We propose an alternative approach to compensation of intermodal interactions in few-mode optical fibers by means of digital backpropagation. Instead of solving the inverse generalized multimode nonlinear Schrödinger equation, we accomplish backpropagation of the multimode signals with help of their near-field intensity distributions captured by a camera. We demonstrate that this task can successfully be handled by a deep neural network and provide a proof of concept by training an autoencoder for backpropagation of six linearly polarized (LP) modes of a step-index fiber.