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..
Orthogonal Contrastive Learning for Multi-Representation fMRI Analysis
Authors: Tony Yousefnezhad
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Across extensive experiments on multi-subject and multi-site f MRI benchmarks, OCL consistently outperforms state-of-the-art alignment and analysis methods in both representation quality and downstream classification accuracy. |
| Researcher Affiliation | Industry | Tony Muhammad Yousefnezhad1,2, 1Learning By Machine 2Information Management, National Bank of Canada Edmonton AB Canada EMAIL |
| Pseudocode | Yes | We provide the pseudocode for the proposed OCL algorithm in the supplementary material. |
| Open Source Code | Yes | Our proposed OCL algorithm is available on Git Hub 9. |
| Open Datasets | Yes | All datasets are publicly available (via Open NEURO 4, except CMU 5) and were preprocessed with our GUI-based toolbox called easy f MRI 6 and FSL 6.0.15 7, including spatial normalization, smoothing, anatomical alignment; for those alignment techniques that require it, temporal realignment was also applied (see Section 4.1). |
| Dataset Splits | Yes | We use a leave-one-subject-out nested cross-validation: in each outer fold, one subject is held out for testing; within each, another subject serves as validation (inner fold), and the rest form the training set. |
| Hardware Specification | Yes | All experiments were run on two PCs with the specifications listed in the Footnote 8. ... 8 OS: Fedora 42, Python: 3.11.9, Py Torch: 2.6, CUDA: 12.6; Connection: 2 40Gb E CX314A Mellanox (PC1) CPU: AMD EPYC 7551P (64 cores), RAM: 256G GPU:2 NVIDIA 4060Ti 16G; (PC2) CPU: AMD Threadripper 2990WX (64 cores), RAM: 128G, GPU:2 NVIDIA 4060Ti 16G. |
| Software Dependencies | Yes | 8 OS: Fedora 42, Python: 3.11.9, Py Torch: 2.6, CUDA: 12.6; Connection: 2 40Gb E CX314A Mellanox (PC1) CPU: AMD EPYC 7551P (64 cores), RAM: 256G GPU:2 NVIDIA 4060Ti 16G; (PC2) CPU: AMD Threadripper 2990WX (64 cores), RAM: 128G, GPU:2 NVIDIA 4060Ti 16G. ... and FSL 6.0.15 7 |
| Experiment Setup | Yes | We set the embedding dimension to d = 256, employ Encoding Transformer with 16 attention heads, and use N = 32 network layers. ... We train (and pretrain) OCL for up to ฯ = 1000 iterations with automatic early stopping based on validation loss using Adam optimizer [42]. ... We perform grid search over the key OCL hyperparameters temperature ฯ {0.01, 0.1, 0.5, 0.9, 0.99}, margin ยต {0.1, 0.2, 0.5, 0.8, 0.9}, between-class weight ฮป {0.1, 0.2, 0.3, 0.4, 0.5}, learning rate ฮท {0.1, 0.2, 0.3, 0.4}, and quantization granularity w {0.9, 1.0, 1.1, 1.2} and select the combination that maximizes performance on the validation set. |