🧬 biology

Spiking neurons as predictive controllers of linear systems

This paper introduces a scalable, biologically-inspired framework where spiking neurons act as predictive controllers for linear systems by emitting sparse, instantaneous impulses only when they drive the system toward a target, thereby eliminating the need for continuous rate-based representations while maintaining closed-form mathematical derivations for network connectivity and dynamics.

Original authors: Paolo Agliati, André Urbano, Pablo Lanillos, Nasir Ahmad, Marcel van Gerven, Sander Keemink

Published 2026-03-17

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Paolo Agliati, André Urbano, Pablo Lanillos, Nasir Ahmad, Marcel van Gerven, Sander Keemink

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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

The Big Idea: From "Constant Hum" to "Strategic Kicks"

Imagine you are trying to steer a massive, heavy ship through a foggy harbor.

The Old Way (Current Technology):
Most current computer controllers act like a person who never stops talking. They constantly shout instructions to the rudder: "Turn left a tiny bit! Turn left a tiny bit more! Stop! Turn right!" They send a continuous stream of signals. In the world of artificial intelligence, this is like using a "rate code," where the speed of firing represents the signal. It works, but it's energy-hungry and messy. It's like trying to push a heavy boulder by constantly shoving it with your hands.

The New Way (This Paper):
The authors propose a smarter, more biological way: The "Strategic Kick."
Instead of constant shoving, imagine you are a master surfer. You don't paddle constantly; you wait for the perfect moment, then give the board one sharp, precise kick to catch the wave. Once you kick, you let the wave (the physics of the system) carry you forward. You only kick again when you need to correct your path.

This paper teaches a computer brain (a Spiking Neural Network) to do exactly that: Wait, predict, and then kick.

The Core Problem: Why "Kicking" is Hard

In the real world, neurons communicate using spikes—tiny, instantaneous electrical pulses. Think of a spike like a single clap of thunder. It's loud, brief, and happens all at once.

The problem is that controlling a machine with just "claps" is tricky.

If you clap too early, you might push the ship the wrong way.
If you clap too late, you miss the wave.
If you clap randomly, you just make noise.

Previous attempts to use spikes for control tried to smooth out the claps into a continuous hum (like turning a drumbeat into a violin note). But the authors say: "Why turn the drumbeat into a violin? Let's use the drumbeat!"

The Solution: The "Crystal Ball" Rule

The authors created a new rule for when a neuron should fire (clap). It's called Predictive Control.

Here is the analogy:
Imagine you are playing a game of pool (billiards). You want to sink the cue ball into a specific pocket.

Reactive Control (The Dumb Way): You look at where the ball is right now. If it's far from the pocket, you hit it. But you don't think about the table's friction or the angle of the rails. You just hit it. On a real table, this often makes the ball bounce off the rail and miss the pocket.
Predictive Control (The Smart Way): Before you hit the ball, you use a "crystal ball." You ask: "If I hit the ball now, where will it be in 2 seconds? Will it be closer to the pocket then?"
- If the answer is YES, you hit the ball.
- If the answer is NO (maybe hitting it now will make it bounce off a rail and miss), you wait. You let the ball roll on its own momentum until the next perfect moment to strike.

The Paper's Breakthrough:
The authors mathematically proved that if you give a network of neurons this "crystal ball" ability, they can control complex machines (like a spring-mass system or a robotic arm) using only these sparse, instantaneous kicks.

How It Works in Practice

The paper tested this on two main scenarios:

1. The Spring-Mass System (The Bouncy Ball)
Imagine a weight attached to a spring. It bounces up and down.

Without Prediction: If you try to stop the bouncing by pushing it whenever it's far from the center, you might push it while it's already moving away, making it bounce harder. The system fails.
With Prediction: The network waits. It sees the ball is moving up. It knows that if it waits, gravity and the spring will naturally slow it down. It only gives a tiny "kick" to the velocity when it predicts the ball will overshoot the target. It uses the physics of the spring to do the heavy lifting, only intervening when absolutely necessary.

2. The Robotic Arm (The Reaching Task)
Imagine a robotic arm made of muscle-like fibers.

The network controls the "muscles" by sending spikes.
Instead of constantly firing to hold the arm steady, the network sends a few precise spikes to start the movement, then lets the "muscle" (the spring and damping physics) do the work.
The result? The arm moves smoothly to the target using far fewer spikes (less energy) than traditional methods.

Why This Matters

Energy Efficiency: Real brains are incredibly energy-efficient. They don't fire constantly; they fire only when needed. This method mimics that, making it perfect for neuromorphic hardware (chips designed to look like brains) which could run on tiny batteries.
Simplicity: The math behind this is surprisingly clean. You don't need a super-computer to figure out the control; the network figures it out locally, neuron by neuron, just by comparing "What happens if I fire?" vs. "What happens if I don't?"
Robustness: The paper showed that even if you "break" (silence) 36% of the neurons in the network, the system still works perfectly. The remaining neurons just pick up the slack. It's like a team of rowers: if one stops, the others adjust, and the boat keeps going.

The Takeaway

This paper is a bridge between Control Theory (how we steer machines) and Neuroscience (how brains work).

It tells us that we don't need to force biological neurons to act like continuous electrical wires. Instead, we can design systems that embrace the "spiky," "on-off" nature of real neurons. By giving these neurons a simple rule—"Only fire if you can predict that it will help us reach the goal"—we can build controllers that are efficient, robust, and surprisingly smart.

In short: Stop shouting instructions. Start listening to the physics, wait for the perfect moment, and deliver a single, perfect kick.

