Topological Convergence
Topological convergence in Aether means that model or residual behavior is observed through shape signals, not only scalar loss.
Current Internal Signals
The convergence modules use:
- scalar error;
- Betti-number history;
- centroid drift;
- residual sign-change and oscillation heuristics;
- fixed windows and thresholds.
Internal Convergence Shape
For a residual sequence \(r_i = y_i - \hat{y_i}\), the interpreter-level escalating regressor estimates shape using sign changes and oscillation counts. That is a lightweight residual heuristic, not persistent homology.
The persistent-homology path is separate:
let diagram = topology.ph(M, max_dim=2)~
let b = topology.betti(diagram, radius=0.5)~
Claim Boundary
It is accurate to say:
- Aether exposes topology and residual-shape signals for convergence logic.
- Some tests verify parser and interpreter paths for
seal untiland topology calls. - The core crate contains convergence and residual-analysis structures.
It is not yet accurate to say:
- every training loop terminates by persistent homology;
- topology improves model quality on external datasets;
- topological convergence replaces validation metrics;
- convergence behavior is benchmarked across model classes.