ICML paper debunks Platonic Representation convergence, proposes Aristotelian view
phillip_isola · x · 2026-10-10
MIT's Phillip Isola shares his group's ICML 2026 paper 'Revisiting the Platonic Representation Hypothesis: An Aristotelian View' (Gröger, Wen, Brbić). Key findings:
- Existing representational similarity metrics are confounded by scale — deeper/wider networks systematically inflate similarity scores.
- The paper introduces a permutation-based null-calibration framework turning any similarity metric into a calibrated score with statistical guarantees.
- After calibration, global spectral convergence largely disappears; only local neighborhood similarity (not distances) retains significant cross-modal agreement.
- New hypothesis: neural network representations converge to shared local neighborhood relationships. A single global, roughly orthogonal map (with centering/normalization) aligns image and text embeddings well enough for reasonable text2img translation.
Isola says the results surprised him and links alternative perspectives (local relational, coarse-grained, global relational).
More from Research
- Mathematician John Urschel solved a decades-old problem with GPT Astra, staying optimistic on AI research — soumitrashukla9 · 2026-10-10
- State space models bring recurrence back with linear cost and transformer-era tricks — anshulkundaje · 2026-10-10
- IROS 2026 Best Paper: humanoid robot learns tennis rallies from 5 hours of human motion data — qinzytech · 2026-10-10
- AC2 lets you train custom decision models, lifting toxicity-detection F1 from 0.482 to 0.677 — rhythmrg · 2026-10-10
- KordingLab renames Plan Your Science, keeps free AI science-planning tool for researchers — KordingLab · 2026-10-10
- QuestionFirst's AI mentor critiques research plans, not with praise but 'which result would change your mind?' — KordingLab · 2026-10-10