Oliver Stoll
Oliver Stoll
Researcher
Tom Burgert, Oliver Stoll, Paolo Rota, Begüm Demir
ImageNet-trained CNNs are not biased towards texture: Revisiting feature reliance through controlled suppression
September
24,
2025
https://doi.org/10.48550/arXiv.2509.20234
Machine Learning| Dec 01, 2025
Rethinking how models "see"
Congratulations to BIFOLD researchers Tom Burgert, Oliver Stoll, and Begüm Demir from TU Berlin, and Paolo Rota from the University of Trento. They published a new study, that revisits a central claim in computer vision: so-called convolutional neural networks (CNNs) primarily rely on texture, rather than object shape, to recognize images. The publication was accepted as an oral presentation at NeurIPS 2025.