About me
I am a PhD Student at the University of Tübingen and the International Max Planck Research School (IMPRS) for Intelligent Systems (IS). I am co-supervised by Oliver Bringmann and Wieland Brendel. Before my PhD, I have obtained a B.Sc. and a M.Sc. in Physics at the Karlsruhe Institute of Technology. During my Master’s, I have participated in several research projects centered around numerical optics, and my Master thesis has been published at Nature Communications.
During my PhD, I have been working on improving the generalization capabilities of Deep Neural Networks beyond their training distribution. I have explored how we can make vision models more robust to distribution shifts. Beyond investigating different robustification methods, I have also analyzed the benefits of continual learning when the model is allowed to adapt to the encountered distribution shifts.
In my recent works, I am trying to understand how the OOD generalization capabilities of popular foundation models trained on large-scale datasets can be benchmarked. I also find it intriguing to investigate how multi-modality affects the learned representations and their generalizability.
I have completed a research internship at FAIR under the guidance of Ari Morcos and Kamalika Chaudhuri. During that internship, I have been working on pruning of large-scale datasets for CLIP training. This work has been published at ICLR 2024.
Latest publications (ICLR 2024)
Effective pruning of web-scale datasets based on complexity of concept clusters
Amro Abbas*, Evgenia Rusak*, Kushal Tirumala, Wieland Brendel, Kamalika Chaudhuri, Ari S. Morcos, ICLR 2024
We propose a pruning method where we aim to obtain optimal dataset coverage by assessing sample complexity; we report competitive results on the DataComp Medium benchmark and outperform regular OpenCLIP training on LAION with significantly less data. This project has also been presented as an oral contribution at the DataComp Workshop at ICCV 2024.
Does CLIP’s generalization performance mainly stem from high train-test similarity?
Prasanna Mayilvahanan*, Thaddäus Wiedemer*, Evgenia Rusak, Matthias Bethge, Wieland Brendel, ICLR 2024
CLIP’s ability to generalize to standard OOD benchmarks does not mainly stem from highly similar images in its training dataset. This project has also been presented as an oral contribution at the Workshop on Distribution Shifts (DistShift) at NeurIPS 2023.
Removing High Frequency Information Improves DNN Behavioral Alignment
Max Wolff, Evgenia Rusak, Wieland Brendel, Workshop on Representational Alignment, ICLR 2024
Removing high-frequency information by applying blur and resize transformations dramatically improves the model’s alignment with humans according to shape-bias and error-consistency.