A Modified Bottleneck Neural Network for Dimensionality Reduction

Abstract. We present a modification of the bottleneck neural network for dimensionality reduction. We call our scheme a modified bottleneck network (MBNN). Unlike a traditional bottleneck network for dimensionality reduc-tion, an MBNN uses the class information and thus the transformed data can be suitably used for classification problem. We also propose a new tech-nique to create ensembles of neural networks using multiple projections of the same data obtained from different MBNNs. We justify the suitability of the proposed method by some experiments on some classification problems.