SPLBoost: An Improved Robust Boosting Algorithm Based on Self-Paced Learning

It is known that Boosting can be interpreted as a gradient descent technique to minimize an underlying loss function. Specifically, the being minimized by traditional AdaBoost exponential loss, which proved very sensitive random noise/outliers. Therefore, several algorithms, e.g., LogitBoost and SavageBoost, have been proposed improve robustness of replacing with some designed robust functions. In this work, we present new way robustify AdaBoost, i.e., incorporating learning idea Self-paced Learning (SPL) into framework. design algorithm based on SPL regime, SPLBoost, easily implemented slightly modifying off-the-shelf packages. Extensive experiments theoretical characterization are also carried out illustrate merits SPLBoost.