Superstitious Perception: Comparing Perceptual Prediction by Humans and Neural Networks

Area of Science:

• Cognitive Psychology
• Computational Neuroscience
• Artificial Intelligence

Background:

• Convolutional neural networks (CNNs) offer novel models for human vision.
• Previous comparisons between CNNs and human vision focused on internal representations, neglecting behavioral aspects.
• Cognitive strategy's influence on the human-CNN behavioral similarity remains underexplored.

Purpose of the Study:

• To investigate how cognitive strategy affects the similarity between CNN output and human behavior.
• To compare human performance in a superstitious perception task under active versus passive attention with CNN performance.

Main Methods:

• Human participants performed a superstitious perception task (target detection in white noise) under active or passive attentional strategies.
• CNN decision criteria were adjusted to match human false alarm rates in each strategy condition.
• Behavioral responses of humans and CNNs were compared across conditions.

Main Results:

• Human response patterns significantly differed based on attentional strategy.
• False alarms in the passive condition correlated with target presence evidence, unlike in the active condition.
• CNN responses more closely resembled human behavior in the passive attention condition than in the active condition.

Conclusions:

• CNNs better mimic human behavior when central executive functions are disengaged (passive attention).
• Superstitious perception involves dissociable components: detecting weak signals (passive) and ignoring noise for internal signals (active).
• Modeling the full spectrum of human responsiveness, including cognitive strategy, is crucial for understanding perception.