Answers are grounded in this paper and cite their sources.
Parametrisations of one neural network architecture can reach nearly the same error yet route their internal computation very differently. Functional error alone is therefore insufficient to show what learning organises. For this reason, we describe multilayer perceptrons (MLPs) and convolutional networks (CNNs) at three levels: their structural disposition, their realisation under a declared condition and their re-realisation after causal intervention. At each level, a symmetry-invariant path measure yields intrinsic observables, together with a layer-conserved flow and a corresponding multiscale flow geometry. Keeping these explanatory layers apart locates each regularity where it holds and provides an intrinsic frame for explaining and interpreting neural networks. We study 2,847 optimiser trajectories from Bayesian AutoML and 149,040 interventions on 324 networks. Learning, formalised as an exact exponential tilt, concentrates the path measure: median effective occupancy falls to 12.9% of its initial value in all 48 MLPs and to 13.0% in 46 of 48 CNNs. The architecture fixes the structural occupancy level, the environment modulates its magnitude and adaptive optimisers drive it down. Within matched MLPs, Adam raises the entropy deficit by 3.323 nats and SGD by only 0.062. Of Adam’s rise, 96.7% follows the validation optimum, so structural and functional time decouple. The geometry records the same reorganisation, as MLPs contract their scale while CNNs internalise flow within boxes. All these findings concern the disposition, which states only what is made available. A conditional realisation provably reweights this structural support at a given state but never extends it. On the MPI3D and Fashion-MNIST benchmarks, reference conditions realise a median of 0.68% and 7.82% of the structural occupancy, respectively, and the activation class shapes which part is realised. Attenuating the realised carriers moves the response more than activation-matched controls in every network, although the activation law decides whether realised or structural carriers matter more. Learning therefore organises a symmetry-faithful field of computational possibility, of which each condition realises a narrower and load-bearing part.
No takes yet. Share an insight, caveat, or question.
Moharil et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: