Nonnegative autoencoder with simplified random neural network

A new shallow multi-layer auto-encoder that combines the spiking Random Neural Network (RNN) with the network architecture typically used in deep-learning, is proposed with a learning algorithm inspired by non-negative matrix factorization which satisfies the non-negative probability constraints of the RNN. Auto-encoders equipped with this learning algorithm are tested on typical images including the MNIST, Yale face and CIFAR-10 datasets, and also using 16 real-world datasets from different areas, exhibiting the desired high learning and recognition accuracy. Montecarlo simulations of the stochastic spiking behaviour of this RNN auto-encoder have also been carried out, showing that it can be implemented in a highly parallel manner to achieve substantial speed improvements.