1. Problem Statement

Current approaches to using Spiking Neural Networks (SNNs) for control theory face a fundamental disconnect between the discrete nature of biological spikes and the continuous nature of classical control systems.

The Gap: Traditional control methods (e.g., Linear Quadratic Regulator - LQR) rely on continuous signals. Existing SNN control strategies typically filter spike trains to approximate continuous firing rates, which are then decoded into control signals. This approach treats spikes merely as a representation of state rather than the control signal itself, requiring continuous energy input and obscuring the specific impact of individual spikes.
The Challenge: Directly using instantaneous spike events (impulse control) as the control input is mathematically challenging due to low analytical tractability and scalability issues. Furthermore, for physically constrained systems (where position cannot change instantaneously), reactive control (based only on current state) often fails because a spike's effect on velocity does not immediately reduce the distance to a target position.

2. Methodology

The authors propose a novel framework where spikes are treated as direct, instantaneous control inputs ("kicks") rather than filtered signals. The methodology is derived by combining control theory principles with Spike Coding Network (SCN) theory.

Core Principles

Direct Impulse Control: The control signal $u(t)$ is replaced by a sparse spike train $s(t)$ , modeled as Dirac delta functions. The system dynamics are governed by $\dot{x} = Ax + Bs$ , where $B$ maps spikes directly to the system state.
Greedy Local Spiking Rule: Instead of optimizing a global cost function over time (which is computationally expensive for SNNs), each neuron independently decides to spike based on a local "greedy" condition. A neuron $i$ $i$ spikes if and only if the loss (distance to target) after the spike is strictly lower than the loss without the spike.
- Loss Function: $L = \|x(t+f) - z(t+f)\|^2 + \mu\Gamma(s) + \alpha\|r\|^2$ .
- Condition: $L^s_i < L^{ns}$ (Loss with spike < Loss without spike).
Predictive Horizon ( $f$ ): To handle physically constrained systems (like a mass-spring-damper), the algorithm incorporates a predictive window $f$ . Neurons evaluate the system state $f$ seconds into the future to determine if a spike will eventually reduce the error, allowing the system to exploit its intrinsic passive dynamics.

Network Derivation

From the loss comparison inequality, the authors mathematically derive the dynamics of a recurrent Leaky Integrate-and-Fire (LIF) network:
$\dot{V} = -V + G(\dot{z} + z) - Fx_0 - \Omega s$
Where:

$V$ : Membrane potentials.
$G$ : Forward weights encoding the target $z$ .
$F$ : Forward weights encoding the current system state $x_0$ .
$\Omega$ : Recurrent weights.
Low-Rank Connectivity: The derived recurrent connectivity matrix $\Omega$ is shown to be low-rank relative to the dimensionality of the downstream system, a structural property derived rather than assumed.

3. Key Contributions

Direct Spiking Control Paradigm: The paper establishes a scalable method for task-specific spiking control where spikes are the control signal itself, eliminating the need for rate-based approximations or continuous energy input.
Predictive Spiking Mechanism: It demonstrates that for physical systems, predictive control is necessary. Reactive control fails on constrained systems because it cannot account for the time delay between a velocity change and a position change. The predictive window allows the network to "anticipate" the system's trajectory.
Analytical Derivation: The authors provide a closed-form mathematical derivation of the network connectivity, dynamics, and control objectives directly from the control problem, without requiring learning or training.
Scalability and Robustness: The framework is shown to scale to high-dimensional systems (e.g., 500 neurons controlling 10 coupled oscillators) and remains robust to neuron silencing (cell loss) due to the redundancy inherent in the greedy local rule.

4. Results

The authors validated the framework through several simulations:

Linear Systems (Unconstrained vs. Constrained):
- Unconstrained: Reactive control works by "teleporting" the state toward the target.
- Constrained (Spring-Mass-Damper): Reactive control fails (the network remains silent because a velocity kick doesn't immediately reduce position error). Predictive control succeeds, where the network fires to adjust velocity, exploiting the system's natural oscillation to reach the target.
Comparison with LQR and Filtered-Spikes:
- Compared to continuous LQR and filtered-spike approaches, the predictive spiking controller achieves comparable tracking accuracy but with significantly fewer spikes (higher sparsity).
- Unlike filtered approaches that approximate continuous signals, the spiking controller produces discrete "kicks," resulting in a trajectory that orbits the target rather than settling perfectly on it, but with much lower energy expenditure.
Coupled Oscillators: The algorithm successfully controlled a system of 10 coupled masses using 500 neurons. When random neurons were silenced, the remaining neurons redistributed the control effort, maintaining performance.
Bio-inspired Motor Control: The framework was applied to a point-arm reaching task controlled by 8 oscillating spring-mass-damper units (simulating muscle fibers). The network successfully traced complex 2D trajectories using sparse spiking activity.

5. Significance and Implications

Neuroscience Insight: The work provides a biologically plausible mechanism for how real neurons could exert control over downstream actuators (like muscles) using sparse, discrete spikes. It suggests that the "greedy" nature of biological spiking is sufficient for effective control without global optimization.
Neuromorphic Engineering: The paradigm is highly suitable for neuromorphic hardware. Since the control signal is sparse and event-driven, it offers potential for ultra-low-power control systems where continuous analog signals are inefficient.
Theoretical Advancement: It bridges the gap between control theory and computational neuroscience by showing that optimal control objectives can be approximated through local, predictive spiking rules, offering a new perspective on the computational role of spikes in neural circuits.

In summary, this paper presents a rigorous, mathematically derived framework where spiking neurons act as predictive controllers for linear systems, demonstrating that sparse, discrete impulses can effectively control complex physical dynamics without the need for continuous signal approximation.