Exploration and generalization in deep learning with SwitchPath activations

This work provides a comprehensive theoretical and empirical analysis of SwitchPath, a stochastic activation function that improves learning dynamics by probabilistically toggling between a neuron standard activation and its negation. We develop theoretical foundations and demonstrate its impact in multiple scenarios. By maintaining gradient flow and injecting controlled stochasticity, the method improves generalization, uncertainty estimation, and training efficiency. Experiments in classification show consistent gains over ReLU and Leaky ReLU across CNNs and Vision Transformers, with reduced overfitting and better test accuracy. In generative modeling, a novel two-phase training scheme significantly mitigates mode collapse and accelerates convergence. Our theoretical analysis reveals that SwitchPath introduces a form of multiplicative noise that acts as a structural regularizer. Additional empirical investigations show improved information propagation and reduced model complexity. These results establish this activation mechanism as a simple yet effective way to enhance exploration, regularization, and reliability in modern neural networks.