A common assumption in the field of machine learning is that the data used for training and the target data in which the model is applied share the same distribution. While this may hold in many applications, in many other cases the assumption does not hold. We may want to do classification in certain domain in which we do not have enough labeled data but we have enough data in a different but related dataset. The field of transfer learning deal with such cases. The main problem solved by transfer learning is how to use information collected in a different domain in order to improve the inference performance in another domain. A special case is when we want to solve the same task in two datasets that have different marginal distributions. In those cases it is possible to use a method known as covariate shift in which we weight the training distribution using the density ratios in order to make inference on the target distribution. We show how a recently developed method can be used to estimate the weights needed in the covariate shift method and how this can be used in the transfer learning setting obtaining encouraging results.