🧬 biology

How Recurrent Neural Information Flow Can Generate a Setpoint

This paper demonstrates that in recurrent neural networks, a unique and stable activity setpoint emerges not from plasticity or external targets, but from the intrinsic flow of information through excitatory-inhibitory loops, where stability is strictly governed by feedback gain, delay, and a channel capacity that must exceed the rate at which network instability generates information about its own state.

Original authors: luciano fernandes

Published 2026-10-09

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: luciano fernandes

Original paper licensed under CC BY 4.0 (https://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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

The brain is a vast network of cells that fire electrical signals to think, feel, and move. For this system to work, the overall level of activity must stay within a specific range. If the network fires too little, it goes quiet and stops working; if it fires too much, it can spiral into chaotic, uncontrolled bursts. Biologists have long known that the brain has mechanisms to keep this activity in check, often by adjusting the strength of connections between cells over time. However, a precise question has remained unanswered: where does the target level come from in the first place? Does the brain need a separate, dedicated control center that constantly measures activity and sends a "target" signal to tell the neurons how fast to fire? Or can the network itself, through the simple back-and-forth flow of signals between its parts, naturally settle on a stable operating level without any external instruction?

A new study by Luciano Fernandes at Washington State University explores this question using a detailed mathematical model of a neural network. The researchers built a virtual brain circuit consisting of two main groups of neurons: excitatory cells that encourage firing and inhibitory cells that suppress it. In their model, these groups are connected in a loop where the excitatory cells drive the inhibitory ones, which in turn send a signal back to calm the excitatory cells. Crucially, this model contains no external "setpoint" variable, no separate sensor to measure the rate, and no plasticity that changes the connections over time. The only feedback comes from the inhibitory cells themselves. The study asks whether this simple loop, driven only by the flow of information from the inhibitory cells back to the excitatory ones, is enough to create and hold a stable firing rate.

The researchers found that the answer is yes, but only under very specific conditions. The value of the stable firing rate is determined entirely by the balance of activity flowing around the loop. It is a fixed point where the excitation and inhibition perfectly cancel each other out, creating a unique, steady rhythm. However, for this level to be stable—meaning the network returns to it after a small disturbance—the inhibitory signal must carry information about the current state of the network back to the excitatory cells with the right strength and speed. If the feedback is too weak, the network drifts away from the level; if it is too strong, it begins to oscillate wildly. The study proves mathematically that there is a precise window of feedback strength where the level remains stable, and a critical limit on how much delay is allowed in the signal's travel time.

A central discovery of the paper is that maintaining this stability is fundamentally an information problem. The network is inherently unstable on its own; without the inhibitory feedback, any tiny difference in the starting state would grow exponentially, causing the network to run away to a different state. To prevent this, the inhibitory pathway must transmit information about the network's current state back to the excitatory cells. The study calculates the exact minimum rate at which this information must flow to keep the system stable. For the specific network they analyzed, this minimum rate is approximately 28.9 bits per second. This means the feedback channel must be capable of carrying at least this much information every second. If the channel is slower than this, or if the signal does not depend on the current state (for example, if it is a constant, unchanging signal), the network loses its stable level and collapses into chaos or a different, low-activity state.

To test these theoretical limits, the researchers ran computer simulations of the network, including a large-scale version with 10,000 individual neurons firing like real brain cells. They replaced the natural, continuous inhibitory signal with digital codes that transmitted information at different speeds. When they used an efficient coding method that packed information tightly, the network held its stable level as long as the transmission rate was just slightly above the calculated minimum of 28.9 bits per second. When the rate dropped even slightly below this threshold, the network failed to maintain the level, regardless of how the code was structured. Conversely, when they used inefficient coding methods, the network required a much higher transmission rate to stay stable. This confirmed that the limit is not just a theoretical curiosity but a hard constraint on how much information the feedback loop must carry.

The study also examined how the network handles changes in input. While the network can resist small, temporary jolts and return to its stable level, it cannot ignore a sustained, constant change in input. If the researchers applied a continuous extra current to the excitatory cells, the stable level shifted to a new value. This reveals that the level generated by this information flow is a "settling point" rather than a rigid setpoint. It is a stable state that the network naturally returns to after disturbances, but it moves if the underlying conditions change permanently. This distinction is important because it shows that the mechanism described here explains how a network can hold a level against noise, but it does not explain how the brain maintains a fixed level against long-term changes, which likely requires additional biological processes not included in this model.

The findings suggest that the brain's ability to maintain a steady firing rate might not rely on a separate, dedicated controller. Instead, the stability emerges from the very act of the inhibitory neurons sending information back to the excitatory ones. The network holds its level because the inhibitory feedback carries enough information about the current state to counteract the network's natural tendency to drift. This information flow acts as the glue that keeps the system together. The research provides a precise, testable prediction: in any biological circuit that relies on this type of inhibition-stabilized mechanism, the pathway carrying the inhibitory signal must have a capacity to transmit information that exceeds a specific minimum rate. If the signal is too slow, too noisy, or too delayed, the circuit will lose its stability. This offers a new way to think about neural regulation, framing the maintenance of brain activity not as a matter of matching a target, but as a continuous, information-theoretic balancing act where the flow of data itself creates the stability.

