Go to ICLR 2026 Conference homepageCORTEX: A Neuroscientifically Inspired Temporal Architecture with Working-Long-term Memory Collaboration
Keywords: Deep learning, Dual Memory Systems, Temporal sequence modeling, Working Memory-Long Term Memory Collaboration, Spiking Neural Network (SNN), Three-Phase Interaction Paradigm
TL;DR: We propose a brain-inspired neural network that mimics mammalian memory, significantly outperforming existing methods on time series tasks like stock prediction and EEG analysis.
Abstract: Time series modeling faces a critical trade-off between adapting to dynamic patterns and maintaining stable long-term representations. To address this, we introduce CORTEX, a novel temporal modeling framework inspired by the dual-memory system of the mammalian brain. CORTEX operationalizes the synergy between working memory (fast adaptation) and long-term memory (stable consolidation) through a modular, neuro-inspired architecture. The framework comprises four core components: (1) a Heterogeneous Shallow Memory with bio-inspired spiking units for rapid, adaptive signal processing; (2) a Unified Dimensionality Reduction Hub that performs topology-preserving compression, analogous to entorhinal sparse coding; (3) a Hierarchical Deep Memory that consolidates long-range dependencies, akin to hippocampal function; and (4) a Joint Decoding Module that integrates information from both memory pathways. We validate CORTEX across diverse domains, demonstrating state-of-the-art performance. Notably, it achieves a 99.46% accuracy in complex EEG signal classification and an R² of 0.837 in volatile financial forecasting, significantly outperforming a comprehensive suite of modern baselines. Our work establishes a powerful and biologically plausible paradigm for temporal sequence analysis, effectively bridging the gap between rapid responsiveness and long-term stability.
Supplementary Material: zip
Primary Area: applications to neuroscience & cognitive science
Submission Number: 11301
Loading