0
0This research introduces bidirectional recurrent gating, a biologically inspired computational mechanism designed to unify various forms of visual attention and feature binding. The authors developed a neural network architecture that mimics the ventral visual stream, utilizing feedforward pathways for feature extraction and top-down connections to modulate information flow. Through multitask learning on complex datasets, the model successfully performs diverse behaviors such as spatial orienting, visual search, and object tracking. Significantly, the system replicates human psychophysical phenomena, including inattentional blindness and perceptual load effects, while developing internal neural properties consistent with primate physiology. This work suggests that a single, neurally plausible framework can account for how the brain selects relevant information and integrates features into coherent objects. Ultimately, the model serves as a powerful bridge between computational neuroscience and advanced artificial intelligence.
References:
Salehi S, Lei J, Benjamin A S, et al. Modeling attention and binding in the brain through bidirectional recurrent gating[J]. Nature Communications, 2026, 17(1): 4072.
