🌀 nonlinear sciences

Adaptive transitions in FitzHugh-Nagumo networks with Hebb-Oja coupling rules

This study demonstrates that FitzHugh-Nagumo networks with Hebb-Oja adaptive coupling undergo complex transitions between traveling waves, synchronized states, and chimera states, particularly when coupling dynamics are slower than nodal dynamics, ultimately revealing that asymptotic coupling strength follows an inverse power law with respect to the Oja forgetting parameter.

Original authors: Astero Provata, George C. Boulougouris, Johanne Hizanidis

Published 2026-02-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Astero Provata, George C. Boulougouris, Johanne Hizanidis

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a giant, bustling city made up of thousands of tiny, independent dancers (the neurons). In a traditional city plan, the connections between these dancers are fixed: if you hold hands with your neighbor, you hold hands forever, no matter what. But in the real brain, and in the world of artificial intelligence, connections are alive. They strengthen when people dance together and fade when they drift apart. This is the concept of adaptive coupling.

This paper explores what happens when we let these dancers change their hand-holding rules while they are dancing, using a specific set of rules called Hebb-Oja learning.

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

1. The Dancers and the Rules

The "dancers" in this study are mathematical models called FitzHugh-Nagumo oscillators. Think of them as simple neurons that have a "spike" (like a heartbeat or a thought) and a "recovery" period.

The Old Way: In the past, scientists studied these dancers assuming the hand-holding strength (the connection) never changed.
The New Way: This study asks: What if the hand-holding strength changes while they dance?
The Rule (Hebb-Oja): The rule is simple: "Dancers who move together, get closer." If two neurons fire at the same time, their connection gets stronger. However, there's a catch (the Oja rule): if the connection gets too strong, it starts to weaken slightly to prevent the whole system from exploding. It's like a thermostat that keeps the heat from getting too high.

2. The Speed of Change: The "Slow Motion" Camera

The most important discovery in this paper is about speed.

Fast Change: If the connections change as fast as the dancers move, the system settles down immediately. It's like a chaotic party that instantly turns into a quiet room. You don't see the interesting stuff in between.
Slow Change: If the connections change very slowly (much slower than the dancing), something magical happens. The system gets stuck in a "slow motion" phase where it tries out many different dance styles before finally settling down.

The authors found that when the connection rules change slowly, the network doesn't just go from "chaos" to "order." It goes through a journey of transitions.

3. The Journey of Transitions

Imagine the city of dancers going through different phases as the connections slowly evolve:

Phase 1: The Chaos. At the start, everyone is dancing randomly.
Phase 2: The Chimera. Suddenly, the city splits. One half of the city is dancing in perfect unison (like a military march), while the other half is still dancing wildly and randomly. This strange mix is called a Chimera State (named after the Greek monster with parts of a lion, goat, and snake).
Phase 3: Traveling Waves. Sometimes, the "wild" dancers form a wave that travels around the ring, like a ripple in a pond.
Phase 4: The Final State. Eventually, the system settles into a final pattern, usually a few small groups of wild dancers surrounded by a sea of synchronized ones.

The paper shows that these transitions happen abruptly. It's not a smooth slide; it's like a light switch flipping. One moment you have a 6-headed chimera (6 wild groups), and the next moment, it snaps into a 2-headed chimera.

4. The "Forgetting" Parameter

There is a knob in the system called the Oja parameter (α). You can think of this as the "forgetting rate."

If the dancers have a high forgetting rate (high α), they forget their connections quickly. The final connections end up being weak.
If they have a low forgetting rate (low α), they remember their connections well, and the final connections are strong.
The paper found a neat mathematical rule: The final strength of the connection is basically 1 divided by the forgetting rate. It's a perfect inverse relationship.

5. Why Does This Matter? (The Real World Connection)

Why should we care about mathematical dancers? Because this mimics how our brains learn and age.

Learning (Potentiation): When a child learns, their brain connections grow stronger. This is like starting with weak connections and letting them slowly build up. The "slow change" in the model shows that learning isn't a smooth line; it happens in abrupt jumps. You don't just get slightly better at math; you suddenly "get it."
Aging and Disease (Depression): When we age or get diseases like Alzheimer's, connections weaken and disappear. The model shows this as connections slowly deteriorating from strong to weak.
The "Slow" Lesson: The study suggests that for complex learning to happen, the brain (or an AI) needs time. If the connections change too fast, the system skips the interesting intermediate steps (the chimera states) and just settles. To see the full complexity of learning, the "learning speed" must be much slower than the "thinking speed."

