Generating behaviors through spikes that minimize free energy

Biological agents exhibit remarkable efficiency in perception and action, seamlessly integrating sensory input with prior knowledge to regulate internal states and generate purposeful behavior. An influential theory to explain this capability is the free energy principle, which models cognition and control as processes of probabilistic inference aimed at minimizing variational free energy. However, existing implementations remain at the algorithmic level and have largely relied on conventional, biologically implausible architectures. Here, we derive from first principles a novel Spiking Neural Network (SNN) model that perceives and generates behavior by minimizing the free energy through emitting spikes. Thus, it offers a more faithful model of neural computation that operates in a parallel and asynchronous regime. Our results show that our proposed spiking-free-energy constrainers can mimic some adaptive capabilities of biological agents, such as robustness against internal and external perturbations, while maintaining low latency and energy consumption. Particularly, we generalized spike coding networks for control—known for their sparse spiking activity and biological plausibility—to permit quadratic energy functions, like the objective used in continuous active inference. Thus, opening new avenues for advanced model-based probabilistic inference in neuromorphic hardware and its application in artificial intelligence, robotics and cognitive neuroscience.