Comparison of semantic context and inductive biases on self-attention based models

November 22, 2022

One of the main developing fields of modern neural network research is the segmentation and classification of images. This can be applied to specific examples, such as medical image segmentation, or broader, more abstract terms, such as everyday pictures made up of a multitude of objects. This dissertation analyzes the current state of the art when it comes to image segmentation, using three different models that all share a common structure. We will investigate how building upon an attention model via the injection of a priori properties can affect the results of semantic segmentation. The development of this subject can lead to further developments in different scientific research areas, whether by developing new subjects or automating existing ones, from topics as broad as medical research to the development of smart cities.

Authors in the community:

• 

  A

  António Diogo Marques Cartuxo Farinha Pereira

  ist190019

Supervisors of this institution:

• 

  A

  Arlindo Manuel Limede de Oliveira

  ist12282

electrical-engineering-electronic-engineering-information-engineering - Electrical engineering, electronic engineering, information engineering

eng - English

Instituto Superior Técnico