Asymmetric Similarity in Neural and Cognitive Representations

This direction is merged into another direction: asymmetric semantic association. Its last version stays readable and citable here.

Version 1 of 114 papers23 conceptsfrozen 2026-10-04

In geometry, distance is always symmetric: the distance from point A to point B is the same as the distance from B to A. Most modern AI models inherit this assumption by using metrics like cosine similarity, dot products, or Euclidean distance to compare concepts. However, human cognition does not work this way. We find it natural to say "an ellipse is like a circle," but far less natural to say "a circle is like an ellipse."

Recent research across linguistics, vision, and phonetics suggests that for AI to truly align with human perception, it must move beyond these symmetric "geometric" assumptions and embrace asymmetric similarity.

The Failure of Geometric Symmetry

The core problem with symmetric models is that they treat relationships as undirected edges. If a model places the word stork near baby in a vector space, it implicitly assumes that baby evokes stork just as strongly as stork evokes baby.

As Justin Washtell and Katja Markert demonstrate in A Comparison of Windowless and Window-Based Computational Association Measures, nearly 40% of our strongest mental associations are "syntagmatic" and directional. In speech perception, Yair Lakretz and colleagues found that humans exhibit directional biases in phoneme confusion—one sound is often more easily mistaken for another than vice versa—a nuance that standard metric learning fails to capture (Metric Learning for Phoneme Perception).

This asymmetry often stems from the relationship between prototypes and variants. A "circle" is a cognitive prototype; an "ellipse" is a variant. We compare the variant to the prototype, not the other way around. Murray Campbell and Wei Zhang show that while static word embeddings struggle with this, contextual models like BERT begin to capture these "ellipse-to-circle" directional nuances (Circles are like Ellipses, or Ellipses are like Circles?).

Modeling Directional Association

To fix this, researchers have developed several mechanisms to inject asymmetry into computational models:

1. Input-Output Vector Duality

In models like Word2Vec, every word actually has two vectors: an "input" vector (when it is the cue) and a "context" vector (when it is the target). Standard practice discards one or averages them. However, Fatemeh Torabi Asr, Robert Zinkov, and Michael N. Jones argue that by calculating the similarity between the input vector of a cue and the context vector of a target, we can mathematically represent asymmetry (Querying Word Embeddings for Similarity and Relatedness). This allows the model to reflect that stork evokes baby more strongly than the reverse.

2. Hierarchical Specialization

In ontologies, moving from a specific concept (sparrow) to a general one (bird) is cognitively different from the reverse. Vincent Schickel and Boi Faltings developed the OSS metric, which treats similarity as a function of information transfer within a hierarchy (OSS: A Semantic Similarity Function based on Hierarchical Ontologies). Generalizing is "cheaper" than specializing, creating a natural asymmetry that matches human intuition.

3. Tversky Similarity

Moving away from geometric "distance" entirely, Dan Jurafsky and Moussa Doumbouya have introduced Tversky Neural Networks (TNNs) (Tversky Neural Networks). Based on Amos Tversky’s 1977 Contrast Model, this approach treats similarity as a comparison of feature sets. Because the weight of features unique to the "subject" can differ from the weight of features unique to the "referent," the similarity is inherently directional.

Why Asymmetry Works

Asymmetric models work because they acknowledge the context of comparison. In visual perception, Maria Attarian and Michael C. Mozer found that modeling similarity as a directional relationship—where one object is the "query" and the other the "reference"—allows AI to predict human visual choices with over 90% accuracy (Transforming Neural Network Visual Representations).

By breaking the symmetry of the dot product, these models stop viewing the world as a static map of points and start viewing it as a network of directional associations. This shift better captures how humans navigate meaning: we don't just see how "close" two things are; we see how one thing leads to another.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_asymmetric_similarity_in_neural_and_cogn_dbe79385,
  title = {Asymmetric Similarity in Neural and Cognitive Representations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/asymmetric-similarity-in-neural-and-cognitive-representations/rdr_dbe79385a32c48878b1124a91bdcf6d4},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_asymmetric_similarity_in_neural_and_cogn_dbe79385_v1,
  title = {Asymmetric Similarity in Neural and Cognitive Representations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/asymmetric-similarity-in-neural-and-cognitive-representations/rdr_dbe79385a32c48878b1124a91bdcf6d4/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}