Neighborhood linear embedding for intrinsic structure discovery

In this paper, an unsupervised learning algorithm, neighborhood linear embedding (NLE), is proposed to discover the intrinsic structures such as neighborhood relationships, global distributions and clustering property of a given set of input data. This algorithm eases the process of intrinsic structure discovery by avoiding the trial and error operations for neighbor selection, and at the same time, allows the discovery to adapt to the characteristics of the input data. In addition, it is able to explore different intrinsic structures of data simultaneously, and the discovered structures can be used to compute manipulative embeddings for potential data classification and recognition applications. Experiments for image object segmentation are carried out to demonstrate some potential applications of the NLE algorithm.