Plan

 In order to simplify the search problem for the best model, I decided to follow the following approach.

1. Concatenate PC's of activations from a single model's layers and use AutoSklearn to find the best regression parameters. Compute the accuracy for each voxel and distribution for each ROI.

2. Try out different models in this way. Compute which models have the best performance in terms of means and stds of accuracies for each ROI.

Is this too coarse? It is possible that within each ROI there are cells that specialize in different aspects of processing the visual image. However, by concatenating multiple layers as predictors we are reducing the risk of not capturing those aspects. I think this approach has clear benefits in terms of conceptual simplicity. Using different models for different models is too difficult. I want a broad characterization in terms of ROI's. This will give me more time to try out different models.

 It would also be good to fit a ridge regression model with different reg params and visualize the best one to figure out which coefficients are contributing most to the predictions, i.e. which voxels are closest to which predictors. AutoSklearn will probably be more accurate, but for interpretability and understanding Ridge is also necessary.

:)