Fig. 4. Experimental validation.

a Schematic of the SLM phase inference by the CNN. The trained neural network reconstructs SLM phases from unseen images. The reconstructed phases are then sent through the MMF. Outputs of the fiber are then captured by the camera. The network, which is trained to regenerate input phases that output only the Latin alphabet, is also able to regenerate input phases belonging to other categories (transfer learning). Captured images of (b–e) Latin alphabets, (f–i) digits, (j) a cross and k a heart at the output of the fiber. The fidelity for each transmitted image with respect to the same transmitted image obtained using the transmission matrix of the system (sending ground truth SLM phases (labels), which are calculated by the inverse of the transmission matrix through the fiber, and capturing output images by the camera) is also shown