Topology-Aware Attention: Injecting Structural Priors into Transformer Attention Mechanisms

Fuente: Zenodo
Guardado en:
Detalles Bibliográficos
Autor principal: Wang, Mingpei
Formato: Recurso digital
Publicado: Zenodo 2025
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866901540855349248
author Wang, Mingpei
author_facet Wang, Mingpei
contents <p>We propose Topology-Aware Attention, a novel paradigm that injects topological structural priors into Transformer attention mechanisms. While standard Transformers compute attention weights solely based on semantic similarity between tokens, our approach incorporates topological centrality measures to identify and emphasize structurally important tokens. </p> <p>Key findings:<br>- Topological features provide 99.6% independent information from semantic embeddings<br>- Topology-enhanced attention significantly increases the correlation between structural importance and attention received (from r=0.41 to r=0.81)<br>- We propose four injection schemes: Attention Bias, Gated Fusion, Centrality Weighting, and Dual-Stream Attention</p> <p>This work was conducted independently by an 18-year-old researcher, representing the first application of topological priors to Transformer attention mechanisms.</p>
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_18060955
institution Zenodo
language
publishDate 2025
publisher Zenodo
record_format zenodo
spellingShingle Topology-Aware Attention: Injecting Structural Priors into Transformer Attention Mechanisms
Wang, Mingpei
Transformer
Attention Mechanism
Topology
Persistent Homology
Deep Learning
Natural Language Processing
Topological Data Analysis
Machine Learning
<p>We propose Topology-Aware Attention, a novel paradigm that injects topological structural priors into Transformer attention mechanisms. While standard Transformers compute attention weights solely based on semantic similarity between tokens, our approach incorporates topological centrality measures to identify and emphasize structurally important tokens. </p> <p>Key findings:<br>- Topological features provide 99.6% independent information from semantic embeddings<br>- Topology-enhanced attention significantly increases the correlation between structural importance and attention received (from r=0.41 to r=0.81)<br>- We propose four injection schemes: Attention Bias, Gated Fusion, Centrality Weighting, and Dual-Stream Attention</p> <p>This work was conducted independently by an 18-year-old researcher, representing the first application of topological priors to Transformer attention mechanisms.</p>
title Topology-Aware Attention: Injecting Structural Priors into Transformer Attention Mechanisms
topic Transformer
Attention Mechanism
Topology
Persistent Homology
Deep Learning
Natural Language Processing
Topological Data Analysis
Machine Learning
url https://doi.org/10.5281/zenodo.18060955