Independent AI researcher building formal theories of how neural agents learn — and why they fail at systematic composition.
A coupled dynamical-systems framework proving that compositional generalisation failure is a bifurcation phenomenon — not a capacity or data problem.
| Aspect | Detail |
|---|---|
| Core finding | SGD creates a stable low-coherence equilibrium ("σ-trap") that standard training cannot escape |
| Status | Preprint (preparing for peer review) |
| Validation | 45 experimental runs, Cohen's d = 9.08 (additive), 7.57 (multiplicative), p < 0.0001 |
| Theory | 16 complete mathematical proofs (existence, invariance, bifurcation, consistency) |
| Stack | Python, PyTorch, NumPy, SciPy |
Key predictions:
- Depth without coherence produces brittle generalisation
- Curriculum must target σ-dynamics, not just δ-accumulation
- Nine falsifiable predictions distinguishable from depth-only accounts
| Area | Technologies |
|---|---|
| AI/ML | PyTorch, NumPy, SciPy, scikit-learn |
| Theory | Dynamical systems, bifurcation analysis, singular perturbation |
| Dev | Python, Jupyter, Docker, Git |
- ORCID: orcid.org/0009-0004-2159-9371
- LinkedIn: linkedin.com/in/basyirin-amsyar
- GitHub: github.com/basyirin-dev
