Mutual nearest-neighbor alignment

methods / Representation Alignment / Mutual nearest-neighbor alignment

Techniqueintermediate

Measures agreement between two representations' kernels as the mean overlap of their k-nearest-neighbor sets around each datapoint, normalized by k — a rank-based alternative to continuous kernel-alignment statistics like CKA.

Used in (2 observations)

After correcting CKA and mutual k-NN alignment for a proven width/depth-dependent null-baseline confound (O(d/n) for global spectral metrics vs O(k/n) for local-neighborhood metrics), the Platonic Representation Hypothesis's global cross-modal convergence largely vanishes while its local-neighborhood convergence survives, across real image-text and video-language pretrained model pairsmeasured

structure: Aristotelian Representation Hypothesis · models: BLOOMZ-560M, BLOOMZ-1.1B, BLOOMZ-1.7B, BLOOMZ-3B, BLOOMZ-7.1B, OpenLLaMA-3B, OpenLLaMA-7B, OpenLLaMA-13B, LLaMA-7B, LLaMA-13B, LLaMA-30B, LLaMA-65B, Gemma-2-9B-it, ViT-Base, ViT-Small, MAE ViT-Base (Masked Autoencoder), DINOv2 ViT-B/14, DINOv2 ViT-L/14, CLIP ViT-B/16, CLIP ViT-L/14, VideoMAE-base, VideoMAE-v2 · paper: Revisiting the Platonic Representation Hypothesis: An Aristotelian View

Cross-model and cross-modal representational alignment increases with scale and competencemeasured

structure: Platonic Representation Hypothesis · models: LLaMA-7B, LLaMA-13B, LLaMA-33B, LLaMA-65B, Llama-3-8B, Llama-3-70B, BLOOM-560M, BLOOM-1.1B, BLOOM-1.7B, BLOOM-3B, BLOOM-7.1B, OpenLLaMA-3B, OpenLLaMA-7B, OpenLLaMA-13B, Mistral-7B, Mixtral-8x7B, Gemma-2B, Gemma-7B, OLMo-1B, OLMo-7B · paper: The Platonic Representation Hypothesis