acceptodds
Desk rejected at ICLR 2027

A Periodic Table of Machine Intelligence: Periods, Groups, and a Law of Shell Filling

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Scaling laws relate compute to loss. They do not say which qualitative regime a model occupies, or when an auxiliary constraint will help. We formalize the Periodic Table of Machine Intelligence (PTMI) as a product of periods (interaction range on the computational graph: pointwise, local, global, plus iterative, routed, and autonomic lifts) and groups (functional valence: invariant, discriminative, predictive, generative, agentic, latent). Capability mass tags isotopes of a cell; it is an estimator, not a thermodynamic potential. Two theorems make the original “shell filling” claim checkable. A linear residual whose stencil does not vanish on a representable truth strictly biases Tikhonov estimation; an under-receptive surrogate is a different operator, so a mismatched physics term can raise risk. Iteration lifts receptive field by composition, while width inside the pointwise period is a semigroup and plateaus on local tasks. I-Con already tabulated losses; PTMI's second axis is periods. We treat LIRN as a didactic occupant of the already-populated local-generative-iterative cell, not as a claim of first occupancy, and show on 1D/2D/3D Poisson and NN graphs that compatible local iteration beats mismatched PINN-style MLPs and width scaling.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
scaling laws, inductive bias, receptive fields, iterative refinement, physics-informed learning, graph neural networks.

Related papers

Discussion (0)

Sign in to comment.