Predictive-Coding-Inspired Reconstruction of Nonlinear Dynamics in Spiking Neural Networks

Accurate reconstruction of nonlinear time-dependent dynamics is a fundamental challenge for recurrent spiking neural networks and neuromorphic signal-processing systems. In this work, we compare two prediction-error architectures for nonlinear state reconstruction: a model with an external comparator and a model employing a dedicated spiking population to encode prediction error. Both architectures are implemented within the Neural Engineering Framework and trained using the local Prescribed Error Sensitivity learning rule. The proposed approach is evaluated on a two-dimensional periodic oscillator and the nonlinear Lorenz system. The reconstruction fidelity is assessed using the reconstruction error together with amplitude-envelope correlation, instantaneous-frequency correlation, phase-locking value, mean phase difference, and computational efficiency. The results demonstrate that both architectures accurately reconstruct nonlinear dynamics, while the spiking error representation provides a larger stable operating region and a more explicit spike-based representation of the bottom-up prediction-error pathway.