Statement
“Neural networks, trained with different objectives on different data and modalities, are converging to shared local neighborhood relationships” (Gröger, Wen & Brbić, 2026, arXiv:2602.14486, project statement). This is a deliberately narrower claim than the Platonic Representation Hypothesis (PRH): PRH’s own defining evidence uses both global spectral metrics (CKA-style) and local-neighborhood metrics (mutual k-NN) more or less interchangeably as if they measured the same underlying convergence. The Aristotelian view holds that once both metric families are corrected for their scale-dependent null baseline via Null-calibrated representational similarity, only the local- neighborhood family continues to show significant cross-model, cross-modality agreement — “the apparent convergence reported by global spectral measures largely disappears after calibration, while local neighborhood similarity… retains significant agreement across different modalities.”
Intuition
Two independently-trained networks might not agree on the overall shape of their representation spaces (how spread out or curved the whole cloud of points is, which is what global spectral metrics like CKA summarize), while still agreeing on something more local and, in a sense, more basic: which specific points are close neighbors of which. Aristotle’s empiricism emphasized particulars and their immediate relations over an idealized universal Form — the name reflects a parallel move here: from a claim about one shared global geometric Form (Platonic) to a claim about shared local relational structure between particulars (Aristotelian).
Properties
- A calibration story, not a new similarity metric. The hypothesis rests entirely on Null-calibrated representational similarity‘s correction, applied post-hoc to existing metrics (CKA, mutual k-NN, RSA, CCA, Procrustes) — it makes no claim that any of these metrics is individually flawed in its definition, only that raw, uncalibrated scores confound genuine cross-model agreement with a scale-dependent null baseline that itself grows with width and (via layer-selection) depth.
- Narrows, rather than refutes, PRH. The claim is not that representational convergence is illusory, but that it is real and significant specifically at the local-neighborhood level, while the global, “same overall Form” reading of convergence that raw CKA numbers seemed to support is largely an artifact of the uncorrected metric’s scale confound.
- Two theoretically-derived null baselines with different orders. Global spectral metrics have a null baseline of order (grows with representation width); local mutual k-NN has a null baseline of order (stays small since neighborhood size is typically far smaller than ambient dimension ) — this asymmetry in null-baseline growth rate is the specific mathematical reason the two metric families can diverge after calibration even when they agreed (both looked “high”) before it.
- Tested on real, named pretrained models across two modality pairings, following and extending the original PRH’s own image-text experimental setup (BLOOMZ, OpenLLaMA/LLaMA, Gemma-2-9B-it language side; AugReg/MAE/DINOv2/CLIP-LAION2B ViT vision side) to a video-language setting (VideoMAE plus DINOv2/CLIP vision-side encoders).
Key papers
- Huh, Cheung, Wang & Isola (2024). The Platonic Representation Hypothesis. ICML, arXiv:2405.07987 — the hypothesis this page calibrates and narrows; see representational-convergence for its own (uncalibrated) evidence.
- Gröger, Wen & Brbić (2026). Revisiting the Platonic Representation
Hypothesis: An Aristotelian View. arXiv:2602.14486, ICML 2026.
Derives closed-form null baselines for global spectral vs.
local-neighborhood similarity metrics under an independence null model,
introduces a permutation-based calibration framework
(Null-calibrated representational similarity) to correct for them,
and re-runs PRH-style image-text and video-language alignment
experiments through this calibration on real pretrained encoders. See
groger-wen-brbic-2026-null-calibration-shows-global-spectral-representational-convergence-largely-disappears-while-local-neighborhood-convergence-persists.