Parameter and hidden-state inference in mean-field models from partial observations of finite-size neural networks

This paper proposes a methodology to infer unknown parameters and reconstruct hidden macroscopic dynamics of finite-size neural networks by synchronizing a known mean-field model with a single scalar observable using a differential evolution algorithm.

The Gist

Imagine you are trying to figure out how a massive, complex orchestra is playing, but there is a catch: you are standing in a soundproof room, and the only thing you can hear is a single, solitary flute playing a melody.

You want to know two things:

1. The "Sheet Music" (Parameters): What are the rules of the song? How fast is the tempo? How loud is the percussion?
2. The "Hidden Musicians" (Hidden States): Even though you can’t hear the drums or the cellos, can you figure out exactly what they are doing just by listening to how the flute reacts to them?

This is the problem the researchers in this paper are solving for the brain.

The Problem: The "Too Many Neurons" Headache

The human brain has billions of neurons. If you tried to write a mathematical equation for every single neuron and every single connection, your computer would basically explode. It’s too much data.

To solve this, scientists use "Mean-Field Models." Instead of tracking every single musician, they try to track the "average" behavior—like the overall volume or the average pitch of the whole orchestra. This makes the math much simpler.

The catch? In a real experiment, we can usually only measure one tiny thing (like the average electrical voltage in a small area), while all the other important "average" behaviors remain invisible. Plus, real biological networks are "noisy" and messy, unlike perfect mathematical models.

The Solution: The "Master-Slave" Dance

The researchers developed a clever way to bridge the gap between the messy, real-world data and their clean mathematical models. They used two main tricks:

1. The Synchronization Trick (The "Follow the Leader" Method)
If you try to guess the "sheet music" by just starting a model and hoping it matches the flute, you’ll fail because you don't know how the orchestra started playing (the "initial conditions").

To fix this, they used Synchronization. Imagine the mathematical model is a dancer. Instead of letting the dancer move freely, you grab their hand and gently pull them to follow the rhythm of the flute you are hearing.

• The Noninvasive Method: You gently nudge the model to stay in sync with the data.
• The Invasive Method: You play a steady, rhythmic beat (like a metronome) that forces both the real network and the model to march to the same drum.

Once the model is "locked in" to the rhythm of the real data, it "forgets" its own wrong starting point and starts behaving like the real system.

2. The Evolution Trick (The "Survival of the Fittest" Method)
To find the exact "sheet music" (the parameters), they used an algorithm called Differential Evolution.

Think of this like a digital version of natural selection. The computer creates a "population" of different possible sheet musics. Some are too fast, some are too slow, some are too loud. The computer plays them, compares them to the flute you're hearing, and "kills off" the ones that sound wrong. The "survivors" are combined and mutated to create a new generation of even better sheet music. After many generations, you are left with the perfect set of rules.

The Results: It Actually Works!

The researchers tested this on two types of "orchestras":

• The Steady Orchestra (Periodic): One that plays predictable, repeating patterns.
• The Chaotic Orchestra (Chaotic): One that plays wild, unpredictable, and complex patterns.

The verdict? Even with the "chaotic" orchestra, their method was incredibly accurate. Once the network reached a certain size (about 1,000 neurons), they could figure out the "sheet music" with 99% accuracy.

Even more impressively, once they found the right rules, they could "see" the invisible musicians. They successfully reconstructed the behavior of the hidden parts of the network (like the firing rate) just by looking at the one thing they could measure (the voltage).

Why does this matter?

In the future, this could help doctors and scientists look at limited data from a brain scan and work backward to understand the deep, underlying "rules" of how a person's neural circuits are functioning. It’s a way of turning a tiny window of observation into a wide-angle view of the entire system.

Technical Summary

Technical Summary: Parameter and Hidden-State Inference in Mean-Field Models

1. Problem Statement

The paper addresses a fundamental challenge in computational neuroscience: how to bridge the gap between microscopic neural dynamics (individual neurons and synapses) and macroscopic observations (collective activity like mean firing rates).

While Next-Generation Neural Population Models (NGNPM) provide exact low-dimensional descriptions of large neural networks in the thermodynamic limit ($N \to \infty$), applying them to real-world scenarios is difficult because:

• Partial Observability: In experimental settings, only a single scalar macroscopic variable (e.g., mean membrane potential) is typically measurable, while other governing variables remain "hidden."
• Parameter Uncertainty: The specific parameters governing the mean-field equations are often unknown.
• Initial Condition Sensitivity: In chaotic regimes, the loss function used for parameter estimation is extremely sensitive to the unknown initial conditions of the hidden variables.
• Finite-Size Effects: Real networks are finite, meaning their dynamics include "noise" (fluctuations) that deviates from the deterministic mean-field limit, complicating trajectory matching.

2. Methodology

The authors propose a framework to infer unknown mean-field parameters and reconstruct hidden macroscopic variables using time-series data from a single observable.

A. Optimization Algorithm:
The study employs Differential Evolution (DE), a derivative-free, population-based global optimization algorithm. DE was chosen for its robustness in high-dimensional, nonlinear, and non-differentiable parameter spaces, as well as its ability to avoid local minima.

B. Synchronization-Based Strategies:
To decouple the loss function from the unknown initial conditions of hidden variables, the authors implement two synchronization techniques. These methods force the mean-field model to "forget" its initial state and follow the observed trajectory:

• Noninvasive Method: Uses a master-slave control scheme. A feedback term is added to the mean-field equations that drives the observed component ($X_m$) toward the measured signal ($X_{out}$), without perturbing the original network.
• Invasive Method: Applies a common periodic external current to both the microscopic network and the mean-field model. This drives both systems into a synchronized state (often within "Arnold tongues"), making the system less sensitive to initial conditions.

C. Test Models:
The methodology is validated on two classes of Quadratic Integrate-and-Fire (QIF) networks:

1. QIF-IN: An inhibitory network exhibiting periodic collective oscillations.
2. QIF-AD: An excitatory network with spike-frequency adaptation exhibiting chaotic collective dynamics.

3. Key Contributions

• Joint Inference Framework: A unified approach to simultaneously estimate model parameters and reconstruct the time series of unobserved (hidden) macroscopic variables.
• Synchronization for Optimization: The introduction of noninvasive and invasive synchronization to solve the "initial condition problem" in nonlinear dynamical system identification.
• Robustness to Chaos: Demonstrating that synchronization-based optimization can effectively handle chaotic trajectories, which are traditionally difficult to fit due to the "butterfly effect."

4. Results

• Parameter Accuracy: For both periodic (QIF-IN) and chaotic (QIF-AD) networks, the relative error in parameter estimation falls below 1% once the network size $N$ exceeds 1000 neurons.
• Hidden State Reconstruction: The model successfully reconstructed the dynamics of unobserved variables (e.g., mean firing rate $R$ and synaptic activation $S$) with high fidelity. Even in the chaotic QIF-AD model, the reconstructed trajectories closely matched the microscopic ground truth.
• Impact of Network Size: The authors quantified the relationship between $N$ and accuracy, showing that as $N$ increases, finite-size fluctuations decrease, leading to a lower minimum in the loss function and higher precision.
• Method Comparison: The invasive method generally provided a slightly narrower error distribution and higher robustness compared to the noninvasive method.

5. Significance

This work provides a mathematically grounded bridge between microscopic neural models and macroscopic population models. By proving that parameters and hidden states can be recovered from a single scalar observation, the authors offer a blueprint for using data-driven techniques (like SINDy or deep learning) to discover the governing equations of neural populations. This has significant implications for understanding neurological dysfunctions where macroscopic signals are the only available clinical data.