In 2009, we proposed a new way to work on high dimension output problem applied to image labeling. This problem consists in assigning a label to each pixel of an image and has multiple application fields such as organ/tumor detection/segmentation in medical imaging or facial landmark detection. Deep neural networks have proven to be efficient on input of high dimension which involves pre-training directed to lower layers of the network. We proposed a new pre-training strategy which is devoted to highest layers of the network. This strategy opens the door to problems with high dimension output to DNN. Structured output problems has been also addressed and in particular through a facial landmark detection application based on deep architectures trained using a training technique developed previously in our laboratory. We successfully applied our new deep neural network architecture on this problem using our old Tesla C1060. Our results will be presented at ICML 2015. However, due to the limited memory available, hyper-parameters search has been reduced.
