Dendritic Contextual Routing

Dendritic-inspired contextual routing with affine modulation and micro-replay for sequential feature-conflict learning.

This repository is a reproducibility package for experiments on contextual affine modulation, dendritic-inspired routing, and micro-replay in sequential learning.

Developed under OPAL-dev / OPAL.inc as an independent research exploration on dendritic-inspired contextual routing, affine modulation, and micro-replay for continual learning.

Repository: https://github.com/OPAL-dev/dendritic-contextual-routing

---

Core question

When does contextual routing become functionally necessary, and what is required to preserve it under sequential training?

The project studies this question through a controlled benchmark called SDFC shared-head (Same-Dimension Feature Conflict), where the same input dimensions must be interpreted differently across tasks.

This creates explicit feature-sign conflicts and mirror-task pairs, making it possible to test whether a model truly uses contextual control rather than merely learning a shared representation.

---

Main result

On SDFC shared-head, models without useful contextual conditioning remain near chance level, while context-conditioned models solve the benchmark under joint training.

In sequential training, however, even strong context-conditioned models suffer severe interference. A small replay buffer fixes this:

A replay buffer containing only 2% of each task’s training set raises final accuracy from about 64% to 95.4% and reduces forgetting from about 43% to 1%, nearly matching joint training.

The oldest mirror-conflicted task, task 0, recovers from about 28% to 94% with only a 2% replay buffer.

---

Key findings

1. Contextual conditioning is necessary on SDFC shared-head.

2. A simple multiplicative dendritic gate is insufficient.

3. The useful primitive is additive + multiplicative affine modulation:

   h = gamma(context) * h_basal + beta(context)
   

4. A separated affine dendritic variant implements the same functional primitive:

   h = g(context) ⊙ h_basal + a(context)
   

5. film_full and dendritic_affine_separate are statistically indistinguishable across replay budgets.

6. Micro-replay preserves the contextual solution under sequential learning.

In short:

Contextual affine modulation solves the feature-conflict structure; micro-replay preserves it across sequential learning.

---

Final replay budgets

Replay fraction	Examples per task
0%	0
2%	200
5%	500
10%	1000

---

Final models

The final comparison focuses on:

• film_full
• dendritic_affine_separate

Earlier experimental branches also included MLP baselines, no-context dendritic controls, additive/multiplicative ablations, apical unlock diagnostics, PermutedMNIST, and SplitMNIST controls.

---

Repository structure

.
├── src/
├── scripts/
├── configs/
├── artifacts/
├── results/
│   ├── raw_csv/
│   ├── processed/
│   └── main_tables/
├── paper/
│   ├── figures/
│   └── results_section_dendritic_v2.md
├── docs/
│   ├── README_REPRODUCIBILITY.md
│   ├── EXPERIMENT_LOG.md
│   └── RELEASE_CHECKLIST.md
├── CITATION.cff
├── LICENSE
└── README.md

---

Quick reproduction

From the repository root:

python -m src.main --make-benchmark --benchmark-seed 12345

powershell -ExecutionPolicy Bypass -File .\scripts\run_sdfc_replay_joint_multiseed.ps1

powershell -ExecutionPolicy Bypass -File .\scripts\run_sdfc_replay_microbuffer_multiseed.ps1

Curated final CSVs are stored in:

results/raw_csv/

Final paper tables are stored in:

results/main_tables/

Final figures are stored in:

paper/figures/

---

Main final results

Model	Replay	Accuracy	Forgetting
film_full	0%	~63.9%	~43.2%
film_full	2%	~95.4%	~1.1%
film_full	5%	~95.9%	~0.5%
film_full	10%	~96.0%	~0.3%
dendritic_affine_separate	0%	~63.8%	~43.2%
dendritic_affine_separate	2%	~95.4%	~1.1%
dendritic_affine_separate	5%	~95.9%	~0.4%
dendritic_affine_separate	10%	~96.0%	~0.4%

---

Attribution

This project was developed under OPAL-dev / OPAL.inc as an independent research exploration on contextual routing, continual learning, and dendritic-inspired architectures.

Main research and implementation: MantHalo / OPAL-dev.

Experimental design and analysis were assisted by multiple AI systems and cross-checked through iterative review.

---

Citation

If you use this repository, please cite it using the metadata in CITATION.cff.

A release DOI can be generated through Zenodo by archiving a GitHub release.

---

License

This project is released under the MIT License. See LICENSE.