We propose a novel approach for the referenceless mode decomposition of multimode fibers. A deep neural network with the interaction of a physical model achieves decomposition of 55 modes without pre-training. This paradigm shift is of great importance for space division multiplexing.