Perceive Fast, Think Slow: A Cognitive-Inspired Framework for Time Series Analysis

ICLR 2026 Conference Submission15602 Authors

Keywords: Time-series modeling; Multivariate time series; Event-driven segmentation; Dual-Axis Adaptive Attention

TL;DR: PeCo-TS is a cognitive-inspired framework that uses event-driven tokenization and a dual Fast–Slow pathway to achieve state-of-the-art accuracy and efficiency across diverse time-series tasks.

Abstract: Time series modeling faces persistent challenges: fixed-window tokenization misaligns with natural event boundaries, uniform computation wastes capacity on simple patterns, and static architectures cannot adapt to diverse temporal dependencies. We propose **PeCo-TS**, a cognitive-inspired framework that instantiates the principle of “perceive fast, think slow” through three key innovations: (1) *event-driven dynamic-length tokenization* that aligns tokens with semantic boundaries and reduces redundancy, (2) a *Slow–Fast dual-pathway architecture* that separates rapid perception of fine-grained variations from slower abstraction of event-level structures, and (3) *Dual-Axis Adaptive (DA²) attention* that dynamically balances intra-series and inter-series dependencies via learnable gating. Extensive experiments across forecasting, classification, anomaly detection, and imputation demonstrate the broad applicability of PeCo-TS, yielding consistent improvements over Transformer and linear baselines, including 5.6% lower forecasting MSE, 9.3% lower imputation error, higher classification accuracy across UCR/UEA benchmarks, and a 6.7% relative F1 gain in anomaly detection. Beyond accuracy, PeCo-TS achieves favorable efficiency–performance trade-offs by leveraging event-level abstraction and complementary pathway synergy, while its learned boundaries align with real-world regime shifts, providing interpretability. These results establish PeCo-TS as a versatile backbone that unifies efficiency, adaptability, and semantic alignment for diverse time-series applications.

Primary Area: learning on time series and dynamical systems

Submission Number: 15602