This repository contains experimental code, data, and figures for the Dual-Gradient Learning research program. The central result is a regime-structured view of robust learning under label noise, formalized and empirically validated in Paper A, with extensions across vision and NLP.
The repository is organized to support full numerical reproducibility of all reported results, figures, and statistics.
HIDEKI
Independent Researcher, Japan
ORCID: 0009-0002-0019-6608
Contact: hideki@r3776.jp
| ID | Title | Status / DOI |
|---|---|---|
| A | Designing Robust Learning as Phase Control: Identical Training Produces Stable Success or Irreversible Failure | DOI to be assigned |
| B‑I | The Sub‑Additivity Principle: How Integration and Value Alignment Jointly Determine Cognitive Capacity | https://doi.org/10.5281/zenodo.17970754 |
| B‑II | Sub‑additive Interaction between Integration and Value Alignment in Learning Systems: Evidence from 14,400 Simulations | https://doi.org/10.5281/zenodo.17940361 |
| C | ρ‑Design: Orthogonal Control of Gradient Direction and Mixing Strength in Learning Experiments | https://doi.org/10.5281/zenodo.17982236 |
| D | Beyond Dual Flatness: Curvature Emergence via Anisotropic Metric Perturbations | https://doi.org/10.5281/zenodo.18027618 |
| E | Natural Gradient Descent: Characterizing Local Efficiency and Coordinate Invariance in Controlled Settings | https://doi.org/10.5281/zenodo.18163401 |
.
├── notebooks/
│ ├── Paper-A/
│ │ ├── README.md # Detailed experimental ledger (v4.1)
│ │ ├── notebooks/ # Training / analysis notebooks
│ │ ├── results/ # CSV / JSON summaries
│ │ └── figures/
│ │ ├── main/ # Figure 1–5 (paper)
│ │ └── extended_data/ # Extended Data Fig. 1–8
│ │
│ ├── Paper-B/
│ ├── Paper-C/
│ ├── Paper-D/
│ └── Paper-E/
│
└── README.md (this file)
Problem
Robust learning under label noise is typically assumed to degrade smoothly as noise increases.
Core Finding
Dual‑gradient learning exhibits regime structure:
identical training conditions can produce either stable low‑error solutions or persistent high‑error failures.
Key phenomena
- Sharp boundary and variance peak as a function of mixing coefficient λ
- Bimodality, hysteresis, and metastability in vision
- Absorbing high‑error failure states at high noise
- Cross‑domain validation in NLP with domain‑specific phenomenology
- Causal decomposition of non‑recovery into time‑driven corrosion and path‑dependent damage
This motivates reframing robust learning as phase control rather than loss optimization.
| Domain | Series | Runs |
|---|---|---|
| Vision | A–Q | 1,773 |
| Vision | R–W | 1,273 |
| NLP | X | 790 |
| NLP (controls) | Y | 80 |
| Total | 3,916 |
All run counts are derived from released data archives and exclude archived or failed runs.
The complete census and definitions are documented in notebooks/Paper-A/README.md (v4.1).
This repository constitutes the complete reproducibility package for Paper A.
All numerical values and figures in the manuscript can be regenerated from:
- raw data (CSV / JSON),
- analysis notebooks, and
- fixed experimental protocols documented in Paper‑A/README.md.
Deterministic Execution Contract
torch.manual_seed(seed)
np.random.seed(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
torch.set_default_dtype(torch.float64)All experiments were executed on Google Colab (A100 GPU).
MIT License
This research was conducted independently. The theoretical foundations build on information geometry and related work by Shun‑ichi Amari and collaborators.
Last updated: March 2026