We can think of these inputs to layer 1 as the name of a song (input from above) and where we are in a song (delayed activity from active columns in the same region). Thus layer 1 carries much of the information we need to predict when a column should be active—the sequence name and where we are within the sequence.