Geometric Deep Learning for Computer Vision
||Geometric Deep Learning for Computer Vision|
||Georg Bökman <email@example.com>|
||Chalmers tekniska högskola|
||2023-09-01 – 2024-03-01|
- Recent research has shown that group equivariant CNNs can outperform ordinary CNNs on many tasks, including basic vision tasks such as image classification.
We are interested in investigating the differences and similarities between data augmentation and group equivariance or partial group equivariance. Data augmentation is used in all highly performant vision neural networks, so a better understanding of the interplay between equivariance and data augmentation could be of high value.
- Furthermore we want to investigate the relative benefits of group equivariance when scaling up to huge data sets such as ImageNet. If successful, this could be influential on the current debate on how to design large scale neural networks for vision.
- Finally, we want to investigate novel ways of enforcing group equivariance in neural networks.