Emergent neuronal mechanisms mediating covert attention in convolutional neural networks

Abstract

Covert visual attention allows the brain to select different regions of the visual world without eye movements. Predictive cues of a target location orient covert attention and improve perceptual performance. In most computational models, researchers explicitly incorporate an attentional mechanism that alters processing at the attended location (gain, noise reduction, divisive normalization, biased competition, Bayesian priors). Here, we assess the emergent neuronal mechanisms of Convolutional Neural Networks (CNNs) that exhibit behavioral signatures of covert attention, despite lacking a built-in attention mechanism. We use neuroscience-inspired approaches to analyze 1.8 M units of CNNs trained on the cueing paradigm. Consistent with neurophysiology, we show early layers with retinotopic neurons separately tuned to the target or cue, and later layers with neurons with joint tuning and increased cue influence on target responses. CNN computational stages mirror a Bayesian ideal observer (BIO), but with more gradual transitions. The cue influences the target sensitivity through four mechanisms. A BIO-like cue-weighted location summation, and three mechanisms absent in the BIO: an opponency across locations, a summation/opponency location combination, and interaction with the thresholding Rectified Linear Unit. Reanalyses of mice’s superior colliculus neuronal activity during a cueing task show CNN-predicted but previously unreported cue-inhibitory, location-summation, and location-opponent cells in addition to the commonly reported cue (attention) excitatory cells. The single-unit CNN analysis approach establishes a likely system-wide characterization mediating covert attention and a framework to identify new neuron types and emergent computational mechanisms contributing to perceptual behavior.