SUSNN: A Self-Unifying Spiking Neural Network Architecture Driven by Prediction-Error Minimization

We present SUSNN (Self-Unifying Spiking Neural Network), a spiking neural network architecture organized in three-dimensional space and governed by a single organizing principle: intelligence as the minimization of prediction error. The network comprises an input face, a prediction face, and a spatially structured intermediate bulk whose initial coordinates may be derived from scaled large-scale cosmic structure, yielding a mixture of local clustering and global scattering. Neurons follow an integrate-and-fire variant with subtractive reset and dynamic thresholds maintaining firing-rate homeostasis. Learning operates through spike-timing-dependent plasticity (STDP) in a wake state, while a sleep state enables structural plasticity—growth and pruning of connections—driven by Hebbian co-activation traces. Crucially, motor neurons receive no error signal directly; they must learn through the influence of their actions on the environment, forming an indirect closed loop that endows the system with an intrinsic embodiment requirement. A minimal 55-neuron implementation on a 1-D foraging task exhibits above-baseline behavioral change (38.7% vs. 35.0% success), emergent motor asymmetries, and self-organized structural growth (400 → ~700 connections). We discuss the credit-assignment challenges of indirect closed-loop learning and outline scaling experiments enabled by a vectorized engine implementation. Keywords: spiking neural network; STDP; predictive coding; structural plasticity; embodied intelligence; neuromorphic computing Code: https://github.com/GMUnitX/SUSNN