The float in a toilet tank is an ideal controller: sensing the water level, it instantly brings the system into equilibrium without looking back at the past. Engineers applied this principle to neural networks. They rewrote the standard architecture so that instead of a cascade of computations, it solves a single equilibrium equation. The control signal is generated immediately — like the float's position. Conventional neural networks push the signal through many layers, accumulating error, but here the solution is found instantly. This provides a mathematical guarantee of stability: the analysis resembles checking whether ripples will subside in jelly. The error energy melts away, disorder does not grow. A clever training algorithm keeps the network within safe limits, then tests it separately — a technical inspection for an artificial brain. During training, strict conditions are observed, and the final check leaves no room for failure. A surprise awaits in the physical world: for a shaky construction with a barely alive motor, this 'brainless' float outperformed all schemes with memory. The reason is simple — power constraints prevent smart algorithms from unfolding, and instantaneous reaction without hesitation turns out to be the only correct one. The result: safe drones and robots incapable of fatal error.
as written by the authors arXiv:2607.11122
Implicit neural controllers (INCs) are static feedback laws that are evaluated through an algebraic fixed point {equation}; they include as special cases neural network controllers. We propose a so-called implicit representation of neural networks as a key enabling device that exposes the controller as a trainable linear interconnection closed through a known static activation map, thereby making well-posedness and Lyapunov/IQC analysis mathematically easy to handle. For finite-dimensional LTI plants, we first develop a rigorous analysis theory for a given INC, including Perron--Frobenius and norm conditions for well posedness, LMI/IQC certificates for exponential stability, and LMIs for discounted infinite-horizon quadratic performance. We then formulate synthesis as a certification-compatible heuristic search: training is carried out under explicit well-posedness constraints, implicit-differentiation formulas provide gradients, and the resulting controller is accepted only after independent post-training LMIs or regional admissibility checks are feasible. Finally, we establish constrained-control separation results: for a specific scalar unstable plant with hard actuator bounds, an INC achieves a strictly smaller discounted infinite-horizon cost than any admissible finite-order dynamic linear controller. Additional results cover quadratic state-input costs, comparison with linear static output feedback, and computable upper/lower-bound certificates. Numerical examples illustrate the mechanism and the resulting certified performance.
The controller's lack of memory is not a bug, but a feature. Under severe power constraints, a 'brainless' regulator without history outperforms complex dynamic schemes, like a rookie instantly reacting to a surprise while the expert hesitates, sifting through experience.