DIRU: dendrite-inspired recurrent units for learning chaotic dynamics and neonatal epilepsy time series

Abstract

We introduce the Dendrite-Inspired Recurrent Unit (DIRU), a novel recurrent neural network architecture that incorporates compartmentalized computation inspired by the hierarchical processing of biological dendrites. Unlike classical recurrent models, which encode temporal dynamics within a single homogeneous hidden state, DIRU implements multiple interacting nonlinear compartments within each recurrent unit, enabling structured internal computation, hierarchical state formation, and adaptive dynamical processing. We evaluate DIRU on canonical chaotic systems, including the Lorenz attractor and the Mackey–Glass time-delay system, using multi-step forecasting tasks that explicitly probe long-horizon stability, error growth consistent with bounded chaotic dynamics, and regime preservation. Across all benchmarks, DIRU demonstrates consistently lower prediction error and more stable training dynamics than baseline models across the evaluated settings, with statistically supported performance differences under controlled capacity conditions. To assess real-world biological relevance, we further evaluate DIRU on neonatal electroencephalography (EEG) seizure classification using the Helsinki Neonatal EEG dataset. Operating directly on raw EEG time-series windows under extreme class imbalance, DIRU achieves improved discriminative performance relative to baseline architectures (AUC = 0.884), demonstrating robust detection of transient pathological dynamics without aggressive feature engineering or data augmentation. Together, these results establish multi-compartment recurrent computation as a powerful inductive bias for learning complex nonlinear systems. DIRU provides a promising framework for modeling chaotic dynamics and biological time series, bridging neuroscience-inspired computation with modern deep sequence learning, and offering a principled pathway toward structured recurrent architectures for complex dynamical data.