Abstract
The detection of a sequence of events requires memory. The detection of visual motion is a well-studied example; there the memory allows the comparison of current with earlier visual input. This comparison results in an estimate of direction and speed of motion. The dominant model of motion detection in primates—the motion energy model—assumes that this memory resides in subclasses of cells with slower temporal dynamics. It is not clear, however, how such slow dynamics could arise. We used extracellularly recorded responses of neurons in the macaque middle temporal area to train an artificial neural network with recurrent connectivity. The trained network successfully reproduced the population response, and had many properties also found in the visual cortex (e.g., Gabor-like receptive fields, a hierarchy of simple and complex cells, motion opponency). When probed with reverse-correlation methods, the network's response was very similar to that of a feed-forward motion energy model, even though recurrent feedback is an essential part of its architecture. These findings show that a strongly recurrent network can masquerade as a feed-forward network. Moreover, they suggest a conceptually novel role for recurrent network connectivity: the creation of flexible temporal delays to implement short term memory and compute velocity.