Fuzzy Deep Stack of Autoencoders for Dealing with Data Uncertainty

This paper addresses the problem of dealing with uncertainty on neural networks for the specific case of Autoencoders. The recently introduced concepts of `Autoencoders' and `Deep Stacks of Autoencoders' (DSAE) and their use for dimensionality reduction and data compression problems have gained considerable attention and reached very promising results. However, similarly to the traditional neural networks, Autoencoders are deterministic structures that are not very suitable for dealing with data uncertainty, a very important aspect of the real-world applications. In this paper, we propose a fuzzy approach to reduce uncertainty on stacks of Autoencoders by automatically adding qualitative fuzzy information about the data to the input layer. This can be accomplished by adding a fuzzy layer-0 to the stack of Autoencoders that extracts fuzzy knowledge from the crisp data set and includes that knowledge as extra information to the network input. The approach is completely transparent to the network and to the user and, theoretically, can be generalized to any architecture of neural network, including convolutional neural networks. The results presented here are very encouraging and present substantial improvement, especially when dealing with noisy data.