MULTI-LABEL METRIC LEARNING WITH BIDIRECTIONAL REPRESENTATION DEEP NEURAL NETWORKS

Tao Zheng, Ivor Tsang, Xin Yao

Abstract:

Multi-Label Learning task simultaneously predicting multiple labels has attracted researchers' interest for its wide application.
Metric Learning crucially determines the performance of the k nearest neighbor algorithms, the most popular framework handling the multi-label problem.
However, the existing advanced multiple-label metric learning suffers the inferior capacity and application restriction.
We propose an extendable and end-to-end deep representation approach for metric learning on multi-label data set that is based on neural networks able to operate on feature data or directly on raw image data.
We motivate the choice of our network architecture via a Bidirectional Representation learning where the label dependency is also integrated and deep convolutional networks that handle image data.
In multi-label metric learning, instances with the more different labels will be dragged the more far away, but ones with identical labels will concentrate together.
Our model scales linearly in the number of instances and trains deep neural networks that encode both input data and output labels, then, obtains a metric space for testing data.
In a number of experiments on multi-labels tasks, we demonstrate that our approach is better than related methods based on the systematic metric and its extendability.

Keywords: metric learning, representation learning, multi-label classification, multi-output