Multilayer bootstrap networks
share
Multilayer bootstrap network builds a gradually narrowed multilayer nonlinear network from bottom up for unsupervised nonlinear dimensionality reduction. Each layer of the network is a group of k-centers clusterings. Each clustering uses randomly sampled data points with randomly selected features as its centers, and learns a one-of-k encoding by one-nearest-neighbor optimization. Thanks to the binarized encoding, the similarity of two data points is measured by the number of the nearest centers they share in common, which is an adaptive similarity metric in the discrete space that needs no model assumption and parameter tuning. Thanks to the network structure, larger and larger local variations of data are gradually reduced from bottom up. The information loss caused by the binarized encoding is proportional to the correlation of the clusterings, both of which are reduced by the randomization steps.
READ FULL TEXT