I ran a SpaceNetClassifier on brain activity preceding movement to try to predict upcoming actions (A or B). As expected, the fitted classifier returns coefficient maps that are sparse and blobby (and even blobby in a priori plausible regions!). However, I noticed there is variation across voxels in both the sign and magnitude of the coefficients. Is there any intuitive way to interpret the sign or magnitude of the coefficients? Or should all non-zero coefficients be treated equally, interpreted simply as indicators that a voxel helped the classifier predict the class label (hence, that the voxel is predictive of or has information about the class label). Thanks!