Neural timescales from a computational perspective (2409.02684v2)

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.

• The paper introduces a triad of research pathways—data analysis, mechanistic modeling, and task-optimized neural networks—to quantify neural timescales.
• The paper demonstrates that diverse recording modalities and modeling approaches capture distinct biophysical properties and network dynamics.
• The paper shows that task performance benefits from varied neural timescales, offering practical insights into cognitive and behavioral functions.

Neural Timescales from a Computational Perspective

This paper presents an overview of computational perspectives on the concept of neural timescales, providing a structured examination of how these timescales can be distilled into quantitative and testable hypotheses. The authors elucidate a triad of complementary research pathways: data analysis, mechanistic modeling, and task-optimized models in machine learning. These explorations collectively aim to provide a more rigorous understanding of the relationship between neural timescales, brain dynamics, and behavior.

Estimating Timescales from Neural Recordings

The paper begins by addressing the various methodologies used to estimate neural timescales from different recording modalities, such as spiking activity, LFP, ECoG, calcium imaging, and fMRI BOLD. It underscores that these modalities vary in their spatiotemporal resolutions and physiological bases (e.g., neuronal membrane potential vs. metabolic processes), which impacts the estimation and interpretation of neural timescales.

Timescale estimation methods can broadly be categorized into "model-free" and "model-based" approaches. Model-free methods make minimal assumptions about the autocorrelation shape of neural signals and are straightforward to compute. In contrast, model-based methods assume specific mathematical forms for the autocorrelation function, accounting for multiple timescales and oscillatory components. These latter methods can provide a more detailed and nuanced understanding of the timescales at different levels of neural activity.

Mechanistic Models of Timescales in the Brain

Mechanistic models provide insights into the origins of neural timescales, both at the level of individual neurons and network interactions. Cellular and synaptic mechanisms contribute significantly to timescales, where various biophysical time constants (e.g., membrane, synaptic, and adaptation time constants) shape the dynamics of single neurons. However, the paper suggests that these cellular properties alone are insufficient to explain the full complexity of observed neural timescales. Network connectivity, which includes both local and long-range interactions, plays a pivotal role.

The strength of connectivity, presence of clustered connections, and balance of excitation and inhibition create a rich tapestry of dynamic states, leading to varied and often longer timescales. It further highlights the role of neuromodulatory inputs in dynamically shaping these timescales, allowing neural circuits to adapt to different computational and behavioral demands.

Computational Benefits of Diverse and Flexible Timescales

The functional implications of the diversity and flexibility of neural timescales are explored using task-optimized artificial neural networks (ANNs). These ANNs, including recurrent neural networks (RNNs) and spiking neural networks (SNNs), are trained to perform tasks mimicking cognitive and behavioral experiments. By analyzing the emergent dynamics of these networks, the paper demonstrates that task performance benefits from a repertoire of neural timescales. These emergent timescales enable efficient representation and processing of temporal information, which is vital for tasks such as working memory, decision-making, and sensory processing.

Furthermore, the paper shows that incorporating and optimizing timescale-related parameters (e.g., synaptic and membrane time constants) improves the computational capacity of these networks. The findings from ANN models offer insights into how neural circuits might naturally organize themselves to perform complex, time-dependent computations.

Implications and Future Directions

The paper's integrative approach provides a comprehensive framework for understanding neural timescales, highlighting the need for combining empirical and computational methods. One key implication is that neural timescales can serve as a standardized metric for characterizing neural dynamics across different levels and modalities of brain activity. This standardization could facilitate cross-paper comparisons and meta-analyses.

Future research directions include the development of more sophisticated timescale estimation methods and the creation of curated databases of neural timescales across various conditions and species. The paper also advocates for further exploration of how different experimental and computational models can be constrained and validated against empirical timescale measurements, enhancing our understanding of the underlying mechanisms and functions of neural timescales.

Overall, this paper underscores the importance of neural timescales as a fundamental aspect of neural computation, bridging the gap between microscopic cellular processes and macroscopic cognitive functions. The insights gained from these computational approaches are poised to advance our knowledge of brain dynamics and inform the development of more effective artificial neural systems.