Summary Analogy

Imagine a group of people trying to organize a flash mob.

Static Network: They are told who to stand next to, and they can't move. They either dance perfectly or not at all.
Adaptive Network (Fast): They can change partners instantly. They immediately find the best formation and lock in.
Adaptive Network (Slow - This Paper): They are allowed to change partners, but only once every hour. Because they change so slowly, they spend hours trying out weird formations (some people dancing together, some alone, waves moving around). They go through a chaotic, beautiful, and complex journey of "almost right" states before finally finding the perfect formation.

The Takeaway: Complexity and interesting behaviors (like the sudden "aha!" moments of learning) often happen when the rules of the game change slowly, allowing the system to explore many different possibilities before settling on a final answer.

1. Problem Statement

The study addresses the dynamics of complex networks where the connectivity (synaptic weights) is not static but evolves over time based on the system's activity. While traditional network science often assumes static topologies, biological systems (like the brain) and artificial neural networks rely on synaptic plasticity (learning).

The Gap: Previous studies on Chimera states (hybrid states where coherent and incoherent domains coexist) in FitzHugh-Nagumo (FHN) networks largely focused on static coupling. The influence of adaptive coupling (specifically Hebbian learning with Oja's normalization) on the formation, evolution, and transition between different synchronization regimes (traveling waves, synchronized states, and multi-chimera states) in FHN networks was not fully explored.
The Core Question: How does the timescale of coupling adaptation relative to the nodal dynamics affect the network's trajectory toward its asymptotic state? Does the network pass through intermediate synchronization regimes, and how do parameters like the Oja "forgetting" parameter ( $\alpha$ ) and the coupling timescale ( $\tau_\sigma$ ) influence these transitions?

2. Methodology

The authors employed numerical simulations on a ring network of $N$ interacting FitzHugh-Nagumo oscillators with periodic boundary conditions.

Mathematical Model

The system is governed by three coupled differential equations:

Membrane Potential ( $u_j$ ): $\epsilon \dot{u}_j = u_j - u_j^3/3 - v_j + \text{coupling terms}$
Recovery Potential ( $v_j$ ): $\dot{v}_j = u_j + \gamma + \text{coupling terms}$
Adaptive Coupling ( $\sigma_{jk}$ ): $\tau_\sigma \dot{\sigma}_{jk} = u_j u_k - \alpha u_j^2 \sigma_{jk}$

Hebbian Term ( $u_j u_k$ ): Represents "neurons that fire together, wire together," strengthening connections between active nodes.
Oja Forgetting Term ( $-\alpha u_j^2 \sigma_{jk}$ ): Prevents unbounded growth of weights, ensuring they converge to finite values.
Parameters:
- $\epsilon = 0.01$ : Separation of timescales between membrane and recovery potentials.
- $\gamma = 0.5$ : Sets the node dynamics to an oscillatory regime.
- $\sigma_c$ : Controls the sign of coupling (excitatory/inhibitory).
- $\alpha$ : Oja parameter (learning/forgetting balance).
- $\tau_\sigma$ : Timescale of coupling evolution (varied from fast to slow).

Quantitative Measures

To analyze the transitions, the authors utilized:

Kuramoto Order Parameter ( $z$ ): Measures global synchronization ( $z=0$ for asynchrony, $z=1$ for full synchrony). Intermediate values indicate chimera states.
Average Effective Coupling ( $\langle \sigma_{eff} \rangle$ ): The mean strength of active links.
Coupling Fluctuations ( $D_\sigma$ ): The spatial deviation of coupling weights from the mean, used to detect instability and transitions.
Spacetime Plots: Visualizing the membrane potential $u_j(t)$ over time and node index to identify coherent/incoherent domains.

