Modeling attention and binding in the brain through bidirectional recurrent gating

Area of Science:

• Cognitive Neuroscience
• Computational Neuroscience
• Artificial Intelligence

Background:

• Attention is crucial for cognition, enabling information selection, feature binding, and behavior guidance.
• Existing models lack a unified framework for diverse attentional phenomena like spatial, feature-based, and object-based attention.
• A neurally plausible computational model is needed to integrate these aspects.

Purpose of the Study:

• To propose a unifying computational model for attention based on a bidirectional recurrent gating mechanism.
• To integrate this mechanism within the ventral visual stream architecture.
• To demonstrate the model's ability to replicate attentional tasks and neural properties.

Main Methods:

• Developed a computational model with feedforward pathways for feature extraction and recurrent connections for modulatory signals.
• Trained the model on recognition and segmentation tasks.
• Evaluated the model on canonical attention tasks (orienting, filtering, visual search) and psychophysical phenomena.

Main Results:

• The model successfully performed orienting, filtering, and visual search on complex scenes.
• It replicated key psychophysical phenomena, including perceptual load and inattentional blindness.
• Internal model units exhibited neural properties consistent with primate physiology (e.g., gain modulation, border-ownership coding).

Conclusions:

• Diverse attentional and binding phenomena can emerge from error-backpropagation within specific architectural constraints.
• The proposed model offers a powerful tool for neuroscience research.
• Presents a bio-inspired alternative to conventional AI architectures.