Monday, May 18, 2026, 2:00 – 5:30 pm, Blue Heron

Organizers: Luca Kämmer, Josefine Zerbe, Martin Hebart
Speakers: Luca Kämmer, Josefine Zerbe, Martin Hebart

The introduction of large, condition-rich datasets (such as BOLD5000, NSD, and THINGS) has led to substantial advances in our understanding of visual and semantic representations in the human brain. Despite these advances, it has recently been shown that these datasets may not cover the whole breadth of visual experience (Roth et al., 2025) and may not generalize out of the training distribution. This has led to two challenges. First, for many findings it is currently unknown to what degree these results would replicate in other datasets or whether these findings are dataset-specific. Second, for most findings, the degree to which results will generalize out-of-distribution is unknown.

To address these challenges, we are launching Re:Vision, a community driven initiative to systematically replicate high-impact findings from condition-rich visual fMRI research and generalize findings beyond the training distribution. To this end, we will use a new large-scale fMRI dataset in which 5 participants went through 47 fMRI sessions including >30 image viewing sessions, looking at ~7,000 images each (Zerbe et al., in prep.). This dataset has been shown to sample the world of visual stimuli much broader than existing datasets, thus offering a higher potential for generalizable findings. Participating researchers will use this dataset to replicate relevant findings, address their generalizability, and write a short report of their findings, which will serve as the basis for a joint summary paper on the state of replicability in visual neuroscience. Additionally, prize money will be awarded to submissions of especially relevant replication and generalization attempts.

This VSS satellite is planned as a kickoff event for the Re:Vision initiative, which will run for several months. The target audience of the satellite consist of (1) researchers interested in existing challenges with large-scale datasets and how they can be addressed, (2) attendees who are interested in learning more about the new dataset and how to use it effectively, and (3) researchers who would like to participate in the initiative, either by having their own work replicated or by actively replicating work.  In the first half of this satellite event, attendees will be provided with general information about the dataset and the structure of the initiative. The second half of the event will be structured like a hackathon to give participants the opportunity to get a first look at our dataset and potentially start with their replication attempts. The organizers will be present to answer questions and help the participants get acquainted with the dataset. Attendance is welcome regardless of whether attendees ultimately participate in the initiative. Attendees are welcome to reach out to authors whose work they intend to replicate already prior to the satellite event.

We additionally would like to point out another VSS Satellite event called “In-silico replications and hypothesis testing for model benchmarking” that will complement the Re:Vision initiative.