Marginalizing Corrupted Features

摘要: The goal of machine learning is to develop predictors that generalize well test data. Ideally, this achieved by training on an almost infinitely large data set captures all variations in the distribution. In practical settings, however, we do not have infinite and our may overfit. Overfitting be combatted, for example, adding a regularizer objective or defining prior over model parameters performing Bayesian inference. paper, propose third, alternative approach combat overfitting: extend with many artificial examples are obtained corrupting original We show efficient range corruption models. Our approach, called marginalized corrupted features (MCF), trains robust minimizing expected value loss function under model. empirically variety sets MCF classifiers can trained efficiently, substantially better data, also more feature deletion at time.