← Back to papers

Form as Endpoint Not Starting Point: An Interpretive Reversal of the Platonic Representation Hypothesis

Target: Minds and Machines (Springer)

Abstract

The Platonic Representation Hypothesis (PRH) reports a fact: independently trained neural networks converge on geometrically similar internal representations. Its authors interpret this as evidence that learning systems approximate a pre-existing, mind-independent abstract structure. The paper rebuts that interpretation on two interlocking levels.

Level one (parsimony). Platonic ontology is explanatorily redundant — convergent-attractor theory under shared physical constraints already provides a complete account of the convergence phenomenon, without postulating any independent abstract realm.

Level two (explanatory reduction). The paper explains why the Platonic reading remains so compelling — the extreme robustness of certain attractors (which is unavoidable in any sufficiently expressive optimisation process) generates a phenomenology of necessity and mind-independence, naturally — but mistakenly — inviting the inference to ontological independence. Robustness is mistaken for ontological independence; the impulse to posit an independent realm is the natural but unnecessary response to attractors that cannot be escaped.

The paper proposes the Reversed Platonic Representation Hypothesis (RPRH): convergent representations are emergent fixed points of optimisation processes under shared constraints, not ontologically prior structures being approximated; the intuition that they must independently exist is itself a structurally predictable cognitive consequence of their robustness.

Keywords: Platonic Representation Hypothesis; abstract objects; structural realism; convergent attractors; philosophy of mathematics; naturalism; neural-network convergence