Personal Page

Abstract:

Recent work has shown that the attention maps of the widely popular DINOv2 model exhibit artifacts, which hurt both model interpretability and performance on dense image tasks. These artifacts emerge due to the model repurposing patch tokens with redundant local information for the storage of global image information. To address this problem, additional register tokens have been incorporated in which the model can store such information instead. We carefully examine the influence of these register tokens on the relationship between global and local image features, showing that while register tokens yield cleaner attention maps, these maps do not accurately reflect the integration of local image information in large models. Instead, global information is dominated by information extracted from register tokens, leading to a disconnect between local and global features. Inspired by these findings, we show that the CLS token itself, which can be interpreted as a register, leads to a very similar phenomenon in models without explicit register tokens. Our work shows that care must be taken when interpreting attention maps of large ViTs. Further, by clearly attributing the faulty behaviour to register and CLS tokens, we show a path towards more interpretable vision models.

Authors: Lappe, Alexander; Giese, Martin A.

Research Areas: Neural and Computational Principles of Action and Social Processing

Type of Publication: Article

Journal: arXiv

Year: 2025

Abstract:

Previous work has shown that neurons from body patches in macaque superior temporal sulcus (STS) respond selectively to images of bodies. However, the visual features leading to this body selectivity remain unclear. METHODS: We conducted experiments using 720 stimuli presenting a monkey avatar in various poses and viewpoints. Spiking activity was recorded from mid-STS (MSB) and anterior-STS (ASB) body patches, previously identified using fMRI. To identify visual features driving the neural responses, we used a model with a deep network as frontend and a linear readout model that was fitted to predict the neuron activities. Computing the gradients of the outputs backwards along the neural network, we identified the image regions that were most influential for the model neuron output. Since previous work suggests that neurons from this area also respond to some extent to images of objects, we used a similar approach to visualize object parts eliciting responses from the model neurons. Based on an object dataset, we identified the shapes that activate each model unit maximally. Computing and combining the pixel-wise gradients of model activations from object and body processing, we were able to identify common visual features driving neural activity in the model. RESULTS: Linear models fit the data well, with mean noise-corrected correlations with neural data of 0.8 in ASB and 0.94 in MSB. Gradient analysis on the body stimuli did not reveal clear preferences of certain body parts and were difficult to interpret visually. However, the joint gradients between objects and bodies traced visually similar features in both images. CONCLUSION: Deep neural networks model STS data well, even though for all tested models, explained variance was substantially lower in the more anterior region. Further work will test if the features that the deep network relies on are also used by body patch neurons.

Authors: Lappe, Alexander; Bognár, Anna Nejad, Ghazaleh Ghamkhari Raman, Rajani Mukovskiy, Albert; Martini, Lucas M.; Vogels, Rufin Giese, Martin A.

Research Areas: Neural and Computational Principles of Action and Social Processing

Type of Publication: In Collection

Authors: Smekal, Vojtěch Solanas, Tamás Szűcs Marta Poyo Lappe, Alexander; Giese, Martin A.; de Gelder, Beatrice

Research Areas: Neural and Computational Principles of Action and Social Processing

Type of Publication: In Collection

Authors: Lappe, Alexander; Bognár, Anna Nejad, Ghazaleh Ghamkhari Mukovskiy, Albert; Martini, Lucas M.; Giese, Martin A.; Vogels, Rufin

Research Areas: Biomedical and Biologically Motivated Technical Applications

Type of Publication: Article