My research focuses on deep learning with few labels and to leverage prior knowledge in the learning process.

Efficient Learning

In practice, the data and/or labels are often scarce or even absent. To enable AI algorithms for such limited settings, we develop efficient learning techniques that can deal with few labeled training samples. Examples of such techniques are: domain transfer, few-shot object detection, zero-shot recognition, out-of-distribution detection, self-supervised learning, and (simulated) data generation.

Incorporating Knowledge

To improve robustness and enable higher-level analysis, we leverage domain and expert knowledge and combine it with deep learning. Knowledge can fill the gaps of learning on limited datasets, making predictions more robust. Examples of such techniques that we research and develop are: graph networks, taxonomy-based learning, and, learning based on physical models (3D, geometry, augmentations).

More details about my research activities and projects can be found in the Blog posts.