Dynamic shifts in brain criticality support cognitive processing

This study demonstrates that the hippocampus dynamically regulates its proximity to criticality—maintaining a critical state for flexible learning and coordination, while shifting to a subcritical, ordered state during sleep for memory replay—thereby providing a biophysical basis for optimizing both biological and artificial cognitive systems.

The Gist

Imagine your brain is a bustling city. For this city to function perfectly, it needs to be in a very specific "Goldilocks zone"—not too chaotic, not too rigid, but just right. In physics and computer science, this sweet spot is called criticality.

Think of criticality like a tightrope walker. If the walker is too stiff (too ordered), they can't adjust to a sudden gust of wind. If they are too loose (too chaotic), they fall off immediately. But right in the middle, they can react instantly, adapt to changes, and process information with maximum efficiency.

This paper, by researchers at Cornell University, explores how the brain's "memory center" (the hippocampus) constantly shifts between being a flexible tightrope walker (for learning) and a rigid, organized library (for storing memories), and how it switches between these two modes.

Here is the story of their discovery, broken down into simple concepts:

1. The Two Modes of the Brain

The researchers found that the brain doesn't stay in one mode. It dynamically shifts depending on what it's doing:

• Mode A: The Flexible Tightrope (Learning & Waking)
  When you are awake and trying to learn something new (like a new route to work or a new language), your brain operates close to criticality.

  • The Analogy: Imagine a jazz band improvising. Everyone is listening to each other, ready to jump in with a new idea. The music is fluid, adaptable, and can go in any direction.
  • Why it helps: This state gives the brain a huge "dynamic range." It can hear very quiet signals and shout loud ones, allowing it to pick up on new information quickly and coordinate with other brain regions to build new memories.

• Mode B: The Rigid Library (Sleep & Memory Storage)
  When you fall asleep, especially during deep sleep, the brain shifts away from criticality into a more "ordered" or "subcritical" state.

  • The Analogy: Imagine the jazz band stops improvising and starts playing a strict, rehearsed symphony. The music is predictable, repetitive, and highly organized.
  • Why it helps: This rigidity is actually good for sleep! It allows the brain to play back the day's events (called "memory replay") without getting distracted by new noise. It's like a librarian carefully filing away books so they don't get lost or mixed up.

2. The "Switch" Mechanism: How the Brain Changes Gears

The most exciting part of the study is how the brain switches from the "Jazz Band" (learning) to the "Symphony" (sleeping) and back again.

• The Problem: During sleep, the brain gets very "ordered" (subcritical) to replay memories. But if it stays too ordered, it can't learn anything new the next day. It needs a way to reset.
• The Solution: The researchers discovered a specific type of brain cell (called CCK interneurons) that acts like a reset button.
  • The Analogy: Imagine the "Symphony" (sleep replay) is getting too quiet and rigid. Suddenly, a specific group of conductors (the CCK cells) starts waving their batons vigorously. This creates a burst of activity (called BARRs) that shakes the system up, breaking the rigid order and returning the brain to the "Goldilocks zone" of criticality.
  • The Result: This ensures that when you wake up, your brain is once again flexible and ready to learn, rather than stuck in a rigid loop.

3. What This Means for You (and AI)

The study confirms that learning requires flexibility, while storing memories requires order. The brain is a master of switching between these two states.

• For Humans: It explains why sleep is so crucial. You can't just "learn" while you sleep; you need to first learn while awake (flexible mode), then switch to the rigid mode to file those memories away, and finally use the "reset button" to wake up ready for the next day.
• For Artificial Intelligence (AI): The authors suggest that our current AI models (like Large Language Models) might be missing this trick. They are often trained to be either too rigid or too chaotic. By designing AI that can dynamically shift between "flexible learning" and "rigid storage" states—just like our brains—we could create smarter, more efficient machines that learn faster and remember better.

Summary

Think of your brain as a smart thermostat:

1. Daytime (Learning): It keeps the temperature just right (critical) so you can react to everything.
2. Nighttime (Sleep): It turns the heat down (subcritical) to organize the house and clean up the mess.
3. The Reset: A special mechanism (CCK cells) kicks in to warm things back up so you aren't frozen when you wake up.

This dynamic shifting is the secret sauce that allows us to learn, remember, and stay sane!

Technical Summary

1. Problem Statement

The "criticality hypothesis" posits that neural networks operate near a critical point (a phase transition between order and disorder) to maximize computational advantages such as dynamic range, information capacity, and flexibility. While signatures of criticality (e.g., power-law distributions) have been observed in the brain, two fundamental questions remain unresolved:

1. Biophysical Mechanism: What specific circuit mechanisms regulate the brain's proximity to criticality?
2. Functional Specificity: How does the brain dynamically shift between critical and subcritical regimes to support distinct cognitive processes (e.g., learning vs. memory consolidation), given that these processes have opposing computational demands?

2. Methodology

The authors employed a multi-modal approach combining large-scale electrophysiology, computational modeling, and machine learning analysis in behaving rodents (mice and rats).

