Structured Output Prediction

Supervised machine learning attempts to find a mapping between an input and output space. Recently, many deep learning methods are introduced in this area because of their superiority in finding the non-linear relation between different spaces. In many supervised problems, we should find a mapping from an input space to a structured output space. Structured variables are a set of random variables which have effect on each other. In this situation, considering the relations between different random variables can improve the strength of a machine learning model. Most of deep learning techniques which have been introduced up to now, are suitable for a simple output space. However, some works are introduced in deep structured output prediction. In this work, we investigate the effect of considering the relation between output variables in a deep learning model. We try to lead the deep model to learn a suitable multimodal distribution on the output variables. Structured output prediction has different applications in real problems. One of these important problems is protein and gene function prediction which is a structured multi label classification problem. Considering the relations between labels can help to extract interpretable features from protein and gene sequences.

Structured Output Prediction

People Involved: Fatemeh Seyyedsalehi, Hamid R. Rabiee, Mahdieh Soleymani

Sharif University of Technology