Event:

Humans and other primates are able to parse the complex flow of visual information into meaningful objects that can be recognized and acted upon. Recently, we have witnessed two major shifts in the study of the neural basis of this and other aspects of visual perception. First, more and more studies go for rodents instead of monkeys in order to unravel the details of the neural circuitry. Second, computer models such as deep convolutional neural networks (convNets) have been proposed that can reach unprecedented performance levels in object recognition. Here I will connect the dots between these different approaches, and report on correspondences and discrepancies between how humans, rodents, and convNets represent objects. In the rat "ventral" cortical pathway, we find selectivity for basic visual features as well as shape but without any evidence for selectivity for object categories. Rat behavior and neural representations are modeled by middle layers in a convNet architecture. In the human ventral object vision pathway, we find a so-called feature-based categorical code that shows a mixture of selectivity for basic visual features, more high-level shape features, as well as for object categories. Similar representations emerge in the highest layers of convNets. Nevertheless, human and convNet representations diverge in a stimulus set in which appearance and object identity are dissociated. I conclude that the nonhuman models are very useful to reveal and eventually manipulate the ways in which the human brain is special.