<ul><li>Retraining a model using its own predictions and original labels can improve model performance.</li><li>This paper focuses on optimizing the combination of model predictions and labels for binary classification tasks.</li><li>A framework based on approximate message passing (AMP) is developed for analyzing retraining procedures.</li><li>The study derives the Bayes optimal aggregator function for retraining models, showing its superiority in minimizing prediction error.</li></ul>

Self-Boost via Optimal Retraining: An Analysis via Approximate Message Passing

Discover more