Possible Attention Mechanism (Pam) accepted for publication
An article by the team of authors of the Department of Software Systems and Technologies was accepted for publication in the journal indexed in Scopus, devoted to the development of a new approach to attention mechanisms in neural networks.
Possibilistic Attention Mechanism
The work presents a new attention mechanism based on the theory of possibilities, as an alternative to the traditional softmax. The approach allows to solve the «paradox of attention» when merging on the principle of «product-of-experts» and to model situations with contradictory evidence more correctly.
The results are demonstrated on machine translation and image classification benchmarks. The method shows competitive results with better interpretability.
(This is a draft example. Add journal title, year, DOI link, full list of contributors.)