Technical Summary: How Recurrent Neural Information Flow Can Generate a Setpoint

Problem Statement
The paper addresses a fundamental question in neural systems: what determines the specific firing rate level (the "setpoint") to which neural activity returns after perturbation? While existing models explain how activity is maintained once a reference is specified (often via explicit target-rate variables or plasticity rules), they do not explain the origin of the reference itself. The author investigates whether an attracting operating level can emerge solely from the flow of activity and information through a recurrent network with fixed connections, without an explicit target-rate variable, a separately added sensing/eﬀector circuit, or synaptic plasticity.

Methodology
The study employs a rigorous combination of exact mathematical proofs and finite-time simulations:

Exact Mean-Field Model: The author utilizes the exact population reduction of heterogeneous quadratic integrate-and-fire (QIF) neurons (Montbrió et al., 2015). This reduces the network dynamics to a closed system of six ordinary differential equations describing the firing rates and voltage centers of excitatory (E) and inhibitory (I) populations.
Two Network Configurations:
- Exact Construction: A specific parameter set chosen for algebraic tractability to prove exact bounds on stability and information rates.
- Reference Network: A biologically plausible parameter set used for simulations, featuring an excitatory population of 8,000 and an inhibitory population of 2,000 neurons.
- Random Networks: 68 randomly generated inhibition-stabilized networks (ISNs) to test the generality of the findings.
Control and Information Theory Tools: The analysis applies the Data-Rate Theorem (establishing minimum channel capacity for stabilizing unstable systems), Bode's Sensitivity Integral (relating feedback gain to disturbance attenuation), and Shannon information theory.
Interventions: The inhibitory signal received by excitatory cells is manipulated via:
- Feedback Gain and Delay: Scaling the state-dependent part of the signal and introducing time lags.
- Finite-Rate Coding: Replacing the analog inhibitory signal with efficient "zooming" coders or static quantizers (e.g., one-bit codes) to test capacity limits.
- Noise Injection: Adding white Gaussian noise to the inhibitory pathway to measure information transmission rates.
- State-Independent Signals: Replacing the dynamic signal with constant or replayed signals to test the necessity of state dependence.

Key Contributions and Results

Existence of a Unique Level: The author proves that for networks with fixed connections, the balance of activity flowing around the excitatory-inhibitory loop fixes a unique, positive equilibrium firing rate. This level is determined by the network's structural parameters (coupling strengths, excitability distributions) rather than an external reference.
Conditions for Attraction (Stability):
- Gain Window: The level is locally exponentially stable if and only if the inhibitory feedback gain ( $g$ ) lies within an exact window (e.g., $7/16 < g < 1187/512$ for the exact construction). Below the lower bound, the system diverges; above the upper bound, it undergoes a Hopf bifurcation into oscillation.
- Critical Delay: At unit gain, stability requires the feedback delay ( $d$ ) to be below a critical value ( $d_c \approx 1.5675\tau$ ).
- State Dependence: If the inhibitory signal is state-independent (constant or replayed), the level is unstable. The signal must carry information about the current state of the network.
Information Rate Bound ( $h$ ):
- Applying the Data-Rate Theorem, the author proves that any feedback mechanism stabilizing the level must have a channel capacity $C$ at least equal to $h$ , where $h$ is the rate at which the network's instability generates information about its state (the sum of the positive real parts of the unstable eigenvalues of the clamped system).
- For the reference network, $h \approx 28.9$ bits/s.
- Efficiency: An efficient "zooming" coder can hold the level at rates just above $h$ (e.g., $1.02h$ ), whereas static codes require significantly higher rates (often $>2h$ to $8h$ ) due to quantization inefficiencies.
Information Identity and Regulation Budget:
- Theorem 4: For a linearized network with a white Gaussian channel, the rate at which the received signal accumulates information about the initial state is exactly $h$ , regardless of noise amplitude or gain (provided $g \neq 0$ ). This establishes $h$ as a signature of the system's instability.
- Bode's Integral: The same rate $h$ fixes the "regulation budget." The integral of the log-sensitivity function over frequency equals $h/2$ . This implies a trade-off: stronger feedback attenuates slow inputs only by amplifying faster ones.
Settling Point vs. Setpoint: The level is a "settling point" in the physiological sense: it resists transient perturbations but shifts under sustained input. The network lacks an integrating motif (e.g., slow calcium-dependent regulation) required to reject sustained biases, meaning the level moves in proportion to the input.

Significance and Claims
The paper claims to provide exact, testable statements regarding how a recurrent neural network can generate an attracting operating level without explicit setpoint variables or plasticity. The significance lies in:

Mechanistic Origin: Demonstrating that the "setpoint" arises from the flow of information through the recurrent loop itself, specifically the balance of excitation and inhibition.
Quantitative Limits: Establishing that maintaining this level imposes a strict, quantitative requirement on the information capacity of the inhibitory pathway ( $C \ge h$ ).
Distinction from Other Concepts: Clarifying that "information flow" in this context refers to the channel capacity required for stabilization, distinct from pairwise predictability statistics like transfer entropy, which may exist even when the level is not held.
Modest Scope: The author explicitly states that these results establish a possibility within a specific model class (fixed connections, no plasticity). They do not claim to prove that biological setpoints always arise this way, but rather that this mechanism is sufficient to generate an attracting level. The mapping to physiological setpoints remains a hypothesis to be tested experimentally.

The work bridges control theory and neuroscience, showing that the stability of neural firing rates is fundamentally constrained by the rate at which the network can transmit information about its own state through inhibitory feedback.

More like this