Can JEPA Bridge Graphs and Language?
→
Europe/Berlin
https://uni-bonn.zoom-x.de/j/64634098374?pwd=XtRxaiL6cdXbebSuRXoCnTwbSWoeJ5.1
Meeting ID: 646 3409 8374
Passcode: 509820
Description
- Speaker: Nicolo Brandizzi
- Abstract: This presentation will introduce the core concepts of the Joint Embedding Predictive Architecture (JEPA) framework proposed by Yann LeCun, highlighting its shift from reconstruction-based to predictive latent modeling as a foundation for autonomous and energy-efficient AI systems.
Building on these ideas, I will outline an experimental extension that applies JEPA principles to align graph and text representations. The approach combines a graph encoder (e.g., Graph Neural Network) and a text encoder (e.g., MiniLM or LLM embedding layer) trained with a predictive latent objective, e.g. encouraging mutual consistency between structural and semantic information.
The agenda of this meeting is empty