Integrating fuzzy neural networks and elastic weight consolidation for continual learning under non-stationary data streams

Neuro-fuzzy systems combine neural network learning with fuzzy logic to robustly represent qualitative, uncertain, and imprecise knowledge, making them especially effective for data mining and knowledge discovery. With ongoing digitalization, the volume and velocity of data streams have increased, creating a pressing need for neuro-fuzzy methods that operate reliably in non-stationary environments. While conventional neuro-fuzzy models exhibit strong performance on individual tasks, they struggle to adapt in dynamic environments where task distributions evolve over time, leading to significant forgetting of previously learned knowledge. To address this limitation, a novel approach that combines Fuzzy Neural Networks (FNN) with Elastic Weight Consolidation (EWC) is proposed to address catastrophic forgetting in continual learning scenarios. The Fuzzy Radial Basis Function Neural Networks (FRBFNN) form fuzzy neural networks, which represent a class of fuzzy neural systems that combine the reasoning mechanism of fuzzy inference with the learning capability of neural networks. By leveraging this fusion structure, FRBFNN are particularly effective in modeling complex and highly nonlinear relationships within data. The EWC mechanism is employed to preserve knowledge by restricting changes to parameters critical for previously learned tasks. Furthermore, the fuzzy rule parameters—including centers, widths, and membership functions—are designed to evolve dynamically, enabling self-organization in response to new data distributions. This method addresses a key gap in existing research, where neuro-fuzzy models lack the capacity for continuous learning. Experimental results demonstrate that EWCFNN outperforms conventional neuro-fuzzy models, highlighting its effectiveness in mitigating catastrophic forgetting and interpretability.