SUPR
Training and Understanding Modern Deep Networks
Dnr:

NAISS 2024/5-279

Type:

NAISS Medium Compute

Principal Investigator:

Hossein Azizpour

Affiliation:

Kungliga Tekniska högskolan

Start Date:

2024-06-01

End Date:

2025-06-01

Primary Classification:

10207: Computer Vision and Robotics (Autonomous Systems)

Webpage:

Allocation

Abstract

This proposal is for the scientific studies within the group of Hossein Azizpour. In Azizpour's group we try to understand modern deep networks and train them for impactful applications. Therefore, there are two main types of projects in the group. One is on the fundamentals of deep learning and the other on the application side. Fundamental of deep networks: here three main tracks are pursued for the foundations of trustworthy deep network: (i) interpreting and understanding trained deep networks especially from the lens of a functional analysis, (ii) explaining individual decisions of a trained deep network in a human-interpretable way, and (iii) quantification and robustness to uncertainty in the decisions of a trained deep network. Applications: on the application side, we apply the findings of the fundamental research on a few impactful applications including: (i) general computer vision, (ii) physics simulation, particularly fluid dynamics (iii) breast cancer diagnosis and prognosis specifically estimating risk of cancer and explaining the predictions, (iv) earth observation, especially uncertainty in the developments for detecting urbanization and forest fires, and (v) protein structure modelling, especially the interpretable models and uncertainty of predictions. As such all of the directions are either purely empirical or require empirical validations of the theories. Such empirical investigations for modern deep networks such as large ResNets and visual transformers can only be enabled with the help of a large GPU cluster such as Alvis or Berzelius. For the record, a similar proposal has allocation at Berzelius but due to the complementarity in the setup and uptime we apply for a parallel allocation at Alvis.