Deep Probabilistic NMF Using Denoising Autoencoders

Abstract

Abstract-Non Negative Matrix Factorization (NMF) has received considerable attention due to its application in pattern recognition and computer vision. However, the algorithm is sensitive to noise and assumes that the signals in the data can be linearly reconstructed. In this paper, we propose a robust non-linear probabilistic model and develop its optimization algorithm. The proposed model reduces the data to a lower dimensional manifold to get a more meaningful representation and takes into account the noisy nature of the data to improve the clustering performance of NMF. Additionally, our empirical study validates the effectiveness of the proposed method on some benchmark datasets.

A Deep Non-negative Matrix Factorization Model for Big Data Representation Learning

“…For instance, Ye et al ( 2018 ) proposed a deep autoencoder-like non-negative matrix factorization by combining an encoder of matrix factorization and a decoder of the symmetric architecture. Bhattamishra ( 2018 ) designed a deep probabilistic NMF on the basis of autoencoders, in which a probabilistic NMF is built on deep representations of a deep autoencoder with an alternate manner of learning. Bando et al ( 2018 ) proposed a deep variational NMF by interpreting the spectrogram of input data as the sum of a speech spectrogram and a non-negative noisy spectrogram, which is modeled as the variational prior distribution.…”