Forward:
data -> progressively corrupted states -> noise

Reverse:
noise -> learned probabilistic transitions -> data