GENERATION OF EFFICIENT REPRESENTATIONS IN NEURAL NET ARCHITECTURES USING HIGH ORDER CORRELATIONS.

Summary form only given. The learning and generalization capabilities of single and multiple slab high order architectures were measured on test problems such as contiguity, symmetry detection, and parity. It was found that for certain sets of problems a single slab of high-order units generalize nearly perfectly in cases in which back-propagation shows very little generalization capacity. High-order cascaded-slab architectures are capable of learning very-high-order problems (such as ninth-order parity) which are not handled efficiently by single-slab architectures. These architectures require only integer arithmetic and converge much faster than back-propagation.