3. Key Contributions

Discovery of Adaptive Transitions: The study demonstrates that FHN networks with Hebb-Oja coupling do not jump directly to a final state. Instead, they undergo a sequence of abrupt transitions between different synchronization regimes (e.g., from asynchrony $\to$ multi-chimera $\to$ traveling waves $\to$ final chimera) before reaching asymptotic stability.
Timescale Dependence: A critical finding is that these transitions are only observable when the coupling timescale ( $\tau_\sigma$ ) is significantly slower (orders of magnitude) than the nodal dynamics. If coupling evolves too fast, the system bypasses intermediate states and reaches the asymptotic state immediately.
Inverse Power Law Relationship: The authors derived that the asymptotic average coupling strength follows an inverse power law with respect to the Oja parameter: $\langle \sigma_{eff} \rangle_{asymp} \propto \alpha^{-1.073}$ . This confirms the theoretical mean-field prediction ( $\sigma \approx 1/\alpha$ ).
Bifurcation in Timescales: The paper identifies a critical threshold for $\tau_\sigma$ (around 15–250 in their units). Below this, coupling distributions remain heterogeneous (mixed positive/negative); above this, they converge to a single positive distribution, fundamentally changing the network's asymptotic behavior.

4. Key Results

A. Transitions with Increasing Coupling (Excitatory)

Starting from random initial conditions with negative/weak coupling, the network evolves through distinct phases:
1. Disorganized State: Low $z$ , random potentials.
2. Multi-Chimera State: Emergence of multiple coherent and incoherent domains (e.g., a 6-headed chimera).
3. Partial Synchronization: Incoherent domains shrink as coupling strengthens.
4. Final Asymptotic State: A stable chimera state (e.g., 2 traveling incoherent domains in a sea of coherence).
Observation: These transitions are marked by abrupt jumps in the average coupling strength and spikes in coupling fluctuations ( $D_\sigma$ ).

B. Transitions with Decreasing Coupling (Inhibitory/Decay)

Starting with high coupling strengths (simulating decay or pruning), the network transitions from full synchrony to chimera states as coupling weakens.
Result: Even in highly synchronized states, the adaptive mechanism can destabilize the system, leading to the emergence of chimera states.

C. Influence of Parameters

Oja Parameter ( $\alpha$ ): Higher $\alpha$ leads to lower asymptotic coupling strengths. The route to the asymptote (smooth vs. abrupt) depends on the initial conditions relative to the target $1/\alpha$ .
Coupling Timescale ( $\tau_\sigma$ ):
- Fast ( $\tau_\sigma \approx 1$ ): The system reaches the asymptotic state almost instantly; no intermediate transitions are observed.
- Slow ( $\tau_\sigma \geq 100$ ): The system exhibits rich dynamics, traversing multiple synchronization regimes.
- Distribution Shift: At low $\tau_\sigma$ , the final coupling distribution is heterogeneous (positive and negative weights). At high $\tau_\sigma$ , the distribution collapses to a single positive peak (delta-like), indicating a shift from mixed to purely excitatory effective coupling.

5. Significance and Implications

Biological Relevance: The results align with biological phenomena such as Long-Term Potentiation (LTP) and Long-Term Depression (LTD). The "slow" coupling regime ( $\tau_\sigma \gg 1$ ) mimics the biological timescale of synaptic plasticity (seconds to minutes) compared to neuronal firing (milliseconds). The findings suggest that learning and memory formation in neural networks may not be smooth processes but rather occur via abrupt, successive steps of reorganization.
Neurodegeneration Modeling: The "decreasing coupling" scenario models aging or neurodegenerative diseases (e.g., Alzheimer's), where synaptic strength deteriorates, potentially leading to a loss of coherence and the emergence of pathological synchronization patterns.
Generalizability: The authors note that similar adaptive transitions were found in Leaky Integrate-and-Fire models, suggesting this behavior might be generic to neuronal networks. Conversely, they note that Kuramoto models do not show these multiplicity transitions under Hebb-Oja rules, highlighting the importance of the specific oscillator dynamics (FHN) in driving these complex transitions.
Artificial Intelligence: The study provides insights for designing adaptive neural networks in AI, suggesting that the timescale of weight updates is a critical hyperparameter for controlling the complexity and stability of learning processes.

In conclusion, the paper establishes that adaptive coupling acts as a dynamic driver that forces FHN networks through a landscape of synchronization states. The observation of these transitions is strictly contingent on the separation of timescales between the network's structural evolution and its nodal dynamics, offering a new perspective on how complex synchronization patterns emerge in plastic systems.