Research on New Neural Networks: Local-World Model Neural Network (LWM-Net)

Traditional Deep Neural Networks (DNNs) rely heavily on constructing a unified, globally shared representation space to handle complex, high-dimensional, and dynamic data. Although this global convergence mechanism has achieved tremendous success on static datasets, it exhibits severe bottlenecks—such as catastrophic forgetting, massive computational overhead, and a lack of real-time adaptability—when confronted with non-stationary environments, incremental learning tasks, and high energy-efficiency requirements. To overcome these limitations, this paper proposes a novel neural network architecture: the Local-World Model Neural Network (LWM-Net). The core philosophy of LWM-Net is to break the traditional paradigm of global representation by treating every individual neuron within the network as an autonomous "micro-world simulator" capable of self-learning. Instead of merely performing linear weighting and non-linear activation on input signals, each neuron internally maintains a dynamic model of its local input environment (i.e., a local-world model). This internal model actively predicts the spatiotemporal distribution and state evolution of local inputs, self-organizingly updating its internal states based on the resulting prediction errors. This paper provides a detailed exposition of the micro-neuron architecture, the macro-network topology, and the dual-loop learning algorithms encompassing forward multi-level evolution and backward error localized emergence. Furthermore, we mathematically prove the superiority of LWM-Net in terms of information entropy minimization and local stability. Finally, we explore the potential application scenarios of LWM-Net in time-series forecasting, continual learning, and edge computing. The research demonstrates that LWM-Net offers a promising new pathway for constructing next-generation brain-inspired artificial intelligence systems characterized by high robustness, low power consumption, and powerful adaptability.