Do our findings replicate?
A replication initiative for image-related neuroscience
Do our findings replicate?
A replication initiative for image-related neuroscience
Many of the recent advancements in our understanding of visual and semantic representations in the human brain are based on a few, condition-rich datasets, which only capture a fraction of the visual world. Our goal is to find out if these findings replicate in our new LAION-fMRI dataset which is optimized to broadly sample the visual world.
Existing fMRI image datasets represent a small, biased slice of visual experience, which risks overfitting. It is unclear whether findings based on them will replicate in a more broadly sampled dataset.
Our main goal is to replicate impactful findings from image-related neuroscience on a broader fMRI dataset.
Our secondary goal is to specifically test if these findings generalize to the full image distribution of LAION-fMRI, while removing potential overlap.
Contains fMRI responses to images from LAION-natural (sampled from the internet), MSCOCO (natural scenes), THINGS (object images), and out-of-distribution images. See here the semantic coverage in CLIP embedding space
the effective dimensionality of NSD or THINGS in CLIP embedding space, which measures how many independent axes of variation a stimulus set covers.
LAION-fMRI contains densely-sampled 7T fMRI data with 30 image-viewing sessions, extensive retinotopy, resting-state, and diffusion imaging.
Cortical flatmaps showing retinotopic and category-selective ROI definitions used in the dataset. Surfaces reconstructed with FreeSurfer, flatmaps generated via pycortex.
Awarded to the team that demonstrates the most rigorous and well-executed replication of a published finding.
Awarded to the team that provides the most compelling evidence for (or against) generalization to the LAION-fMRI dataset.
Awarded to original authors of replicated papers that cooperate with and support the replication team.
These cash prizes will be awarded by the re:vision board after the final submission of the report. They will be split between the members of the winning team.
The winners of the prizes are chosen based on how rigorous the submission was, how much work was put into it, and how well it was presented in the report. They are independent of whether the replication / generalization was successful.
Submit proposal
Submit a short proposal on how to replicate / generalize the study you signed up for.
Due July 15th 2026
Conduct replication
Replicate the main findings using LAION-fMRI. You will not have to collect any data.
Submit report
Submit a report on your replication.
Due January 15th 2027
You will be part of the consortium paper and have a chance to win one of our prizes.
Initiative kickoff at our VSS satellite + dataset release. Includes a 3.5-hour hands-on session with Q&A and dataset tutorials.
Deadline for participants to submit a short proposal.
Deadline for participants to submit the report on their replication attempt.
1st and only round of reviews from the re:vision board.
Deadline for the final revised reports.
The organizers and board will synthesize the reports we receive to write a summary paper. Every participant that submitted a valid report will be on the paper.
Practical details on how to participate in our initiative. If your question is not covered, you can contact us via the e-mail address below.
Anyone with an interest in imaging neuroscience. We advise that at least one member of your replication team has experience with fMRI analysis.
No. All fMRI data, preprocessed betas, retinotopic maps, and image annotations are provided through our Python package. You may need to generate some meta data using a model depending on the study you replicate.
The initiative board evaluates reports for methodological clarity and rigor of replication. All valid replications will be included in the final summary paper, irrespective of the result of the replication.
Any published result generated with a condition-rich dataset. We provide a list with suggested studies but you are welcome to choose a different study as long as at least some of its main findings are in principle replicable using LAION-fMRI.
You cannot replicate your own findings but if you have a result you would like to see replicated we are happy to connect you with someone to conduct the replication.
Teams that submit a valid replication are invited to join the re:vision consortium and will be authors on the final paper. They will also have a chance to win one of the prizes.
Yes, we allow up to 2 replication attempts of the same study. So sign up quickly if you have a specific study in mind that you would like to replicate. If you sign up for a study we will reserve it for you for 2 weeks so you have time to write your proposal.
We allow 1-3 researchers to conduct a replication together. All team members will be part of the final paper if the team submits a valid replication to the board. Additionally, you can contact the original authors of the paper to ask for help with specific methods or code.
Contact the original authors for advice on how they generated this data. In many cases models can generate a wide variety of metadata. Contact us if you have specific questions.
Yes. All teams that submit a valid replication report will be invited to join the re:vision consortium and be listed as co-authors on the final summary paper. This applies to every team member (up to 3 people per team).
Negative results are equally valuable and will be included in the consortium paper. The goal of this initiative is to get an honest picture of which findings hold and which do not.
Yes. The initiative is open to anyone with the relevant expertise, regardless of institutional affiliation. We only ask that at least one team member has experience with fMRI data analysis.
We strongly encourage open science practices. Teams are expected to share their analysis code (e.g., on GitHub) as part of their final report. The fMRI dataset itself is already publicly available through our Python package.
Yes. The suggested list covers the most widely-cited studies in the field, but you may choose a different published study as long as its main findings are in principle replicable using the LAION-fMRI dataset.
Organized by researchers at the Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig and the Justus-Liebig-University, Giessen, Germany.
Expert reviewers who will evaluate submissions and provide feedback on replication reports.
For inquiries, contact us at re-vision-initiative@uni-giessen.de
If you wish to stay updated on the re:vision initiative, join our mailing list,
or follow us on social media:
Bluesky
