iiinova — where your digital self keeps evolving

Lenny van Dyck

@levandyck.bsky.social

224 followers 286 following 27 posts

PhD candidate in CogCompNeuro at JLU Giessen Exploring brains, minds, and worlds 🧠💭🗺️ levandyck.github.io

Posts Replies Media

Lenny van Dyck @levandyck.bsky.social · 05/08/2026

And since we're at CCN, here's a little preview of all the models we tested so far 🚀

120

Lenny van Dyck @levandyck.bsky.social · 05/08/2026

Excited to be at #CCN2026 in New York to present brand-new work with @kathadobs.bsky.social 🥳 By testing hundreds of categories, we found striking differences in functional specialization between visual cortex and DNNs 🧠 Come find me at poster F81 on Thursday 1:45-3:30 p.m. Happy to chat!

1235

Lenny van Dyck @levandyck.bsky.social · 24/02/2026

Finally, we examined how DNNs integrate faces and bodies: DNN responses to whole persons approximated the sum of their responses to isolated faces and bodies, suggesting part-based rather than holistic integration. 8/n

100

Lenny van Dyck @levandyck.bsky.social · 24/02/2026

Back in DNNs, we tested whether mixed selectivity is functionally relevant: We trained a DNN on different person perception tasks and lesioned each unit type. Lesioning mixed-selective units led to deficits across tasks, confirming these units actively support person perception. 7/n

110

Lenny van Dyck @levandyck.bsky.social · 24/02/2026

While face- and body-selective units explained substantial unique variance in their corresponding regions, they mainly explained shared variance. Crucially, this shared component increased from posterior to anterior regions, suggesting that integration grows along the cortical hierarchy. 6/n

110

Lenny van Dyck @levandyck.bsky.social · 24/02/2026

Let’s now link this to visual cortex: Mixed-selective units best predicted activity in both face- and body-selective regions, suggesting that they encode more integrated person information than generally assumed. 🪢 5/n

100

Lenny van Dyck @levandyck.bsky.social · 24/02/2026

Let’s first look at computational models of visual cortex: Diverse DNNs developed distinct face- and body-selective units, but also mixed-selective units responding to both categories. This suggests that face-body integration naturally arises from learning to recognize whole persons. 🕺 4/n

120

Lenny van Dyck @levandyck.bsky.social · 11/08/2025

Really looking forward to #CCN2025! On Tuesday, I'm presenting new work with @kathadobs.bsky.social on segregated vs. integrated face & body processing in visual cortex 😊🧍🧠 Using DNNs & fMRI, we test competing hypotheses, finding both distinct & shared selectivity. Come by Poster A64 for more.

0145

Lenny van Dyck @levandyck.bsky.social · 18/06/2025

Individually, they followed a striking topography. 📍 Distinct subclusters within category-selective areas 🌐 But sparsely distributed maps across cortex Local specialization meets global distribution. 8/n

Functional tuning maps of individual dimensions from category-selective areas across cortex.

150

Lenny van Dyck @levandyck.bsky.social · 18/06/2025

Collectively, the dimensions from each area explained activity both within their area but also across broader regions of visual cortex. 7/n

Prediction performance of dimensions from category-selective areas across cortex.

130

Lenny van Dyck @levandyck.bsky.social · 18/06/2025

These dimensions captured diverse information. 🎯 Many aligned with each area’s preferred category (e.g., bodies in EBA) 🧩 Others encoded finer subcategory features (e.g., body parts) 🔄 Some even reflected cross-category distinctions (e.g., food vs. text) 6/n

Interpretability of dimensions in category-selective areas.

140

Lenny van Dyck @levandyck.bsky.social · 18/06/2025

We found that each area encoded multiple interpretable dimensions, consistent across individuals and primarily tuned to high-level semantic content. Strikingly, even the most category-selective voxels showed this multidimensional tuning. 5/n

Consistency of dimensions in category-selective areas.

141

Lenny van Dyck @levandyck.bsky.social · 18/06/2025

To test this, we analyzed fMRI responses to thousands of natural images within classical category-selective areas using a data-driven decomposition approach. Would the resulting organization look modular, continuous, or like something in between? 4/n

Data-driven fMRI voxel decomposition of category-selective areas.

140