Input/Output Deep Architecture for Structured Output Problems

Pre-training of input layers has shown to be efficient for learning deep architectures, solving the vanishing gradient issues. In this paper, we propose to extend the use of pre-training to output layers in order to address structured output problems, which are characterized by internal dependencies between the outputs (e.g. the classes of pixels in an image labeling problem). Whereas the output structure is generally modeled using graphical models, we propose a fully neural-based model called IODA (Input Output Deep Architecture) that learns both input and output dependencies. We apply IODA on facial landmark detection problems where the output is a strongly structured regression vector. We evaluate IODA on two public challenging datasets: LFPW and HELEN. We show that IODA outperforms the traditional pre-training approach.