HiperLearn: A High Performance Channel Learning Architecture

A learning system architecture using a channel information representation has been developed. This representation implies that signals are monopolar and local. Monopolar means that data only utilizes one polarity, e.g. positive values, allowing zero to represent no information. Locality derives from a partially overlapping mapping of signals into a higherdimensional space. This implies that a signal is non-zero only within a limited range of the input domain. This combination of monopolarity and locality leads to an efficient sparse representation. Locality of features allows the generation of highly nonlinear functions using a linear mapping. The averaging properties of channel functions allow representation of discontinuities, while implementing an interpolating mapping in other regions. Mapping from an input channel set onto an output channel set, allows the use of confidence statements in data, leading to a low sensitivity to noise in features. The optimization uses a modified projected Landweber method. The sparse monopolar representation together with locality, using individual learning rates, allows a fast optimization, as the system exhibits linear complexity. The monopolarity implies a regularization, providing a better generalization. Experiments on functionality and noise sensitivity are presented. Keywords— Channel representation, monopolar, association, sparse, confidence, learning, learning rate, optimization, neural networks, RBF, SVM, local features, classification, linkage matrix, non-linear mapping, wavelets, function approximation, computer vision, signal processing.