Constructive neural-network learning algorithms for pattern classification

摘要: Constructive learning algorithms offer an attractive approach for the incremental construction of near-minimal neural-network architectures pattern classification. They help overcome need ad hoc and often inappropriate choices network topology in that search suitable weights a priori fixed architectures. Several such are proposed literature shown to converge zero classification errors (under certain assumptions) on tasks involve binary mapping (i.e., problems involving binary-valued input attributes two output categories). We present constructive algorithms, MPyramid-real MTiling-real, extend pyramid tiling respectively, real M-ary mappings real-valued multiple classes). prove convergence these empirically demonstrate their applicability practical problems. Additionally, we show how incorporation local pruning step can eliminate several redundant neurons from MTiling-real networks.