CNN-based search model fails to account for human attention guidance by simple visual features

Area of Science:

• Cognitive Science
• Computer Vision
• Neuroscience

Background:

• Convolutional Neural Networks (CNNs) are increasingly used to model visual attention.
• Zhang et al. proposed a CNN model for attention guidance based on learned visual features for object classification.

Purpose of the Study:

• To adapt the CNN model for visual search experiments and evaluate its performance against human attention guidance.
• To identify limitations of the CNN model in replicating human visual search behavior.

Main Methods:

• Adapted Zhang et al.'s CNN model for simulating feature and conjunction visual search experiments.
• Measured model performance using accuracy as the primary metric.
• Compared model predictions with human performance data.

Main Results:

• The CNN-based model significantly underestimated human attention guidance for simple visual features.
• Performance improvements were suggested by using target-distractor differences or lower-layer feature maps.
• The model failed to reproduce qualitative regularities observed in human visual search.

Conclusions:

• Standard CNNs trained for image classification lack the necessary medium- and high-level features for human-like attention guidance.
• Further model development is needed to incorporate more sophisticated feature representations for accurate visual search modeling.