Abstract
Covert visual attention has been shown repeatedly to enhance performance on tasks involving the features and spatial locations to which it is deployed. Many neural correlates of covert attention have been found, but given the complexity of the visual system, connecting these neural effects to performance changes is challenging. Here, we use a deep convolutional neural network as a large-scale model of the visual system to test the effects of applying attention-like neural changes. Particularly, we explore variants of the feature similarity gain model (FSGM) of attention—which relates a cell’s tuning to its attentional modulation. We show that neural modulation of the type and magnitude observed experimentally can lead to performance changes of the type and magnitude observed experimentally. Furthermore, performance enhancements from attention occur for a diversity of tasks: high level object category detection and classification, low level orientation detection, and cross-modal color classification of an attended orientation. Utilizing the full observability of the model we also determine how activity should change to best enhance performance and how activity changes propagate through the network. Through this we find that, for attention applied at certain layers, modulating activity according to tuning performs as well as attentional modulations determined by backpropagation. At other layers, attention applied according to tuning does not successfully propagate through the network, and has a weaker impact on performance than attention determined by backpropagation. This thus highlights a discrepancy between neural tuning and function.
