An AI model trained to learn changing patterns in resting-state brain scans reconstructed held-out signals from both UK Biobank and Human Connectome Project data, with reported R2 values as high as 0.987. It also produced strong displayed results on tests of age, sex, cognitive traits, diagnosis and amyloid status. The work is a preprint evaluating MnemoDyn, a dynamical-systems model trained on roughly 40,000 resting-state fMRI sequences.
These are benchmark results, not clinical validation. The evaluation used the tested resting-state fMRI datasets and tasks, and the report gives means and standard deviations in downstream tables without confidence intervals or formal significance tests.
Reconstruction was strong across datasets
On validation reconstruction, the UK Biobank-pretrained model reached an R2 of 0.985 on UK Biobank data and 0.934 on HCP data. The HCP-pretrained model reached 0.969 on UK Biobank data and 0.987 on HCP data. R2 is a score indicating how closely the reconstructed signal tracks the original; the reported mean-squared-error values ranged from 1.86e-9 to 2.36e-5 across these comparisons.
The masked MnemoDyn variant reported 96.12% accuracy and an F1 score of 95.98% for distinguishing cognitively normal participants from those with mild cognitive impairment in ADNI. For amyloid-status classification, it reported 95.27% accuracy and 95.61% F1. On UK Biobank tasks, its age mean-squared error was 0.44, while sex-classification accuracy and F1 were 88.40% and 88.27%.
On the 656-participant HCP-Aging dataset, the same masked variant reported an age mean-squared error of 0.90, sex-classification accuracy of 83.10% and F1 of 82.77%. It also reported mean-squared errors of 0.90 for neuroticism prediction and 0.60 for a flanker-task measure.
Transfer held up unevenly when the data changed
With HCP pretraining, the masked MnemoDyn-JEPA variant reached 89.87% accuracy and an F1 score of 89.74% on ADNI normal-cognition-versus-mild-cognitive-impairment classification. On the ADNI amyloid-status task, it reported 93.87% accuracy and 93.48% F1.
In tests without adaptation, UK Biobank pretraining reconstructed HBN data at R2 0.98 and HCP-Aging data at R2 0.96. A prediction setup transferred from HCP-Aging to HBN returned a sex accuracy of 0.6393, a sex F1 score of 0.6390 and an age mean-squared error of 0.8516. The analysis describes age prediction as more limited than reconstruction in these zero-shot tests.
On HBN, the masked MnemoDyn row reported an age mean-squared error of 0.80, compared with 1.02 for Brain-JEPA. The base MnemoDyn row reported sex accuracy of 82.37% and F1 of 82.19%, compared with 58.52% and 29.12% for Brain-JEPA. On ADHD-200 ADHD/TDC classification, the masked MnemoDyn-JEPA variant reached 54.70% accuracy and 54.55% F1, versus 45.7% and 46.5% for Brain-JEPA.
On NKIR sex classification, MnemoDyn-Mask reported 88.37% accuracy and 88.36% F1, compared with 66.52% and 63.97% for Brain-JEPA. On ABIDE Autism/TSC classification, MnemoDyn reported 60.32% accuracy and 59.26% F1, while its masked version reported 58.93% and 58.57%. A reliable Brain-JEPA baseline was not available for the ABIDE comparison.
How the model was trained
MnemoDyn uses an operator-based architecture to model activity changing over time across parcellated brain regions at multiple resolutions. The researchers converted raw BIDS fMRI volumes into a common 91,282-grayordinate representation, then grouped them into 450 regional time series. The data were robustly normalized using the training set’s median and interquartile range.
Pretraining evaluated denoising, masked autoencoding and a Brain-JEPA-style objective. In the masked versions, 70% of the observed signal was hidden across time and space and the model was trained to recover it. For downstream tasks, the backbone was frozen and an MLP head was trained using mean-squared error for regression or cross-entropy for classification.
The displayed scaling tests showed lower age error and higher sex-classification scores with greater model depth. Moving from one to two operator blocks changed age mean-squared error from 0.914 to 0.903 and sex accuracy from 0.768 to 0.801. Increasing pretraining from 1,000 to 10,000 to 20,000 subjects was accompanied by age errors of 1.023, 1.021 and 1.012, while sex accuracy rose from 0.706 to 0.753 to 0.793.
An ablation that removed the wavelet component performed worse on every listed task. Normal-cognition-versus-mild-cognitive-impairment accuracy fell from 96.12% with the wavelet-containing masked model to 75.12% without it, while amyloid-status accuracy fell from 95.27% to 72.19%. Age error rose from 0.44 to 0.67, and sex accuracy fell from 88.40% to 76.16%.
Promising benchmark, limited evidence
The study used archival, de-identified, parcellated resting-state fMRI data. It did not test voxel-level or multimodal inputs, longitudinal scans, or transfer to EEG or MEG. The learned structure was not validated as a neurophysiological mechanism, and the evaluation did not establish clinical utility or support direct medical decision-making. Participant counts and detailed splits for HBN, ADHD-200, ABIDE and NKIR were not reported in the supplied analysis.
The paper says its architecture, training and evaluation details are provided in the text and appendix. It also says that code, preprocessing scripts and pretrained weights are planned for public release through Hugging Face upon acceptance. Partial support came from a contract to UW-Madison under the DARPA Strengthen program, and UK Biobank data were used under Application Number 96217.
The supplied metadata classifies the work as a preprint, while its front matter states that it was published as a conference paper at ICLR 2026.
Paper data and sources
Original title: MnemoDyn: Learning Resting State Dynamics from 40K FMRI sequences
Authors: Sourav Pal, Viet Luong, Hoseok Lee et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text