🧬 biology

Soliton-like Waves in a Two-Dimensional Recurrent Spiking Neural Network with Weighted Spike-Timing-Dependent Plasticity

This paper demonstrates that a biologically plausible, two-dimensional recurrent spiking neural network incorporating weighted spike-timing-dependent plasticity and divisive normalization spontaneously generates stable, self-propagating soliton-like waves that learn directional propagation and enable spatial memory through local plasticity rules.

Original authors: Ch. Meessen

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ch. Meessen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine a vast, flat city made of 40,000 tiny, binary light switches (neurons). Half of them are "Excitatory" (they want to turn the lights on), and half are "Inhibitory" (they want to turn the lights off). These switches are arranged in a grid, and they are all connected to their neighbors by invisible wires.

This paper describes a computer simulation of this city to see what happens when we poke it in one spot. The result? The city spontaneously learns to send out perfect, self-sustaining ripples of light that behave like solitons—a special kind of wave that keeps its shape and speed without fading away.

Here is how the magic happens, broken down into simple concepts:

1. The Rules of the Game

The city operates on a few strict rules:

The "Mexican Hat" Shape: The Excitatory switches can shout to neighbors far away (a radius of 9 blocks), but the Inhibitory switches can only shout to neighbors nearby (a radius of 5 blocks). This creates a natural "funnel" where a wave can spread out, but the nearby "stop signs" (inhibition) keep it from exploding everywhere.
The Learning Rule (WSTDP): This is the most important part. The connections (wires) between switches change strength based on timing. If Switch A fires just before Switch B, the wire between them gets stronger. If Switch B fires before Switch A, the wire gets weaker. It's like a "cause-and-effect" memory: the network learns which direction the wave is moving and reinforces that path.
The "Volume Knob" (Normalization): To prevent the city from getting too loud, the total input a switch receives is normalized. Think of it as a volume knob that automatically adjusts so that even if everyone shouts at once, the signal stays within a manageable range. The paper suggests this could happen naturally in a brain's dendrites (the branches of a neuron) using only local information.

2. The Experiment: Poking the City

The researchers started with a "blank" city where all wires had the same strength. Then, they repeatedly tapped a small circle in the center, forcing those switches to flash.

What happened?
At first, the light just spread out in a messy circle. But very quickly, the network "learned." The wires in the direction the wave was moving got stronger, and the wires in the opposite direction got weaker.

The Result: A perfect, ring-shaped wave of light emerged. It moved at a constant speed, kept its shape, and kept going even after the researchers stopped tapping the center. It was a self-propagating wave packet.

3. The "Soliton" Behavior

These waves act like dissipative solitons. In simple terms:

They are stable: They don't wobble or fade.
They are self-sustaining: Once started, they carry their own "fuel" (the strengthened wires) to keep moving.
They annihilate on collision: This is the coolest part. If you start two waves from opposite sides, they don't pass through each other like ghosts. When they meet, they crash and disappear. The "inhibitory trail" (the aftermath of the wave) acts like a wall that stops the other wave from entering.

4. The "Semi-Permanent" Boundary

When two waves crash and die, they leave behind a boundary.

If you stop one of the light sources, the remaining wave tries to cross the middle, but it hits this invisible wall and stops.
Why? The network has "memorized" the boundary. The wires pointing toward the crash site are weak, and the wires pointing away are strong. The wave literally cannot cross because the path is broken.
Encoding Information: If the two light sources flash at slightly different times or speeds, the crash point moves. The network effectively "remembers" the timing difference by placing the boundary in a new spot. It's a form of spatial memory written into the strength of the wires.

5. What Breaks the Magic?

The paper found that this delicate balance requires specific conditions:

Too much Excitation (Epilepsy-like): If the Inhibitory switches aren't strong enough to counter the Excitatory ones, the whole city lights up at once in a chaotic, synchronized flash. No waves form; just a global seizure-like state.
Wrong Geometry: If the Inhibitory switches can shout as far as the Excitatory ones, the waves fail to form. The "funnel" shape is essential.
Frequency Wars: If one light source flashes much faster than the other, its wave eventually overwrites the other's territory, pushing the boundary all the way to the other side.

The Big Picture

The paper concludes that you don't need a complex, pre-wired brain to create these traveling waves. You just need:

A simple grid of neurons.
A slight difference in how far excitatory and inhibitory signals travel.
A simple rule that strengthens connections based on timing.

From these simple local rules, complex, stable, traveling patterns emerge on their own. The network essentially "learns" to sustain a wave and creates permanent boundaries where waves collide, offering a minimal model for how the brain might organize activity and remember spatial relationships.

Technical Summary: Soliton-like Waves in a Two-Dimensional Recurrent Spiking Neural Network with Weighted Spike-Timing-Dependent Plasticity

Problem Statement
Traveling waves of neural activity are ubiquitous in the cortex, yet the mechanisms by which they emerge and sustain themselves in recurrent circuits remain incompletely understood. While neural field models have predicted traveling waves and solitary bumps based on "Mexican-hat" connectivity profiles (short-range excitation balanced by long-range inhibition), these models typically abstract away individual spikes and synaptic plasticity. Conversely, while Spike-Timing-Dependent Plasticity (STDP) is known to shape receptive fields and strengthen pathways, it is unclear whether a recurrent excitatory-inhibitory (E/I) network equipped with STDP can autonomously generate stable, self-propagating wave packets (soliton-like structures) from a uniform state, and whether these structures exhibit specific collision dynamics such as annihilation.

