The paper proposes several methods to enhance the performance and training of neural networks for classification tasks. Cross-validation is used to optimize network parameters and architecture, while ensembles of similar networks are employed to reduce residual "generalization" errors. The authors argue that cross-validation is a valuable tool for assessing network performance and optimizing its characteristics. They also introduce the concept of ensembles, where multiple networks trained on the same dataset are used to make collective decisions, which can lead to better generalization and fault tolerance. The paper discusses the challenges of local minima in the optimization process and presents models to predict the performance of ensembles, including the majority voting rule and plurality voting rule. Experimental results on two examples—Generalized XOR and a model problem—support the theoretical findings, demonstrating that ensembles can achieve significantly better performance than single networks. The paper concludes that using ensembles with a plurality consensus scheme can significantly improve performance, and that cross-validation is crucial for optimizing network architecture and assessing generalization capabilities.The paper proposes several methods to enhance the performance and training of neural networks for classification tasks. Cross-validation is used to optimize network parameters and architecture, while ensembles of similar networks are employed to reduce residual "generalization" errors. The authors argue that cross-validation is a valuable tool for assessing network performance and optimizing its characteristics. They also introduce the concept of ensembles, where multiple networks trained on the same dataset are used to make collective decisions, which can lead to better generalization and fault tolerance. The paper discusses the challenges of local minima in the optimization process and presents models to predict the performance of ensembles, including the majority voting rule and plurality voting rule. Experimental results on two examples—Generalized XOR and a model problem—support the theoretical findings, demonstrating that ensembles can achieve significantly better performance than single networks. The paper concludes that using ensembles with a plurality consensus scheme can significantly improve performance, and that cross-validation is crucial for optimizing network architecture and assessing generalization capabilities.