A Novel Sparse Autoencoder for Modeling High-dimensional Sensory Data

Sparse autoencoders are used to extract important features that can be used in classification or regression applications. In this paper we present a novel sparse autoencoder for modeling high-dimensional sensory data that allows the user to set the sparsity level and can be used for both off-line or on-line learning applications. The encoder starts by generating random basis functions and adjusts the parameters of the basis functions as data arrives for training. After training a sensory data can be represented by a linear combination of a few number of basis functions. Unlike other
autoencoders our sparse encoder does not require special preprocessing of the sensory data. Potential applications of the autoencoder among others include the realization of advanced feature detectors and signal processing methods. We evaluated the performance of the method on standard image data from the literature and found that our autoencoder gives results comparable to the results reported in the literature.