PhD and PostDoc positions

No openings currently.

Student researcher positions

We currently have several openings for Master's or Bachelor's students. Each position is paid and can be adapted to a thesis or a research assistant (HiWi) project. We consider applications until positions are filled. Please reach out to the listed contacts. Physical presence in Munich and enrollment in a German university are required.

Semantic label smoothing
Contact: luca.schulze-buschoff@helmholtz-munich.de, can.demircan@helmholtz-munich.de
Tags: machine learning, computer vision, object recognition, representation learning
Description: Computer vision models are often trained on large scale visual object recognition data sets such as ImageNet. These data sets rely on a one-hot encoding approach for the target labels: the correct image class is set to 1, while all other classes are assigned 0. Therefore, if a model assigns the label “Dalmatian” to a “Hungarian Viszla”, it is penalized in the same way as if it had assigned the label “House” instead. This does not consider semantic structure at all, and semantic information can improve model performance and robustness if used correctly [1].
In this project, you will combine semantic information with an engineering trick called label smoothing [2], which is commonly used to train image models. You will come up with a loss function that combines the two concepts and train computer vision models on object recognition datasets (such as CIFAR). Finally, you will evaluate the models' performances and representations.

[1] Muttenthaler, L., Linhardt, L., Dippel, J., Vandermeulen, R. A., Hermann, K., Lampinen, A., & Kornblith, S. (2024). Improving neural network representations using human similarity judgments. Advances in Neural Information Processing Systems, 36.
[2] Müller, R., Kornblith, S., & Hinton, G. E. (2019). When does label smoothing help?. Advances in neural information processing systems, 32.