MayNet: A Neural Network Ensemble Approach Based on May's Theorem for Improved Classification

In this study, we explored the possibility of applying May's Theorem to neural networks and proposed a new unified network architecture called MayNet. MayNet achieves category prediction by integrating multiple neural network "voters" and uses majority voting to determine the final classification result. Experimental results show that MayNet outperforms traditional single neural networks on CIFAR-10 and MedMNIST datasets and has high robustness. The paper compares the performance of MayNet with popular convolutional neural networks (such as ResNet18) on various datasets and demonstrates its superior performance. May's Theorem provides a solid theoretical foundation for the majority voting mechanism in neural network ensembles, ensuring improved decision accuracy through collective judgments of independent voters. MayNet’s architecture innovatively integrates multiple independently trained convolutional neural networks as voters, leveraging majority voting to combine their outputs effectively. This design enhances classification accuracy, robustness, and generalization ability.