91ÊÓƵ

Kyoto, Japan -- What if we could peer into the brain and watch how it organizes information as we act, perceive, or make decisions? A new study has introduced a method that does exactly this -- not just by looking at fine-grained neuronal spiking activity, but by characterizing its collective dynamics using principles from thermodynamics.

A team from 91ÊÓƵ and Hokkaido University developed a new statistical framework capable of tracing directional, nonequilibrium neural dynamics directly from large-scale spike recordings, enabling them to show how neurons dissipate entropy as they compute. Their findings reveal how neurons dynamically reshape their interactions during behavior and how the brain¡¯s internal "temporal asymmetry" shifts during task engagement, shedding light on how efficient computation arises.

Traditional approaches to temporal asymmetry often assume that brain signals are relatively steady over time -- a convenient assumption, but one that fails to capture the brain's ever-changing computations. "Real neurons never sit still," says first author Ken Ishihara of Hokkaido University. "Their firing rates and interactions fluctuate from moment to moment. To capture their nonequilibrium behavior, we needed a new kind of model."

To address this, the team extended a classical model from statistical physics -- the kinetic Ising model -- into a flexible, time-varying form. By combining the augmented representation with a new mean-field theory, the framework can estimate entropy flow, a thermodynamic measure of irreversible activity tied to limits on the speed and precision of state changes in dynamical systems.

Applying their method to large datasets from mouse visual cortex, the researchers identified three key signatures of neuronal activity during behavior. The first is sparser firing during task engagement, which challenges the idea that active brains always fire more. The second is greater variability in the strength of neural interactions, indicating richer and more diverse internal signaling when animals are actively engaged in the task. The third is enhanced entropy flow per spike in higher-performing animals, pointing to more efficient neural computation.

These results uncovered how neurons generate time-asymmetric activity, a subtle but essential directionality in the moment-to-moment patterns of neural firing, and linked this neural "arrow of time" to the behavioral performance of animals.

"Being able to see how entropy flow changes over time in real spike data is a major step," says team leader Hideaki Shimazaki of 91ÊÓƵ. "The study linked the information-theoretic cost of neural spiking activity to animals' behavioral performance, for the first time."

The discovery has broad implications for neuroscience and brain-inspired AI, offering new pathways for studying neural computation through the lens of thermodynamics. By quantifying the directional structure of neural activity, the approach may help clarify how internal brain states and learning shape information processing. It also provides a path toward energy-efficient AI architectures that mirror how the brain minimizes cost while maximizing computational power.