I explored whether internal neuron correlation structure in a trained CNN is just a visualization artifact or whether it is causally load-bearing.

Using a ResNet18, I: – mapped neuron–neuron activation correlations into a 3D manifold – ran controls (untrained weights, pixel-shuffled inputs) – performed targeted “soft merges” of neurons based on geometric proximity – compared against random merges – tested downstream plasticity via transfer learning

The geometry predicts which merges are low-impact vs disruptive, and geometry-guided consolidation is consistently safer than random consolidation.

Repo includes notebooks, failed experiments, batch statistics, and PLY artifacts for inspection: https://github.com/boglim1984/functional-geometry-hebbian-ma...

I’m sharing this as an exploratory mechanistic interpretability artifact rather than a polished paper—very open to critique.