<ul><li>Researchers revisit Bayesian model averaging (BMA) to ensemble pre-trained foundation models for improved classification on image and text data.</li><li>They introduce trainable linear classifiers to make BMA tractable under foundation models, helping identify which linear heads and frozen features are best suited for a specific dataset.</li><li>Additionally, they propose an optimizable model averaging scheme (OMA) that directly optimizes model ensemble weights to reduce surprise in predictions.</li><li>These approaches aim to leverage future foundation models for enhanced performance in challenging classification tasks.</li></ul>

Revisiting Bayesian Model Averaging in the Era of Foundation Models

Discover more