Abstract
Artificial neural networks are inspired by neurobiology, implementing distributed computation by representing and manipulating data in the same structure. Deep convolutional networks (CNNs) have shown exceptional performance in a range of visual processing tasks, and an analogy can be made between the structure of CNNs and the visual processing in mammals. In this paper we investigate this analogy within Marr and Poggio's four-level model of information processing, identifying similarities and differences between artificial and real networks for general purpose visual processing. While the overall structure of the real and artificial networks are quite similar, several key differences suggest avenues for further research. In particular, the prevalence of lateral inhibition and recurrent connections in biological networks suggests new architectures that may be useful in future CNN research. We also note that in some respects current CNN architectures are more similar to the visual system of mice than of humans, although whether this is important for specific tasks is an open question.