Origins
The framework didn't arrive in one paper. It emerged from a 2023 neuroscience-led question and was reformulated three times as the empirical scope expanded.
2023 — The seed paper
Brain development dictates energy constraints on neural architecture search: cross-disciplinary insights on optimization strategies — arXiv:2310.03042.
The starting question was practical: AI's neural-architecture-search community was treating prediction error as the optimisation target, but developmental neuroscience suggests that real brains optimise energy first and prediction second. Could energy-first NAS beat prediction-error-first NAS?
2025 — Philosophical formulation
A Framework for Convergent Design in a Computational Universe — PhilSci archive 26949.
The neuroscience question generalised: if energy-first optimisation is the rule for biological networks, the same principle should apply to any system extremising a functional that includes connectivity cost. Schöner and Kelso's dynamic-coordination theory (1988) provided the mathematical bridge — a network-weighted action could subsume both the biological and the engineered cases.
2026 — Quantitative validation
Four papers in 2026 operationalise and test the framework:
- Causal thinking in physiology (Frasch 2026a, J Physiol, DOI 10.1113/JP290762) — formal framework + multi-scale unification. Read →
- Minimum-Action Learning (Frasch 2026b, arXiv:2603.16951) — physics-law discovery at reduced training energy. Read →
- minAction.net: Energy-First Neural Architecture (Frasch 2026c, arXiv:2604.24805) — 2,203-experiment validation at the architecture-design scale. Read →
- Modularity Emerges from Action-Functional Constraints (Frasch 2026d, arXiv preprint in submission) — biology-scale validation in marine metagenomic networks. Code and data: github.com/martinfrasch/tara-modularity. Read →
The site reflects this arc: the same framework, four independent tests, each making the case for the architectural-modularity prediction in its own domain.