friends-s01¶
We first implemented a series of experiments to explore the impact of different parameters on the quality of embeddings in the friends-s01 dataset. We ordered all fMRI runs for each subject by chronological order, and separated the even runs (training data) from the odd runs (validation data). We generated several dynamic parcellations using dypac independently on friends-s01 training and validation datasets. We explored different parameters, both for the number of clusters and the number of states: cluster-20_state-60, cluster-20_state120, cluster-50_state-150, cluster-50_state-300and cluster-300_state-900. We also tested two different levels of spatial smoothing: fwhm=5 and fwhm=8.
The results of these experiments was that cluster-20_state-60 was the most spatially subject-specific, while cluster-300_state-900 provided the best approximation of voxel-level data. This approximation was near-perfect for fwhm=8, and good for fwhm=5. We thus decided to generate final parcellations using the entirety of friends-s01 with fwhm=5, which offers a much better spatial resolution, and two different scales: cluster-20_state-60, as well as cluster-300_state-900.