Right now models are training on NSDGeneral, which is a good selection of voxels for vision, but is naively selected for imagery. Presumably there is a more optimal combination of more or less voxels (probably less) that include areas from outside visual cortex that might be useful for decoding imagery. We should experiment with various voxel selection strategies such as:
- SNR thresholding
- Finding cross-validated masks of voxels with strong correspondence between vision and imagery
- ROI analysis of the contributions of different brain areas to imagery signals
Task will be to mask the betas being used for training, and explore generalization performance on NSD-Imagery.
Right now models are training on NSDGeneral, which is a good selection of voxels for vision, but is naively selected for imagery. Presumably there is a more optimal combination of more or less voxels (probably less) that include areas from outside visual cortex that might be useful for decoding imagery. We should experiment with various voxel selection strategies such as:
Task will be to mask the betas being used for training, and explore generalization performance on NSD-Imagery.