GeomCA: Geometric Evaluation of Data Representations
This addresses the problem of representation evaluation for researchers and practitioners in machine learning, offering a method that is model-agnostic and applicable to any dimension, though it appears incremental as it builds on geometric and topological analysis without claiming broad SOTA impact.
The paper tackles the challenge of evaluating learned representations without downstream tasks by introducing the Geometric Component Analysis (GeomCA) algorithm, which assesses representation spaces based on geometric and topological properties and demonstrates applicability across various models like contrastive learning, generative models, and supervised learning.
Evaluating the quality of learned representations without relying on a downstream task remains one of the challenges in representation learning. In this work, we present Geometric Component Analysis (GeomCA) algorithm that evaluates representation spaces based on their geometric and topological properties. GeomCA can be applied to representations of any dimension, independently of the model that generated them. We demonstrate its applicability by analyzing representations obtained from a variety of scenarios, such as contrastive learning models, generative models and supervised learning models.