Transfer functions: hidden possibilities for better neural networks.

摘要: Sigmoidal or radial transfer functions do not guarantee the best gen- eralization nor fast learning of neural networks. Families parameterized trans- fer provide flexible decision borders. Networks based on such should be small and accurate. Several possibilities using different types in models are discussed, including enhance- ment input features, selection from a fixed pool, optimization parameters general type functions, regularization large networks with heterogeneous nodes constructive approaches. A new taxonomy is proposed, allowing for derivation known by additive multiplicative combination activation output functions.