Neural network cubes (N-cubes) for unsupervised learning in gray-scale noise

We consider a class of auto-associative memories, namely, N-Cubes (neural-network cubes) in which 2D gray-level images and hidden sinusoidal 1D wavelets are stored in cubical memories. First, we develop a learning procedure based upon the least-squares algorithm. Therefore, each 2D training image is mapped into the associated 1D waveform in the training phase. Next, we show how the recall procedure minimizes errors among the orthogonal basis functions in the hidden layer. As a 2D image corrupted by noise is applied to an N-Cube, the nearest one of the originally stored training images would be retrieved in the recall phase. Simulation results confirm the efficiency and the noise-free properties of N-Cubes.