Reduce gap between sequential and statistical pattern recognition in machine learning

Sequence classification has a wide range of applications such as genomic analysis, anomaly detection, and gesture recognition. The majority of classification algorithms from machine learning can not be applied directly to these problems, because the concept of gradient is unknown for distortion-invariant (elastic) functions on sequences. Therefore, the objective is to narrow the gap between sequential and statistical pattern recognition. To achieve our objective, we follow a threefold research agenda: (1) explore the geometry of sequence spaces endowed with an elastic distance function, (2) generalize gradient-based classifiers for feature vectors to elastic classifiers for sequences, and (3) empirically assess the performance of the proposed elastic classifiers by controlled numerical experiments and by comparing them to the state-of-the-art on diverse problems. Based on first results using the simplest elastic classifiers on univariate time series, we expect that the proposed approach has the potential to complement the state-of-the-art in sequence classification for diverse application domains.