Geometry-Aware Adaptation for Pretrained Models

Name: Geometry-Aware Adaptation for Pretrained Models
Uploaded: 2026-10-02T23:52:56.556025+00:00
Duration: 263 s
Description: LOKI adapts pretrained machine learning models to reliably predict new classes when labels are sparse by exploiting the geometric relationships between existing class embeddings. It replaces standard prediction rules with a Fréchet mean estimator, allowing models to generalize better than traditional methods without needing extra training data for every new class.

LOKI adapts pretrained machine learning models to reliably predict new classes when labels are sparse by exploiting the geometric relationships between existing class embeddings. It replaces standard prediction rules with a Fréchet mean estimator, allowing models to generalize better than traditional methods without needing extra training data for every new class.

Open the full episode