MTR-RSNN: Multi-Timescale Recurrent SNNs with Kinetic NMDA Gating and Online Learning

MTR-RSNN is a recurrent spiking neural network that replaces the standard static NMDA gating model with a kinetic two-state Mg2+ unblocking mechanism, allowing the network to retain temporal state across spike intervals in a way the equilibrium approximation cannot. The architecture combines three neuron populations with distinct integration timescales, online e-prop learning without backpropagation through time, and a bare-metal C++/AVX2 implementation achieving sub-5 microsecond inference on commodity hardware. To the author's knowledge this is the first recurrent SNN to combine kinetic NMDA gating with online eligibility-trace learning.