• Data Acquisition:
  • Subjects: Rodents performing spatial memory tasks (Cheeseboard maze) and undergoing sleep.
  • Recording: High-density silicon probe recordings from the hippocampal CA1 region, CA2, CA3, and medial entorhinal cortex (MEC).
  • Manipulations: Optogenetic silencing of CA1 Cholecystokinin-expressing (CCK) interneurons during sleep; pharmacological induction of seizures (for control comparison).
• Quantification of Criticality:
  • Time-Resolved Renormalization Group (tRG): Instead of traditional avalanche-based methods (which require long time windows), the authors used autoregressive (AR) models fitted to multiunit spike counts. This allowed for a time-resolved estimate of the distance to criticality ($d^2$), defined as the Euclidean distance from the fitted AR model to the $\beta=2$ fixed-point hyperplane (the critical manifold).
  • Complementary Metrics: Branching Ratio (BR), avalanche size/duration, and inter-spike-interval (ISI) entropy.
• Computational Modeling:
  • A biophysically realistic CA1 network model incorporating Pyramidal cells (E), Parvalbumin (PV) interneurons, and CCK interneurons.
  • Inputs from CA3 (driving SWRs) and CA2 (driving CCK activity).
• Neural Representation Analysis:
  • CEBRA (Consistent Embedding for Bipartite Relationship learning with Auxiliary variables): A self-supervised machine learning method used to extract low-dimensional latent manifolds of neural activity to assess representational flexibility.
  • Canonical Correlation Analysis (CCA): To measure coordination between CA1 and input regions (CA2, CA3, MEC).

3. Key Contributions

1. Dynamic Regulation of Criticality: Demonstrated that the hippocampus does not maintain a static critical state but dynamically shifts between a near-critical regime (during waking/learning) and a subcritical regime (during sleep/memory replay).
2. Circuit Mechanism Identification: Identified CCK-mediated inhibition (specifically via BARRs—barrages of action potentials driven by CA2) as the homeostatic mechanism that restores the network to criticality after the ordered state of memory replay.
3. Functional Decoupling: Showed that proximity to criticality facilitates learning (via flexibility and coordination) while a subcritical state facilitates memory consolidation (via ordered replay).
4. AI Implications: Proposed that optimal learning systems (biological or artificial) require dynamic regulation between flexible (critical) and rigid (subcritical) states, offering constraints for Large Language Model (LLM) design.

4. Key Results

A. State-Dependent Criticality Fluctuations

• Waking vs. Sleep: The hippocampus operates closer to criticality ($low\ d^2$) during waking (especially active theta) and REM sleep compared to non-theta waking and NREM sleep.
• Dynamic Range: Proximity to criticality correlates with a higher dynamic range of firing rates across all states.

B. Memory Consolidation: Shift to Subcriticality

• Replay and Criticality: During post-task NREM sleep, the distance to criticality ($d^2$) increases (moving away from criticality) as memory replay events occur.
• Correlation: Replay probability is positively correlated with $d^2$ (further from criticality).
• Subcritical State: During Sharp-Wave Ripples (SWRs) associated with replay, the Branching Ratio (BR) drops significantly below 1 (subcritical), and firing entropy decreases.
• Causality: This shift is not merely a byproduct of replay activity; the network enters a global subcritical regime that permits replay.

C. Mechanism of Restoration: CCK Interneurons

• Model Prediction: The CA1 model predicted that increased CA2 input to CCK interneurons would reduce $d^2$ (restore criticality).
• BARRs: During sleep, CA2 drives "BARRs" (barrages) in CA1 CCK interneurons.
• Experimental Validation:
  • BARR rate is inversely correlated with $d^2$ (higher BARRs $\rightarrow$ closer to criticality).
  • Optogenetic silencing of CA1 CCK interneurons during BARRs prevents the return to criticality, leaving the network in a subcritical state.
  • BARRs increase BR and entropy, effectively "resetting" the network.

D. Learning: Proximity to Criticality Facilitates Performance

• Behavioral Correlation: In a spatial learning task, behavioral performance (path length) is positively correlated with proximity to criticality ($low\ d^2$) during learning trials.
• Specificity: This correlation disappears once the task is learned (recall phase), suggesting criticality is specific to the acquisition of new information.
• Mechanisms of Benefit:
  1. Coordination: Near criticality, CA1 shows stronger canonical correlation (coordination) with input regions (MEC, CA3, CA2).
  2. Flexibility: Using CEBRA, the authors showed that distinct neural manifolds (representations) for different reward configurations emerge more clearly when the network is near criticality. This "representational flexibility" allows the integration of new memories without overwriting old ones.

5. Significance and Implications

• Biophysical Substrate: The study moves beyond phenomenological observations of criticality to identify specific cell-type-specific mechanisms (CA2 $\rightarrow$ CCK interneurons) that regulate the brain's computational regime.
• Dual-Regime Hypothesis: It challenges the view that the brain should always be at criticality. Instead, it proposes a dynamic oscillation:
  • Near-Critical: Optimized for plasticity, flexibility, and input integration (Learning).
  • Subcritical: Optimized for stability, order, and precise sequence replay (Consolidation).
• Artificial Intelligence: The findings suggest that current AI models (like LLMs) might benefit from incorporating dynamic criticality regulation. Fixed-point training or inference might be suboptimal; instead, systems could toggle between "exploration" (critical/flexible) and "exploitation" (subcritical/rigid) phases to solve the problem of catastrophic forgetting and information scarcity.
• Clinical Relevance: The paper notes that while the healthy brain oscillates near criticality, epileptic seizures represent a pathological supercritical state, highlighting the importance of maintaining the balance between order and disorder.

In summary, this work provides a comprehensive framework linking the physics of criticality to specific neural circuits and cognitive functions, revealing that the brain's ability to learn and remember relies on its capacity to dynamically shift its operating point.