Albert Kjøller Jacobsen

Albert Kjøller Jacobsen

MSc. Eng. Human-Centered AI @ Technical University of Denmark

Sitemap

A list of all the posts and pages found on the site. For you robots out there is an XML version available for digesting as well.

Pages

Posts

Future Blog Post: Visual Question Answering with Knowledge-based Semantics

less than 1 minute read

Published:

This blogpost will highlight the main ideas, concepts and findings presented in the Bachelor thesis on Visual Question Answering (VQA) with Knowledge-based Semantics that considers training a VQA model with strong semantic and conceptual understanding by learning towards a latent space spanned by the Conceptnet Numberbatch embeddings on the OK-VQA dataset. The blog post will revisit the model formulation, yet keep an emphasis on the proposed idea of presenting VQA models with strong semantic and conceptual understanding rather than solely focusing on accuracy as the performance metric. In the process, explainability tools are exploited.

portfolio

projects

Bachelor Thesis: Visual Question Answering with Knowledge-based Semantics

Published:

This thesis considers an exhaustive state-of-the-art (SOTA) desription of VQA models, theoretical considerations and implementation details for training a VQA model learning towards a semantical- and conceptually strong latent space spanned by the Conceptnet Numberbatch embeddings as well an analysis of model behaviour by exploiting explainability tools.

Recommended citation: Jacobsen, Albert Kjøller; Højbjerg, Phillip Chavarria; Jacobsen, Aron Djurhuus. (2022). "Visual Question Answering with Knowledge-based Semantics." DTU Department of Applied Mathematics and Computer Science . https://findit.dtu.dk/en/catalog/62c6c822d4fccf03d747b3db

Master Thesis: On Riemannian Sharpness-Aware Minimization for General Loss Landscapes

Published:

The thesis investigates the sharpness-aware minimization literature and contributes to the literature by adding a reparameterization-invariant version that works for general loss landscape. The proposed method does not require a probabilistic model formulation or a pre-defined Riemannian manifold for working which FisherSAM and RiemannianSAM does, respectively.

publications

talks

teaching

workshops