Anchor Divergence
A method for dynamically specifying semantic geometry on fixed contrastive learning representations. It models probability distributions over anchor points to select the corresponding Bregman divergence geometry, enabling multi-perspective semantic similarity measurement without requiring retraining of the underlying representation model.
Timeline (last 90 days)
The paper proposes that semantic similarity is inherently context-dependent, and that two images may be similar because they share the same object, visual style, or clinical diagnosis.
The paper proposes that representations obtained through contrastive learning inherently contain a family of geometries rather than a single geometry, and that the same set of fixed vector representations can be specialized to different semantic structures without retraining the model.
The anchor divergence method operates on fixed representations and obtains similarity metrics oriented toward different semantic requirements by specifying different anchor distributions, without retraining the encoder.
The paper conducted experiments on retrieval tasks, and the results show that anchor divergence provides an effective and efficient way to specify context-dependent semantic similarity.
Anchor divergence differs from metric learning in that it does not learn a fixed linear transformation; instead, it allows users to specify semantic context on demand at inference time through a probabilistic model, without needing to maintain a separate distance matrix or retrain an adaptation layer for each context.
Anchor divergence has a post-processing characteristic that enables context-adaptive capability in the representation space without modifying the already-trained representations.
As a newly released preprint, the generalization ability of the anchor divergence method on larger-scale and more diverse tasks remains to be further validated.
All Facts (7)
The paper proposes that semantic similarity is inherently context-dependent, and that two images may be similar because they share the same object, visual style, or clinical diagnosis.
50%UnverifiedThe paper proposes that representations obtained through contrastive learning inherently contain a family of geometries rather than a single geometry, and that the same set of fixed vector representations can be specialized to different semantic structures without retraining the model.
50%UnverifiedThe anchor divergence method operates on fixed representations and obtains similarity metrics oriented toward different semantic requirements by specifying different anchor distributions, without retraining the encoder.
50%UnverifiedThe paper conducted experiments on retrieval tasks, and the results show that anchor divergence provides an effective and efficient way to specify context-dependent semantic similarity.
50%UnverifiedAnchor divergence differs from metric learning in that it does not learn a fixed linear transformation; instead, it allows users to specify semantic context on demand at inference time through a probabilistic model, without needing to maintain a separate distance matrix or retrain an adaptation layer for each context.
50%UnverifiedAnchor divergence has a post-processing characteristic that enables context-adaptive capability in the representation space without modifying the already-trained representations.
50%UnverifiedAs a newly released preprint, the generalization ability of the anchor divergence method on larger-scale and more diverse tasks remains to be further validated.
50%