文摘
This paper presents a highly scalable modular bottleneck neural network and its application to image dimensionality reduction and image transformation. The network is a three-dimensional lattice of modules that implements a complex mapping with full connectivity between two high-dimensional datasets. These datasets correspond to input and output pixel-based images of three airplanes with various spatial orientations. The modules are multilayer perceptrons trained with Levenberg-Marquardt method on GPUs. They are locally connected together in an original manner that allows the gradual elaboration of the global mapping. The lattice of modules is squeezed in its middle into a bottleneck, thereby reducing the dimensionality of images. Afterward, the bottleneck itself is stretched to enforce a specific transformation directly on the reduced data. Analysis of the neural values at the bottleneck shows that we can extract from them robust and discriminative descriptors of the airplanes. The approach compares favorably to other dimensionality reduction techniques.