Leveraging Spiking Deep Neural Networks to Understand the Neural Mechanisms Underlying Selective Attention

Abstract

Spatial attention enhances sensory processing of goal-relevant information and improves perceptual sensitivity. Yet, the specific neural mechanisms underlying the effects of spatial attention on performance are still contested. Here, we examine different attention mechanisms in spiking deep convolutional neural networks. We directly contrast effects of precision (internal noise suppression) and two different gain modulation mechanisms on performance on a visual search task with complex real-world images. Unlike standard artificial neurons, biological neurons have saturating activation functions, permitting implementation of attentional gain as gain on a neuron's input or on its outgoing connection. We show that modulating the connection is most effective in selectively enhancing information processing by redistributing spiking activity and by introducing additional task-relevant information, as shown by representational similarity analyses. Precision only produced minor attentional effects in performance. Our results, which mirror empirical findings, show that it is possible to adjudicate between attention mechanisms using more biologically realistic models and natural stimuli.

Arousal state affects perceptual decision-making by modulating hierarchical sensory processing in a large-scale visual system model

“…We augmented a DCNN with a global gain mechanism to investigate how arousal state changes in sensory areas may affect performance and relate to the Yerkes-Dodson effect more specifically. To this end, we used a ResNet18-architecture [ 30 ] with a biologically-inspired activation function (see Fig 1C for an illustration, [ 31 , 32 ]). The global gain mechanism targeted all activation functions in the network simultaneously.…”

Arousal state affects perceptual decision-making by modulating hierarchical sensory processing in a large-scale visual system model

“…We augmented a DCNN with a global gain mechanism to investigate how arousal state changes in sensory areas may affect performance and relate to the Yerkes-Dodson effect more specifically. To this end, we used a ResNet18-architecture [31] with a biologically-inspired activation function (see Fig 1C for an illustration, [32, 33]). The global gain mechanism targeted all activation functions in the network simultaneously.…”

“…We tested the DCNN across two search contexts (street scenes and food scenes). These datasets were curated to be challenging for visual search, while reducing other informative features such as the background (for a detailed description, see [33]).…”

“…We applied this manipulation in the context of two different object recognition datasets. Both datasets were curated to be challenging visual search tasks by reducing the amount of informative background differences across categories (for further information see [33]). As can be seen in Fig 4C, this manipulation resulted in differences for the baseline networks (with global gain set to 1), as indexed by AUC, thus showing that the manipulation was effective.…”

“…The average images were constructed from a separate set of images than those used as real images, thereby avoiding overlap in image features. All images were taken from a dataset consisting of street scenes, described here [33].…”