PROJECT TITLE :

Unsupervised Feature Learning Architecture with Multi-clustering Integration RBM

ABSTRACT:

In this paper, we present a novel unsupervised feature learning architecture that consists of a multi-clustering integration module and a variant of RBM that we call the multi-clustering integration RBM. Both of these components are referred to collectively as the multi-clustering integration RBM (MIRBM). Without any prior knowledge or labels, the multi-clustering integration module uses three different clusterers (K-means, affinity propagation, and spectral clustering algorithms) to obtain three distinct clustering partitions (CPs). Then, a strategy of voting that requires unanimity is used in order to produce a local clustering partition (LCP). The novel MIRBM model is a central component of the proposed unsupervised feature learning architecture, where it is used for feature encoding. The LCP, which is an unsupervised guidance, is integrated into one step of contrastive divergence (CD1) learning to guide the distribution of the hidden layer features. This is a novel approach, and its novelty lies in this integration. During the training process, the hidden and reconstructed hidden layer features of the MIRBM model in the proposed architecture have a tendency to converge toward one another for the instance that is located in the same LCP cluster. During the training process, the centers of each LCP tend to move farther apart from one another as much as is physically possible within the hidden and reconstructed hidden layers. Experiments show that the proposed unsupervised feature learning architecture has a more powerful capability for feature representation and generalization than the state-of-the-art models for clustering tasks in the Microsoft Research Asia Multimedia (MSRA-MM)2.0 dataset. This is demonstrated by the fact that the proposed architecture was able to outperform the state-of-the-art models.