A Robust Interval Autoencoder
Lev Vladimirovich Utkin, Anna V. Podolskaja, Vladimir S. Zaborovsky · 2017
A new autoencoder dealing with interval-valued or set-valued training data is studied in the paper. The first main idea underlying the autoencoder is based on transforming the interval-valued reconstruction error produced by imprecise data to the reconstruction error produced by an extended set of precise training data. The training set is extended, but every new training element has an unknown probability. The second idea is the robust strategy that is the autoencoder minimizes the upper bound for the set of reconstructed error measures with respect to parameters of the neural network. The autoencoder can be used for the dimensionality reduction and for precise robust representation of interval data. Numerical experiments illustrate the proposed interval autoencoder.