As neural networks (NNs) become integral to advanced applications in smart manufacturing, the demand for models that are both accurate and robust continues to grow. A persistent challenge in NN training lies in avoiding local minima, which can hinder the model’s ability to minimize the loss function effectively, both in fitting training data and generalizing to unseen test data, thereby achieving globally optimal performance. To address this, an extension to traditional backpropagation is proposed, incorporating a self-adaptive mechanism that encourages exploration of underutilized regions of the optimization landscape. The method adds an auxiliary objective to the training process, complementing gradient-based exploitation with an exploration component that dynamically adjusts the network’s internal state. A mathematical formulation of the algorithm is provided, and comparative experiments demonstrate that the method achieves lower training loss and superior accuracy. Connections to existing methods, such as momentum and entropy-based regularization, are analyzed, emphasizing the unique contributions of the approach. Finally, the implications for future industrial applications are discussed, where NNs must perform reliably under dynamic, real-world conditions. By enabling smarter, self-critical models, the proposed method advances the development of more reliable and adaptive NNs for smart manufacturing.
Terziyan et al. (Thu,) studied this question.