Gain, not concomitant changes in spatial receptive field properties, improves task performance in a neural network attention model

Abstract

Attention allows us to focus sensory processing on behaviorally relevant aspects of the visual world. One potential mechanism of attention is a change in the gain of sensory responses. However, changing gain at early stages could have multiple downstream consequences for visual processing. Which, if any, of these effects can account for the benefits of attention for detection and discrimination? Using a model of primate visual cortex we document how a Gaussian-shaped gain modulation results in changes to spatial tuning properties. Forcing the model to use only these changes failed to produce any benefit in task performance. Instead, we found that gain alone was both necessary and sufficient to explain category detection and discrimination during attention. Our results show how gain can give rise to changes in receptive fields which are not necessary for enhancing task performance.

Bridging Neurons and Behavior in a Convolutional Neural Network with Emergent Human-like Covert Attention

“…We did not add biological constraints such as divisive normalization, which has been used to model neuronal properties of attention 16 , or explicit biological attention 19,20 , or feedback connections 31,63,65,66 , into the networks. Our analysis pipeline provides a testing bed to understand the importance of additional architectural constraints.…”

“…Cell electrophysiological studies with spatial cues have been used to show that attention can increase the amplitude 4 , but not the width of the orientation tuning curve at the cued location 4 , reduce inter-neuronal correlations 3 , and change the contrast response functions 18 of neurons. Recently, studies have incorporated neurally-inspired attention mechanisms into convolutional neural networks 19,20 to assess the impact on model perceptual accuracy.…”

“…The network has no explicit knowledge about the cue nor its predictive probability. We also do not explicitly incorporate any attention mechanism (e.g., gain 20 , prior 17,34 , feature similarity 19 ) into the network to act differently at the cued location. We compare the properties of 180K neuronal units at the different layers of ten trained CNNs (1.8 M neurons) to the theoretical computational stages of an image-computable Bayesian ideal observer (BIO) that integrates visual information of the cue, the target, and across locations to maximize perceptual performance detecting the target.…”

Spatial attention alters visual cortical representation during target anticipation

“…It is unknown to what degree similar circuits are involved during target anticipation, which is purely top-down, and attentional modulation of stimulus evoked responses, in which top-down and stimulus-driven effects interact. Moreover, studies of attention during stimulus presentation have led to conflicting conclusions about whether spatial attention primarily works by changing neural position tuning 19 or response amplitude 33 . The finding in macaque V4 that distinct neuronal populations are modulated by pre- and post-stimulus attentional signals 34 suggests that human fMRI evidence on the attentional modulation of responses to an attended stimulus may not generalize to the anticipatory effects of attention.…”

Modulatory feedback determines attentional object segmentation in a model of the ventral stream

“…At first sight, many of these neurophysiological observations may seem disconnected, because previous studies focused on only one or a few of these feedback influences (Carandini et al, 2005), or used models with connectivity that was largely handcrafted (Craft et al, 2007; Deco and Rolls, 2004; Hamker, 2005; Van Der Velde and De Kamps, 2001). We therefore sought to integrate previous findings into a coherent framework, capitalizing on the recent advances in the development of artificial neural networks (ANNs) (Echeveste et al, 2020; Fox et al, 2023; Serre, 2019).…”