Polysemanticity

What is Polysemanticity?

The phenomenon where a single neuron or computational component in a neural network is responsible for representing multiple, often unrelated features or concepts.

In other languages

한국어다의성

하나의 신경 단위가 서로 관련 없는 여러 가지 정보나 기능을 동시에 처리하는 현상

日本語ポリセマンティック性

一つのニューロンやアテンションヘッドが、互いに関連のない複数の異なる意味や機能を同時に表現する現象。

Related Terms

• Neural embeddingA technique that maps discrete data, such as words or tokens, into continuous vector spaces to represent semantic relationships numerically
• Latent SpaceA multi-dimensional mathematical representation of data within a neural network where concepts are stored as continuous vectors.
• Semantic NetworkA graph structure representing concepts as nodes connected by relationships, used to model knowledge and linguistic meaning
• Dense embeddingsNumerical vector representations of data that capture semantic relationships, allowing systems to compare the meaning of different inputs
• Multimodal EmbeddingsMathematical representations that combine multiple types of data, such as text, images, and audio, into a shared vector space for unified analysis.
• Embedding SpaceA multi-dimensional mathematical map where similar pieces of data, like related words or images, are placed closer together to help models understand semantic relationships.
• Semantic spaceA multi-dimensional mathematical representation where data points with similar meanings are located close to each other.
• EmbeddingAn embedding is a numeric vector representation of data (such as a word, sentence, or image) that captures meaning so a model can compare and search by similarity.
• Semantic EmbeddingThe process of converting text into numerical vectors that capture the meaning of words, allowing for efficient similarity-based retrieval.
• Vector EmbeddingA method of converting text, images, or audio into numerical vectors in a multi-dimensional space, allowing systems to measure semantic similarity between concepts.
• Cross-encoderA neural network architecture that processes query-document pairs jointly to output a single relevance score.