Time-of-flight (ToF) imaging has become a widespread technique for depth estimation, allowing affordable off-the-shelf cameras to provide depth maps in real time. However, multipath interference (MPI) resulting from indirect illumination significantly degrades the captured depth. Most previous works have tried to solve this problem by means of complex hardware modifications or costly computations. In this work, we avoid these approaches and propose a new technique to correct errors in depth caused by MPI, which requires no camera modifications and takes just 10 milliseconds per frame. Our observations about the nature of MPI suggest that most of its information is available in image space; this allows us to formulate the depth imaging process as a spatially-varying convolution and use a convolutional neural network to correct MPI errors. Since the input and output data present similar structure, we base our network on an autoencoder, which we train in two stages. First, we use the encoder (convolution filters) to learn a suitable basis to represent MPI-corrupted depth images; then, we train the decoder (deconvolution filters) to correct depth from synthetic scenes, generated by using a physically-based, time-resolved renderer. This approach allows us to tackle a key problem in ToF, the lack of ground-truth data, by using a large-scale captured training set with MPI-corrupted depth to train the encoder, and a smaller synthetic training set with ground truth depth to train the decoder stage of the network. We demonstrate and validate our method on both synthetic and real complex scenarios, using an off-the-shelf ToF camera, and with only the captured, incorrect depth as input.