We extend the standard boosting procedure to train a two-layer classifier dedicated to handwritten character recognition. The scheme we propose relies on a hidden layer which extracts feature vectors on a fixed number of points of interest, and an output layer which combines those feature vectors and the point of interest locations into a final classification decision. Our main contribution is to show that the classical AdaBoost procedure can be extended to train such a multi-layered structure by propagating the error through the output layer. Such an extension allows for the selection of optimal weak learners by minimizing a weighted error, in both the output layer and the hidden layer. We provide experimental results on the MNIST database and compare to a classical unsupervised EM-based feature extraction.