Visualisation of high-dimensional data using an ensemble of neural networks

We present a new application for ensembles on the task of visualisation. Ensemble methods are known forwarding off overfitting in learning tasks. The task of interest in this work is visualisation via dimensionality reduction: we take the view that each high-dimensional data item is the image, under a smooth mapping, of a two-dimensional latent coordinate. Learning the mapping from latent to data space may be viewed as a regression task which we address by employing an ensemble of neural networks. The inputs to the regression are the latent coordinates and the targets are the high-dimensional data items. However, apart from learning the mapping, we must also learn the latent coordinates so that they meaningfully represent the high-dimensional dataset in the latent space. To that purpose, we adopt an iterative scheme that alternates between learning the mapping and learning the latent coordinates. Although, learning both the mapping and the latent coordinates sounds like a recipe for overfitting, the employment of an ensemble successfully prevents it. Moreover, we show how the algorithm can be also used for visualisation of class posterior probabilities. We demonstrate the method on high-dimensional datasets. The results serve as a promising token for ensembles in visualisation tasks.