Deep neural network to extract high-level features and labels in multi-label classification problems

Abstract

Pooling layers help reduce redundancy and the number of parameters in deep neural networks without the need of performing additional learning processes. Although these operators are able to deal with both single-label and multi-label problems they are specifically aimed at reducing feature space. However, in the case of multi-label data, this should also be done in the label space. On the other hand, in spite of their success, existing pooling operators are not ideal when handling (multi-label) datasets that do not have an explicit topological organization. In this paper, we present a deep neural architecture using bidirectional association-based pooling layers to extract high-level features and labels in multi-label classification problems. Our approach uses an association function to detect distinct pairs of neurons that will be aggregated into pooled neurons. In the first pooling layer, our proposal computes the Pearson correlation among the variables as the basis to quantify the association values. In addition, we propose an iterative procedure that allows estimating the association degree among pooled neurons in deeper layers without the need of recomputing the correlation matrix. The main advantage of this deep neural architecture is that it allows extracting high-level features and labels on datasets with no specific topological organization. The numerical results show that our bidirectional neural network helps reduce the number of problem features and labels while preserving network’s discriminatory power.