encoding many features as nearly orthogonal directions in a lower-dimensional space. This is analogous to compressed sensing: if features are sparse (rarely active simultaneously), they can share dimensions.