ATMAI Growing Neural Gas Model Lab

An open research lab for understanding, benchmarking, and extending Growing Neural Gas (GNG).

Status: pre-alpha. The repository contains a research structure, not a released ATMAI model.

What we are investigating

How can a GNG graph learn topology online, adapt to changing data streams, and recover recurring concepts without confusing geometric similarity with semantic identity?

The research charter explains the scientific sequence. We begin by reproducing canonical GNG before adding drift or memory.

Structure

Labs are organized by research question, not by person:

labs/
├── 001-canonical-gng/
│   └── experiments/exp-20260727-ee260fe9-fritzke-baseline/
└── 002-controlled-drift/
    └── experiments/exp-20260727-eee720cf-frozen-baseline-drift/
src/
└── atmai_gng/

Authorship lives in the manifest and Git history. Folder names remain stable if people join, leave, or collaborate.

Experimental work stays inside its experiment. Reusable code enters src/ only after it has tests and reproducible evidence.

Contributing

1. For a new experiment or major change, open an issue first. Small fixes can start with a pull request.

2. Work in a fork or feature branch—never directly on main.

3. Add one focused experiment or change.

4. Run:

   python3 scripts/validate_repository.py
   python3 -m unittest discover -s tests -p 'test_*.py'

5. Open a draft pull request and merge only after the checks pass.

See CONTRIBUTING.md for the short contribution guide.

Public-repository boundaries

• Never commit secrets, private or clinical data, raw datasets, or model checkpoints.
• Record the source and license of code, papers, and datasets.
• Code without an explicit license may be studied but not copied.
• Label experiments as planned, running, complete, failed, or inconclusive.
• Do not present an exploratory result as an ATMAI capability.

Current evidence

The canonical Fritzke baseline is reproduced in labs/001-canonical-gng and its unchanged reference engine is promoted in src/atmai_gng.

The first controlled-drift diagnostic records adaptation under translation, rotation, density reweighting, and A-B-A recurrence. It also preserves the important negative boundary: low return error coexisted with a dormant geometric component, and continuing node growth confounded the recurrence comparison. No semantic memory claim is made.

License

Repository software is available under the MIT License. External code, data, papers, and artifacts retain their own terms.