Efficient Learning through Manifold Untangling and Tangling: A Geometric Perspective
Efficient learning can be achieved by a tangling-untangling cycle that maps context-independent representations to context-dependent representations in high-dimensional space, and then collapses the context variables back to the original low-dimensional space for generalization.