Methodology
The authors constructed a minimal, biologically plausible computational model of a two-dimensional recurrent network consisting of 200 × 100 pairs of excitatory (E) and inhibitory (I) neurons arranged on a discrete grid. The model operates in discrete time steps with synchronous updates.

Neuron Model: Each neuron integrates weighted spikes from presynaptic neighbors. A key feature is divisive normalization of synaptic input, where the total excitatory drive is normalized by the sum of excitatory synaptic weights ( $S_E$ ). This normalization is proposed to have a biologically plausible dendritic implementation using local attenuation at binary junctions. The firing threshold ( $\theta$ ) undergoes homeostatic adaptation: it decreases toward a minimum ( $\theta_{min}$ ) during silence to prevent "stall states" and increases toward a maximum ( $\theta_{max}$ ) upon spiking to implement spike-frequency adaptation.
Connectivity: The network features a geometric asymmetry: excitatory-to-excitatory ( $E \to E$ ) connections have a larger radius ( $r_{EE}=9$ ) than excitatory-to-inhibitory ( $E \to I$ ) and inhibitory-to-excitatory ( $I \to E$ ) connections ( $r_{EI}=r_{IE}=5$ ). No $I \to I$ connections exist. Initial weights are uniform, with inhibitory weights set significantly higher than excitatory ones to compensate for the smaller number of inhibitory inputs.
Learning Rule: Synaptic weights evolve via Weighted STDP (WSTDP), a multiplicative, temporally asymmetric learning rule. The rule ensures weights remain bounded without hard clipping, converging to a fixed point determined by the balance of Long-Term Potentiation (LTP) and Long-Term Depression (LTD) event rates.
Stimulation: The network is subjected to repeated, localized stimulation (forcing a disk of excitatory neurons to fire) at a fixed period. Experiments include single-source stimulation and dual-source stimulation with varying phases and frequencies to study wave interactions.

Key Contributions and Results

Emergence of Dissipative Solitons: Starting from spatially uniform connectivity, repeated localized stimulation induces the spontaneous formation of stable, self-propagating wave packets. These packets maintain a constant spatial profile (a bright excitatory ring followed by an inhibitory trail) and propagate at a constant speed. Crucially, once established, these waves continue to propagate autonomously even after the external stimulation is removed, demonstrating they are self-sustaining structures.
Role of Geometric Asymmetry and WSTDP: The emergence of these waves requires two conditions:
- Geometric Asymmetry: The excitatory connection radius must exceed the inhibitory radius to create a functional Mexican-hat profile.
- WSTDP: The learning rule "engraves" the direction of propagation into the synaptic weight landscape. Synapses in the direction of propagation undergo net LTP, while those in the reverse direction undergo net LTD. This creates a directional bias that prevents backward propagation and stabilizes the wave.
Collision Dynamics and Annihilation: When two wave packets collide frontally, they annihilate rather than passing through each other. This behavior is characteristic of dissipative solitons. The annihilation is caused by the inhibitory trail of one wave suppressing the excitatory front of the other.
Semi-Persistent Spatial Boundaries: In dual-source experiments, the collision point forms a "semi-persistent" boundary encoded in the synaptic weights.
- Phase Encoding: If sources are in phase, the boundary forms at the geometric midpoint. If there is a phase offset, the boundary shifts toward the lagging source.
- Frequency Competition: If sources have different frequencies, the faster source progressively overwrites the synaptic weights of the slower source, causing the boundary to drift. A sufficiently large frequency advantage leads to complete territorial takeover.
- Robustness: These boundaries persist even after one source is silenced, though their stability is modulated by the adaptive threshold parameters.
Necessity of Strong Initial Inhibition: The model demonstrates that strong initial inhibitory weights are a geometric necessity. Reducing these weights leads to global synchronous oscillations (an epileptic-like regime) rather than localized solitons, highlighting the critical role of inhibition in confining the excitatory wave front.

Significance and Claims
The paper claims to provide a minimal computational framework for studying the emergence of cortical traveling waves, activity zone delimitation, and spatial memory derived solely from local plasticity rules.

Mechanism of Emergence: The study suggests that complex spatiotemporal dynamics, such as stable traveling waves and collision annihilation, can emerge from simple local rules (WSTDP and homeostatic threshold adaptation) without pre-wired spatial structures, provided there is a geometric asymmetry in connectivity.
Spatial Memory: The authors propose that the synaptic weight landscape in such networks implements a form of spatial memory. The network "remembers" the relative timing (phase) and rate (frequency) of past inputs through the position and stability of the annihilation boundaries.
Directional Stability: The WSTDP mechanism not only enables propagation but actively stabilizes it by suppressing unstable reverse modes, preventing runaway oscillations.
Biological Plausibility: The model offers a potential explanation for how cortical circuits might segregate and maintain distinct activity domains and how traveling waves might be modulated by competing inputs, all within a framework of local, spike-based plasticity.

The authors note that while the model is minimal (discrete time, binary spikes), the robustness of the soliton emergence suggests these phenomena are fundamental properties of recurrent E/I networks with local plasticity, rather than artifacts of specific parameter choices. Future work is suggested to explore asynchronous propagation and structured synaptic afferents.