Embedding Space Metrics Constrained Fine-Tuning
📅 3–4pm CEST, April 24th, 2026
📍 Online
While fine-tuning Geospatial Foundation Models (GFMs) has become the standard for domain adaptation, it often leads to catastrophic forgetting or 'representation drift,' where the model’s fundamental understanding of semantic relationships is warped. This webinar introduces Embedding Space Metrics Constrained Fine-Tuning (ESM-CFT), an optimisation method in the TerraTorch framework designed to preserve the structural integrity of the latent space during supervised fine-tuning.
Unlike traditional methods that rely solely on cross-entropy loss, ESM-CFT incorporates a geometric constraint based on distance metrics to ensure that the relative positioning of representations in the embedding space remains stable.
The webinar presents promising preliminary experimental results and opens the stage for discussion of future work.
Register here:
Additional Information
Guest speaker: Linh Le, IBM
Linh Le is a data science consultant at IBM Consulting, contributing to customer-focused projects in NLP and computer vision within the customer service sector. Earlier in their career, Linh worked in industrial R&D, developing predictive machine learning solutions for production and maintenance optimisation. The path into remote sensing began during a Master’s thesis at IBM Research, where Linh analysed loss functions for fine-tuning geospatial foundation models, deepening an appreciation for the mechanics behind effective model adaptation.
Moderator: Romeo Klenzer, IBM