Spiking neurons as predictive controllers of linear systems

A research team from Radboud University and the Donders Institute has published a breakthrough paper demonstrating that spiking neurons can function as mathematically rigorous predictive controllers. The work, led by Paolo Agliati, addresses a fundamental challenge in neuroscience and AI: how neurons use brief, sparse electrical pulses (spikes) to control complex systems. Previous approaches filtered spike trains into continuous signals, but this new framework treats spikes directly as control inputs through 'impulse control' theory.

The researchers derived exact mathematical rules for how spiking neural networks (SNNs) should be connected to control linear dynamical systems. Their key innovation is a predictive spiking rule where neurons only fire when doing so brings the controlled system closer to its target state. This creates extremely sparse activity patterns while maintaining precise control, mimicking how biological neurons might operate efficiently.

Importantly, the team showed this approach scales to high-dimensional networks and systems while maintaining full mathematical tractability. For physically constrained systems, they demonstrated that predictive control becomes necessary, with the control signal exploiting the system's own passive dynamics. This work provides the first closed-form derivation of SNN connectivity for control tasks, bridging optimal control theory with neuroscience principles.

The implications extend beyond theoretical neuroscience to practical AI applications. By showing how sparse, event-driven control can be mathematically optimal, this research provides a blueprint for designing more energy-efficient neuromorphic hardware. Such hardware could process information using the brain's efficient spike-based communication rather than the continuous computations of traditional AI systems.