Greedy Compositional Clustering for Unsupervised Learning of Hierarchical Compositional Models

This paper proposes to integrate a feature pursuit learning process into a greedy bottom-up learning scheme. The algorithm combines the benefits of bottom-up and top-down approaches for learning hierarchical models: It allows to induce the hierarchical structure of objects in an unsupervised manner, while avoiding a hard decision on the activation of parts. We follow the principle of compositionality by assembling higher-order parts from elements of lower layers in the hierarchy. The parts are learned greedily with an EM-type process that iterates between image encoding and part re-learning. The process stops when a candidate part is not able to find a free niche in the image. The algorithm proceeds layer by layer in a bottom-up manner until no further compositions are found. A subsequent top-down process composes the learned hierarchical shape vocabulary into a holistic object model. Experimental evaluation of the approach shows state-of-the-art performance on a domain adaptation task. Moreover, we demonstrate the capability of learning complex, semantically meaningful hierarchical compositional models without supervision.