The human ability to learn difficult object categories from just a few views is often explained by an extensive use of knowledge from related classes. In this work we study the use of feature relevance as prior information from similar binary classification tasks. An approach is presented which is capable to use this information to increase the recognition performance for learning with few examples on a new binary classification task. Feature relevance probabilities are estimated by a randomized decision forest of a related task and used as a prior distribution in the construction of a new forest. Experiments in an image categorization scenario show a significant performance gain in the case of few training examples.