Not all data points in a dataset are equally important for the performance of a neural network. As training progresses, loss on some data points might become uninformative since the network already learned what it can from it. As such, it can be advantageous to observe the network training to serve it the right type of data at the right time. However, simply selecting data from a fixed dataset could be problematic when no image with the precise mix of attributes exists. The goal of this PhD project is to develop new learning algorithms for generating relevent data "on demand" in response to the need of the target network. This includes improving training efficiency by synthesizing the most relevant data, enforcing desired invariance by creating example, etc.

• As part of our team, you will develop novel approaches to adapt deep generative models (e.g. diffusion models, GANs, VAEs) as data sources to better train and validate downstream models.
• Furthermore, you exploit the controllability and knowledge present in generative base models to move past seeing datasets as just a collection of images.
• You discuss and develop new ideas within the deep learning and computer vision experts at Bosch Center for AI.
• Last but not least, publications in top-tier journals and at conferences follow.