The Effectiveness of Memory Replay in Large Scale Continual Learning