Katharina Dobs
LightNews LightNews
banner
kathadobs.bsky.social
Katharina Dobs
@kathadobs.bsky.social
460 followers 370 following 37 posts
Professor at JLU | cognitive computational neuroscientist | mom of 3 | she/her
Posts Replies Media Videos
kathadobs.bsky.social
Good morning, Bsky'ers! Some people have asked for a life update, here it is: I finished my last chemotherapy yesterday! 🥳 It’s been a long ride, but I’m feeling happy, relieved, and incredibly grateful. I sort of always knew, but now more than ever: #cancersucks #sciencematters #sciencesaveslives
May 7, 2025 at 6:43 AM
kathadobs.bsky.social
We employed interpretability methods to identify the critical features DNNs focus on. Most relied on the face and mouth regions; however, only the Dual-task DNN, trained on face and object recognition, focused on the eyes.
Left plot: Visualization of the critical features in sample pareidolia stimuli that the Dual-task CNN uses to classify these stimuli as either ‘face’ or ‘object’. Green areas indicate positive class attribution, essential for classification, while red areas signify negative class attribution, detrimental for classification. We calculated the mean pixel ratio within each mask (ranging from 0, indicating the lowest, to 1, the highest) for classifying the stimulus as ‘face’ (red bars) or as ‘object’ (yellow bars). Left plot: This mask-based ratio is calculated for all the CNNs. The analysis revealed that the Dual-task CNN primarily relies on facial features, especially the eye and mouth regions, to classify the pareidolia stimulus as a face.
February 5, 2025 at 5:53 PM
kathadobs.bsky.social
🧐 Does face optimization affect pareidolia? To find out, we ran the same model-based RSA on DNNs across layers. Task optimization had a strong impact. DNNs w/ face identification showed weaker pareidolia effects than those w/out. Overall, dual-task training best mirrored the neural data.
Panels 1-3 display the results for each CNN: 1: Dual-task CNN (trained on face identification and object categorization), 2: Face-identification-and-Object-detection CNN, 3: Object-categorization-and-Face-detection CNN. In CNNs that included object categorization training, the representational similarity between pareidolia faces and real faces (‘Pareidolia ~ Faces’) increased across layers, indicating that these networks progressively represent pareidolia faces more like real faces in higher layers. Conversely, in CNNs lacking object categorization training, this similarity was weak, only reaching significance at early layers, and decreased across layers. The Dual-task CNN showed a unique representational pattern where pareidolia faces were represented closer to matched objects than to real faces (‘Pareidolia ~ Objects’), while still being more similar to real faces than objects (‘Pareidolia ~ Faces’) from mid-layers (layer 5) onward, matching pareidolia representations in MEG.
February 5, 2025 at 5:53 PM
kathadobs.bsky.social
Which task optimization best mirrors neural data? Using RSA, we compared MEG responses to activations of 5 DNNs (penultimate layer), each trained on unique combinations of face and object recognition tasks. DNNs w/ object categorization aligned more closely with neural data compared to those w/out.
Plot showing correlations between all time points of the neural MEG responses and the penultimate layer of each task-optimized CNN. CNNs including optimization for object categorization (i.e., Dual-task CNN, Object-categorization CNN, and Object-categorization-and-Face-detection CNN) showed a higher correlation than CNNs lacking optimization for object categorization (i.e., Face-identification CNN and Face-identification-and-Object-detection CNN).
February 5, 2025 at 5:53 PM
kathadobs.bsky.social
🤯 The pareidolia effect (blue line) peaked earlier but, surprisingly, was generally weaker than the similarity between pareidolia and objects (orange line). This suggests that although pareidolia start between faces and matched objects, they trend closer to matched objects overall.
Top plots: Idealized model RDMs reflecting two hypotheses on the representations of face pareidolia: i) ‘Pareidolia ~ Faces’ hypothesis posits that pareidolia faces are represented similarly to real faces but differ from matched objects (left), and ii) ‘Pareidolia ~ Objects’ hypothesis suggests that pareidolia faces are represented similarly to objects but differ from faces (right). Bottom plot B: Partial correlations between all time points of the neural MEG data and each idealized model RDM, controlling for the other model RDM. These correlations reveal that the ‘Pareidolia ~ Faces’ model uniquely explains variance in the neural data at early time points, indicating that pareidolia faces are initially represented similarly to real faces (peak at 165 ms after stimulus onset). However, the ‘Pareidolia ~ Objects’ model uniquely explains increasing variance at later time points, suggesting that pareidolia faces progressively resemble matched objects and become segregated from real faces (peak at 255 ms after stimulus onset). Throughout the MEG time course, the partial correlation values for the ‘Pareidolia ~ Objects’ model (orange line) are generally higher than those for the ‘Pareidolia ~ Faces’ model (blue line), indicating that pareidolia faces are represented more similarly to objects than to real faces overall.
February 5, 2025 at 5:53 PM
kathadobs.bsky.social
🧵 A bit late, but excited to share insights from our new publication in PLOS Comp Biol by Pranjul Gupta (journals.plos.org/ploscompbiol...). Here, we explored the phenomenon of face pareidolia—like seeing faces in our morning coffee! What might explain this effect?
A top view of a coffee cup filled with coffee. The surface of the coffee features foam that resembles a smiling face, complete with two eyes and a wide mouth.
February 5, 2025 at 5:53 PM