With todays networks and training corpus sizes and compute resources available, it's generally better to let the network learn the best transform rather than tr

With todays networks and training corpus sizes and compute resources available, it's generally better to let the network learn the best transform rather than try to guess what it needs and do it manually first.. unless you have analyzed the problem thoroughly (like I guess they did in this paper). You have several dozens of layers and if you need to "waste" a few low-level layers to learn some transform, the increased training time might offset the man-hours and risk of your meat-based feature-designers to throw something sub-optimal in front of the network..