[thesz]

There are no inserts into larger layers, only merges. Which are long (usually processed in background by separate thread) and that longness justifies training a new net in parallel to merge process.