Core Concepts
Manifold-Aligned Graph Regularization (MAGR) enhances AQA models for continual assessment challenges by aligning old features with evolving data manifolds.
Abstract
Action Quality Assessment (AQA) evaluates diverse skills beyond recognition.
Continual Learning (CL) aids in adapting to non-stationarity but faces challenges like catastrophic forgetting.
Continual AQA (CAQA) refines AQA models using sparse new data without forgetting.
MAGR aligns old features with current manifolds, outperforming baselines in AQA datasets.
A comprehensive benchmark study validates MAGR's effectiveness in continual assessment.
Stats
Experiments zeigen, dass MAGR bis zu 6,56%, 5,66%, 15,64% und 9,05% Korrelationsgewinne auf verschiedenen Datensätzen erzielt.
Quotes
"Wir schlagen Manifold-Aligned Graph Regularization (MAGR) vor, um alte Merkmale mit aktuellen Manigfalten auszurichten."