Image transmission through a multi-mode fiber is a difficult task given the complex interference of light through the fiber that leads to random speckle patterns at the distal end of the fiber. With traditional methods and techniques, it is impractical to reconstruct a high-resolution input image by using the information obtained from the intensity of the corresponding output speckle alone. In this work, we train three Convolutional Neural Networks (CNNs) with input-output couples of a multi-mode fiber and test the learning with images outside the learning set. The three implemented deep learning models have modern UNet, ResNet and VGGNet architectures and are trained with 31,200 grey-scale handwritten letters of the Latin alphabet. After the training, 5,200 images outside the learning set are used for testing and it was shown that the models successfully reconstruct the input images from the output random speckle patterns with average fidelities ranging from 81% to 90%. Our results show the superiority of the ResNet based architecture over UNet and VGGNet in reconstruction accuracy, achieving up to 97% fidelity in a short amount of time. This can be attributed to the success of the ResNet architecture in learning non-linear systems compared to its counterparts. We believe that the implementation of machine learning techniques to imaging, along with its contributions to biophysics, can reshape the telecommunication industry and thus will be a cornerstone in future optics and photonics studies.