April 2022, working under the supervision of Prof. Maks Ovsjanikov. His current research focuses on machine learning, particularly geometric deep learning, with a strong emphasis on graphs and multimodal data.

He earned his PhD in Computer Engineering with a focus on Artificial Intelligence from the University of Isfahan (UI), where he conducted research in the ILS-Lab under the supervision of Prof. Peyman Adibi. His doctoral research centered on geometric multimodal learning, exploring the geometric structure of data for multimodal manifold learning and applying geometric deep learning to graph-based multimodal data. Part of his PhD research was conducted at GIPSA-Lab in Grenoble Institute of Technology, where he was a visiting researcher from 2019 to 2020 under the supervision of Prof. Jocelyn Chanussot.