Neural Topographic Factor Analysis for fMRI Data
Authors: Eli Sennesh, Zulqarnain Khan, Yiyu Wang, J Benjamin Hutchinson, Ajay Satpute, Jennifer Dy, Jan-Willem van de Meent
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate NTFA on data from an in-house pilot experiment, as well as two publicly available datasets. We demonstrate that inferring representations for participants and stimuli improves predictive generalization to unseen data when compared to previous topographic methods. We also demonstrate that the inferred latent factor representations are useful for downstream tasks such as multivoxel pattern analysis and functional connectivity. |
| Researcher Affiliation | Academia | Eli Sennesh 1,3, Zulqarnain Khan 2, Yiyu Wang3 Jennifer Dy2, Ajay Satpute3, J. Benjamin Hutchinson4, Jan-Willem van de Meent1 sennesh.e@northeastern.edu, khan.zu@ece.neu.edu, wang.yiyu@northeastern.edu jdy@ece.neu.edu, a.satpute@northeastern.edu, bhutch@uoregon.edu, j.vandemeent@northeastern.edu 1 Khoury College of Computer Sciences, Northeastern University 2 Department of Electrical and Computer Engineering, Northeastern University 3 Department of Psychology, Northeastern University 4 Department of Psychology, University of Oregon |
| Pseudocode | Yes | Algorithm 1 Neural TFA Generative Model |
| Open Source Code | No | Source code submitted with paper and available upon request. |
| Open Datasets | Yes | We analyze and evaluate two publicly available datasets. In the first, participants with major depressive disorder and controls listened to emotionally valenced sounds and music [Lepping et al., 2016]. In the second, participants viewed images of faces, cats, five categories of man-made objects, and scrambled pictures [Haxby et al., 2001]. |
| Dataset Splits | No | As a sanity check, we also compare predictive performance on a validation set of brain images across NTFA and HTFA. We hold out trials by their stimulus-participant pairs, requiring our model to generalize from other trials in which the same stimulus or participant were seen. PCA, the SRM, and TFA cannot recombine representations to predict such novel combinations in this way. To evaluate generalization, we split datasets into training and test sets, ensuring the training set contains at least one trial for each participant p {1, . . . , P} and each stimulus s {1, . . . , S}. To do so, we construct a matrix of (p, s) {1, . . . , P} {1, . . . , S} with participants as rows and stimuli as columns. We then choose all trials along the matrix s diagonals {n : pn mod S = sn} as our test set. All other trials are used as the training set. |
| Hardware Specification | No | This is somewhat fewer than previously reported for HTFA (K = 700) [Manning et al., 2018] owing to GPU memory limitations. |
| Software Dependencies | No | We optimize this objective using black-box methods provided by Probabilistic Torch, a library for deep generative models that extends the Py Torch deep learning framework [Narayanaswamy et al., 2017]. |
| Experiment Setup | Yes | We employ participant and stimulus embeddings with D = 2 in all experiments. For the synthetic dataset, we analyze the data with the same number of factors as were used to generate it, K = 3. For non-simulated data we use K = 100 factors. This is somewhat fewer than previously reported for HTFA (K = 700) [Manning et al., 2018] owing to GPU memory limitations. We report parameter counts for HTFA and NTFA in Table 2, and provide details on network architectures in Appendix A.7. |