DeepSITH: Efficient Learning via Decomposition of What and When Across Time Scales