Structured prediction for sequential data analysis.

Summary In this thesis, we are interested in machine learning problems in the context of structured outputs with a sequential structure. On the one hand, we consider the problem of learning similarity measures for two tasks: (i) the detection of breaks in multivariate signals and (ii) the problem of temporal distortion between pairs of signals. The methods generally used to solve these two problems depend heavily on a similarity measure. We learn a similarity measure from fully labeled data. We present standard structured prediction algorithms that are efficient for learning. We validate our approach on real data from various domains. On the other hand, we address the problem of weak supervision for the task of aligning an audio recording to the played score. We consider the score as a symbolic representation giving (i) complete information on the order of the symbols and (ii) approximate information on the shape of the expected alignment. We learn a classifier for each symbol with this information. We develop a learning method based on the optimization of a convex function. We demonstrate the validity of the approach on musical data.