These results support an adaptive recurrence theory:
When feedforward suffices → recognition stays ventrally.
When it fails → rapid, transient, cortex-wide recurrence (fronto-parietal feedback) to resolve recognition.
In other words: recurrent loops might be our brain's visual debugger 👀 (fin)

EEG results:
(Cross-)decoding revealed three stages:
- Early shared feedforward stage (90–140 ms)
- Divergence between challenge/control (140–190 ms)
- Late convergence (190ms post) with delayed peak for challenge images
-> extra processing time for challenge images suggests recurrence at work (5/n)

Using RSA across ROIs, we built network similarity matrices quantifying the network configuration across brain regions.
Challenge vs. control comparisons showed:
High similarity in IT/LOC but dissimilarity in dlPFC
-> dLPFC might reconfigure visual representations when feedforward isnt enough (4/n)

fMRI results:
For challenge images, we found significant object decoding not just in ventral visual cortex, but also in frontal (vlPFC, dlPFC) and parietal regions. This suggests cortex-wide representation of object identity when feedforward processing alone isn’t enough. (3/n)

Following the approach from @kohitij.bsky.social, we compared “control” images (solved equally well by CNNs & humans) vs. “challenge” images (where humans succeed but CNNs fail)

We recorded EEG & fMRI responses while participants viewed these images. (2/n)