Relative representations (anchor re-coordinatization)

methods / Representation Alignment / Relative representations (anchor re-coordinatization)

Relative representations (anchor re-coordinatization)

Techniqueintermediate

Re-expresses every embedding as its vector of cosine similarities to a fixed set of anchor points from the same space, producing a coordinate system that is provably invariant to rotation, reflection, and rescaling — turning cross-model geometric convergence into a zero-shot stitching capability rather than only a diagnostic similarity number.

Used in (1 observation)

Re-expressing embeddings as cosine similarity to a fixed anchor set yields a rotation/reflection/rescaling-invariant coordinate system in which decoders and classifiers trained on one independently-trained encoder can be stitched, with zero fine-tuning, onto a completely different independently-trained encoder's representations of the same domainmeasured

structure: Platonic Representation Hypothesis · models: RoBERTa-base, ViT-Small, ViT-Base, RexNet-50, Cora GCN (~2000-model varied-hyperparameter ensemble), Moschella et al. (2022) convolutional (V)AE for MNIST/F-MNIST/CIFAR-10/CIFAR-100 · paper: Relative Representations Enable Zero-Shot Latent Space Communication