The paper describes a method to decode natural images from retinal-like activities, using convolutional neural networks. The retinal-like activities are generated by a constructed lattice of linear-nonlinear-Poisson models (separately fitted to RGC responses to natural scenes in a macaque retina preparation) in response to natural static images. After a simple linear decoding of the images from the retinal-like activities, a convolutional neural network further improves on the reconstruction of the original natural images. The paper is clearly written and the results seem sound. A few comments to clarify the motivation, assumptions and impact of the work: -the method proposed is compared to a linear decoder and shown to perform substantially better. However, the performance of the decoding stage will most likely depend on the performance of the encoding stage.