SIGNAL//DESK
AI/MLsrc: curated AI glossary

Attention mechanism

A way for an AI to read a sentence by deciding which words are most important to look at together, like how you might focus on specific clues to understand the meaning of a complex paragraph.

A mechanism that calculates dynamic relevance scores between tokens in a sequence, allowing the model to weigh the importance of different inputs relative to one another when generating an output.

The operation that computes context-dependent weights between elements of a sequence, letting a model focus on the most relevant tokens; self-attention is the core of the Transformer.

evolution

  1. 2014-09 · history
    Bahdanau et al. introduction

    The attention mechanism was first introduced to allow neural machine translation models to focus on relevant parts of the input sequence.

  2. 2015-02 · history
    Luong et al. refinement

    Introduced global and local attention mechanisms, standardizing the dot-product and general scoring functions.

  3. 2017-06 · history
    Attention Is All You Need

    The Transformer architecture replaced recurrent layers with multi-head self-attention, enabling parallel processing and long-range dependency modeling.

  4. 2020-10 · history
    Vision Transformer (ViT)

    Demonstrated that pure attention mechanisms could be applied to image patches, effectively replacing convolutional neural networks in computer vision.


← all terms