Overview
A direct empirical follow-up to the ONN + ORTSF framework paper. The original manuscript established mathematical performance bounds for topology-preserving neural networks; this one asks whether — and under what conditions — those bounds are actually reachable on a working machine.
What the paper shows
- Enhanced regime (𝓛_topo = 0.0792) reaches 99.14% of the theoretically predicted optimum.
- Advanced regime (𝓛_topo = 0.0234) reaches 99.75%.
- Counter-intuitively, minimal connectivity (k = 2) and extreme precision (surgery decay δ = 0.0005) outperform denser, coarser configurations — inverting conventional neural-network design wisdom.
- Results transfer to transformer architectures (14.7% perplexity reduction) and graph neural networks (2.3× faster convergence on WikiText-103).
Where it sits
Downstream of the accepted ONN + ORTSF paper, which provides the theory this one validates. Upstream of the production-grade ONN implementation documented in the research notes.