Understanding shortcut learning in training deep neural networks for computer vision

Recent years have seen a great amount of success in using deep learning in publicly available and standardized datasets such as ImageNet. However, deep learning methods fail to perform as well in a lot of real-world applications such as medical imaging datasets. One potential reason for this challenge is that deep learning models tend to learn superficial features (aka spurious features) such as texture, color, etc, which merely correlate with the labels instead of discriminative features that can explain the labels. In medical imaging in particular, shortcut learning could potentially have fatal implications. Therefore, understanding this phenomenon and taking measures to prevent them can improve neural networks in terms of performance, generalizability to unseen demographics and interpretability. In this project, we will explore representation learning techniques for learning generalizable features. We will study the efficacy of these techniques on benchmarks that better reflect properties of real-world medical imaging problems.

Faculty Supervisor:

Samira E Kahou

Student:

Partner:

Imagia

Discipline:

Computer science

Sector:

Information and cultural industries; Professional, scientific and technical services

University:

École de technologie supérieure

Program: