We demonstrate a new deep learning autoencoder network, trained by anonnegativity constraint algorithm (NCAE), that learns features which showpart-based representation of data. The learning algorithm is based onconstraining negative weights. The performance of the algorithm is assessedbased on decomposing data into parts and its prediction performance is testedon three standard image data sets and one text dataset. The results indicatethat the nonnegativity constraint forces the autoencoder to learn features thatamount to a part-based representation of data, while improving sparsity andreconstruction quality in comparison with the traditional sparse autoencoderand Nonnegative Matrix Factorization. It is also shown that this newly acquiredrepresentation improves the prediction performance of a deep neural network.
展开▼