It has become popular because they found a better way to initialize deep networks (by incrementally fitting auto-encoder layers), where simple gradient-based op

It has become popular because they found a better way to initialize deep networks (by incrementally fitting auto-encoder layers), where simple gradient-based optimizations of yesteryear would not converge from scratch.