Orthographic-to-Semantic Abstraction in Matched Spiking and Artificial Recurrent Networks

This independent research manuscript investigates how small recurrent neural systems transform elementary graphemic events into semantically organized population trajectories. A three-layer recurrent spiking neural network (SNN; 18,128 parameters) is compared against an exactly parameter-matched continuous recurrent ANN.The experiments examine hierarchical orthographic-to-semantic organization, sensitivity to input order, precise spike timing, heterogeneous neuronal dynamics, hard orthographic distractors, relation-disjoint generalization, supervised contextual transfer, and post-stimulus recurrent dynamics. Word-aligned settling analyses and causal interventions are used to distinguish semantic organization expressed across an integrated trajectory from organization present in any single terminal state.Results show that semantic organization becomes increasingly dominant with depth under semantic supervision. In the SNN, ordered temporal input is important for constructing this organization, whereas precise output spike timing is not required for semantic discrimination. After stimulus offset, layer-3 semantic separation continues to reorganize during recurrent evolution, and removing recurrent interaction weakens this transformation. The resulting organization is expressed most clearly in the integrated stimulus-plus-settling trajectory rather than in a stable instantaneous semantic state.The matched ANN also undergoes substantial post-stimulus reorganization, but with a slower and quantitatively different temporal profile. Across the broader evaluation, the ANN shows stronger relation-disjoint and contextual generalization, whereas the SNN shows greater sequence dependence, operational sparsity, and robustness to orthographically deceptive negatives.The accompanying source code, datasets, experiment configurations, per-seed metrics, causal intervention analyses, and reproducibility scripts are publicly available through the associated software artifact.