Organization stays static
Gradient descent can produce specialization, but conventional layers do not explicitly sense that structure or reorganize their available pathways around it.
Adaptive networks / vision–language
FORDE vision–language model.
Functional Organisation via Route-Density Estimation: a stateful network that learns weights quickly and periodically reorganizes its own computational routes.
Interactive architecture
Fast learning inside slow reorganization.
The return edge is the idea: a cached functional map changes how the next fast-loop update is routed.
Case study
A model with a memory of its own organization
Couple two timescales
A CLIP-style dual encoder trains every step while a periodic Sense → Cluster → Smooth → Actuate loop turns accumulated statistics into route assignments.
Expose the internal map
The JAX/Flax repository includes the adaptive model, a baseline ablation, streaming input, tests, brain scans, feature-space plots, and a functional census.
Slow-loop detail
Sense, organize, stabilize, actuate
Sense
Cache gradient and activation statistics from ordinary contrastive training.
Cluster
Group units in a learned feature space to expose emerging functional roles.
Smooth
Regularize assignments so noisy short-term changes do not thrash the architecture.
Actuate
Write the stable map back into stateful layers and alter subsequent routing.
Evidence boundary
An experimental architecture with an explicit ablation path.
The repository establishes an implementation and diagnostic framework for adaptive organization. The baseline makes the architectural contribution testable; broader empirical superiority remains a question for sustained evaluation.