Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Dimensionality Mismatch Between Brains and Artificial Neural Networks
Authors: Santiago Galella, Maren Wehrheim, Matthias Kaschube
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this work, we systematically quantify and compare the linear and nonlinear dimensionality of human brain activity (f MRI) and artificial neural networks (ANNs) during natural image viewing. ... We compute both linear (effective) and nonlinear (intrinsic) dimensionality across the human visual hierarchy... We find that in the human ventral stream, both effective and intrinsic dimensionality increase along the hierarchy... |
| Researcher Affiliation | Academia | Santiago Galella FIAS & Institute of Computer Science Goethe University Frankfurt EMAIL Maren Wehrheim Mila & Department of Biology York University EMAIL Matthias Kaschube* FIAS & Institute of Computer Science Goethe University Frankfurt EMAIL |
| Pseudocode | No | The paper does not contain pseudocode or algorithm blocks; methods are described in narrative text and mathematical formulas. |
| Open Source Code | Yes | All code and scripts necessary to reproduce the experiments and figures are publicly available at https://github.com/sgalella/dimensionality-brains-models. |
| Open Datasets | Yes | We used the THINGS f MRI dataset [34] to study dimensionality of brain activity while viewing natural images. ... Additionally, we use the BOLD5000 dataset, which contains f MRI data... [7]. The data analyzed in this study is publicly available. |
| Dataset Splits | Yes | All decoding analyses were performed using leave-one-session-out cross-validation. ... The plots show the mean and standard deviation across 10 train-test splits (80/20), each using a subset of 50 averaged feature maps. |
| Hardware Specification | No | The code is designed to run on a local multicore machine (e.g., an 8-core setup) and supports parallel computing to accelerate processing. While running the analysis on a computing cluster can significantly reduce runtime, most of analyses can still be completed within a few hours on a standard local machine. |
| Software Dependencies | No | We use the thingsvision implementation of Ecoset-trained models [56]. |
| Experiment Setup | Yes | Here, we set k to be 20. (for ID) For each vertex, we selected a neighborhood of the 50 closest vertices (including the vertex itself). (for searchlight) For decoding categories we use logistic regression, and for decoding stimuli, components and concepts we use ridge regression. (for decoding models) The error bars represent the standard deviation across 100 bootstrapped samples from the results of 100 randomly selected vertices per region. (statistical details) The plots show the mean and standard deviation across 10 train-test splits (80/20), each using a subset of 50 averaged feature maps. (statistical details) |