Deep hierarchical sensory processing accounts for effects of arousal state on perceptual decision-making

An organism's level of arousal strongly affects task performance. Yet, what level of arousal is optimal for performance depends on task difficulty. For easy tasks, performance is best at higher arousal levels, whereas arousal levels show an inverted-U-shaped relationship with performance for difficult tasks, with best performance at medium arousal levels. This interaction between arousal and task difficulty is known as the Yerkes-Dodson effect (1908) and is thought to reflect sensory decision-making in the locus coeruleus and associated widespread release of noradrenaline. Yet, this account does not explain why perceptual performance decays with high levels of arousal in difficult, but not in simple tasks. Recent studies suggest that arousal may also affect performance by modulating sensory processes. Here, we augment a deep convolutional neural network (DCNN) with a global gain mechanism to mimic the effects of arousal on sensory processing. This allowed us to reproduce the Yerkes-Dodson effect in the model's performance. Investigating our network furthermore revealed that for easy tasks, early network features contained most task-relevant information during high global gain states, resulting in model performance on easy tasks being best at high global gain states. In contrast, later layers featured most information at medium global gain states and were essential for performance on challenging tasks. Our results therefore establish a novel account of the Yerkes-Dodson effect, where the interaction between arousal state and task difficulty directly results from an interaction between arousal states and hierarchical sensory processing.