Self-Attention Neural Bag-of-Features

  • 2022-01-26 17:54:14
  • Kateryna Chumachenko, Alexandros Iosifidis, Moncef Gabbouj
  • 3

Abstract

In this work, we propose several attention formulations for multivariatesequence data. We build on top of the recently introduced 2D-Attention andreformulate the attention learning methodology by quantifying the relevance offeature/temporal dimensions through latent spaces based on self-attentionrather than learning them directly. In addition, we propose a jointfeature-temporal attention mechanism that learns a joint 2D attention maskhighlighting relevant information without treating feature and temporalrepresentations independently. The proposed approaches can be used in variousarchitectures and we specifically evaluate their application together withNeural Bag of Features feature extraction module. Experiments on severalsequence data analysis tasks show the improved performance yielded by ourapproach compared to standard methods.

 

Quick Read (beta)

loading the full paper ...