Learning Multiple Latent Variables with Self-Organizing Maps

Inference of latent variables from complicated data is one important problem in data mining. The high dimensionality and high complexity of real world data often make accurate inference difficult. We approach this challenge with a neural architecture we call Conjoined Twins, which is a two-layer feed forward network with a Self-Organizing Map (SOM) as its hidden layer. Its output layer can preferentially use different numbers (k) of SOM winners for the inference of different latent variables. We introduced this architecture in our previous work. In this paper we propose an automated procedure for the customization of k and demonstrate the effectiveness of the method by the inference of two physical parameters of icy planetary surfaces from spectroscopic data.