Neural timescales from a computational perspective

TL;DR

This review addresses how neural timescales are defined and measured across brain regions and tasks, arguing that timescales reflect multi-scale dynamics tied to environment and behavior. It integrates three computational directions—measurement methods, mechanistic circuit models, and functional models from machine learning—to distill diverse empirical findings into quantitative theories. It highlights biophysical, cellular, and network mechanisms that generate timescale diversity, and demonstrates how task-performing networks reveal the computational roles of these timescales, including the benefits of heterogeneous time constants and multi-timescale representations. The work suggests that constraining models with observed timescales can improve biological realism in AI systems and sharpen our understanding of dynamic neural computation in naturalistic settings.

Abstract

Neural activity fluctuates over a wide range of timescales within and across brain areas. Experimental observations suggest that diverse neural timescales reflect information in dynamic environments. However, how timescales are defined and measured from brain recordings vary across the literature. Moreover, these observations do not specify the mechanisms underlying timescale variations, nor whether specific timescales are necessary for neural computation and brain function. Here, we synthesize three directions where computational approaches can distill the broad set of empirical observations into quantitative and testable theories: We review (i) how different data analysis methods quantify timescales across distinct behavioral states and recording modalities, (ii) how biophysical models provide mechanistic explanations for the emergence of diverse timescales, and (iii) how task-performing networks and machine learning models uncover the functional relevance of neural timescales. This integrative computational perspective thus complements experimental investigations, providing a holistic view on how neural timescales reflect the relationship between brain structure, dynamics, and behavior.

Figures (5)

• Figure 1: Graphical summary. This review discusses three main directions, demonstrating conceptually (i) how estimation methods contribute to advancing the measurement of timescales from different experimental paradigms (e.g., recording modalities), (ii) how mechanistic models can help in uncovering the underlying circuit origins of different timescales, and (iii) how functional modeling facilitates investigation of their computational relevance.
• Figure 2: Diversity of neural timescales across and within brain areas.a. Average estimated timescales from spontaneous neural activity measured by different recording modalities consistently increase along the cortical hierarchy, but their values differ by orders of magnitude (data from murray_hierarchy_2014gao_neuronal_2020manea_intrinsic_2022). b. Neural timescales are also diverse across individual neurons within one brain area (data from cavanagh_autocorrelation_2016). MT: medial-temporal area, LIP: lateral intraparietal area, OFC: orbitofrontal cortex, LPFC: lateral prefrontal cortex, DLPFC: dorsolateral prefrontal cortex, ACC: anterior cingulate cortex.
• Figure 3: Different methods for estimating neural timescales.a. Neural activity fluctuations can be characterized by various summary statistics such as the autocorrelation or delayed mutual information (AC / MI, top), or information content in neural activity about a preceding stimulus (bottom) (adapted from runyan_distinct_2017rudelt_embedding_2021rudelt_signatures_2023). b. Methods for estimating timescales ($\tau$) from summary statistics generally vary along two axes: (i) Computational complexity (vertical), (ii) Number of features they capture in data (horizontal): from the approximate decay rate to capturing autocorrelation shape by including a baseline representing slow processes across trials or oscillatory components, to including task-related components and prior knowledge about the data (panels are adapted from watanabe_atypical_2019murray_hierarchy_2014donoghue_how_2022spitmaan_multiple_2020zeraati_flexible_2022).
• Figure 4: Mechanisms underlying generation and modulation of neural timescales. Computational models show that neural timescales are shaped by cellular biophysical properties (a, e.g., membrane, synaptic, and adaptation time constants) and network interactions (b, e.g., spatially structured or clustered connectivity, inhibition). Additionally, timescales can be modulated by neuromodulatory signals and sensory inputs from the environment (c). SFA: spike-frequency adaptation; STD/F: short-term synaptic depression/facilitation.
• Figure 5: Task-performing models reveal functional relevance of timescales.a. Network models are set up to achieve input-output transformations similar to tasks performed by animals and humans. b. Model parameters, such as network weights and neuronal time constants, can be manually tuned when feasible or computationally optimized via gradient descent. c. A certain geometry of population dynamics is desirable for network computation, such as fixed points, line attractors, and limit cycles. d. Activity timescales and fitted time constants in network models match experimental data.