There are successful deep networks with pure feed forward non-convolution layers. There are also deep layers of other exotic non-convolutions flavors like LSTMs and GRUs, particularly useful for sequence-to-sequence tasks like machine translation.