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..
$i$MIND: Insightful Multi-subject Invariant Neural Decoding
Authors: Zixiang Yin, Jiarui Li, Zhengming Ding
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that i MIND achieves state-of-the-art decoding performance with minimal scalability limitations. Furthermore, i MIND empirically generates voxel-object activation fingerprints that reveal object-specific neural patterns and enable investigation of subject-specific variations in attention to identical stimuli. |
| Researcher Affiliation | Academia | Zixiang Yin, Jiarui Li, Zhengming Ding Department of Computer Science, Tulane University EMAIL |
| Pseudocode | No | The paper includes mathematical formulations and descriptions of the method, but no explicitly labeled 'Pseudocode' or 'Algorithm' blocks or figures. |
| Open Source Code | No | All details necessary for reproducing the experimental results are fully described in either the main paper or appendix. The code will be released upon acceptance. |
| Open Datasets | Yes | Dataset. We evaluate our i MIND framework using the Natural Scenes Dataset (NSD) [2], a comprehensive, publicly available f MRI dataset capturing brain responses from 8 human subjects viewing natural scenes from MS-COCO [23]. |
| Dataset Splits | Yes | Among these, neural recordings corresponding to 1, 000 images viewed by all subjects are allocated to the testing set, comprising a total of 21, 118 test trials. The remaining neural recordings, corresponding to images viewed exclusively by individual subjects, are included in the training set, resulting in 191, 882 training trials. Both training and testing trials are standardized voxel-wise using the mean and standard deviation calculated from the training set. Since each image is presented to a subject three times, we average the f MRI responses across repetitions for each image within each subject. This results in 69, 566 training samples and 7, 674 testing samples, allowing us to train and evaluate a single multi-subject model across all eight subjects. |
| Hardware Specification | Yes | All experiments are conducted on two Nvidia RTX 6000 Ada GPUs, with the first stage taking approximately 1.5 hours and the second stage around 2 hours to complete. |
| Software Dependencies | No | The paper mentions 'Adam W optimizer' and that 'The CLIP visual encoder we used is clip-vit-base-patch16 released by Open AI', but does not specify version numbers for general software dependencies like Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | In this stage, we choose a patch size of 64 voxels with a masking ratio of 0.75. The encoder has a hidden dimension of 768 and consists of 12 layers of 6-head self-attention, while the decoder has a hidden dimension of 512 and 8 layers of 8-head self-attention. We set the object neural space dimension dobj = 700 and choose a 4-head cross-attention module for f MRI-vision feature interactions. The CLIP visual encoder we used is clip-vit-base-patch16 released by Open AI, which remains frozen at all stages of the proposed framework. A trade-off parameter λ of 0.1 is set by default to enforce the orthonormal constraint of the learnable basis B. During this stage, all parameters are optimized end-to-end for subject and object classification tasks. For either stage, we train the model for 100 epochs, including 10 warm-up epochs. The learning rate is initialized at 7.5 × 10−4 and terminated at zero adjusted dynamically by a cosine scheduler. The batch size is set to 200. Optimization is performed via the Adam W optimizer with a weight decay of 0.05. |