Well-M $^3$ N: A Maximum-Margin Approach to Unsupervised Structured Prediction

Unsupervised structured prediction is of fundamental importance for the clustering and classification of unannotated structured data. To date, its most common approach still relies on the use of structural probabilistic models and the expectation-maximization (EM) algorithm. Conversely, structural maximum-margin approaches, despite their extensive success in supervised and semi-supervised classification, have not raised equivalent attention in the unsupervised case. For this reason, in this paper, we propose a novel approach that extends the maximum-margin Markov networks (M $^3$ N) to an unsupervised training framework. The main contributions of our extension are new formulations for the feature map and loss function of M $^3$ N that decouple the labels from the measurements and support multiple ground-truth training. Experiments on two challenging segmentation datasets have achieved competitive accuracy and generalization compared to other unsupervised algorithms such as $k$ -means, EM and unsupervised structural SVM, and comparable performance to a contemporary deep learning-based approach.