Emergent brain-like representations in a goal-directed neural network model of visual search
Visual search, the act of locating a target among distractors, is a fundamental cognitive behavior and a core paradigm for studying visual attention. While its behavioral properties are well characterized in humans and non-human primates, the underlying neural mechanisms remain largely unspecified. To address this gap, we developed a biologically aligned neural network model trained to perform visual search directly from pixels in natural scenes. This model exhibits strong generalization to novel scenes and objects, produces human-like scanpaths, and replicates previously known behavioral biases in humans. By analyzing the internal representations of the model, we found that it naturally develops a retinocentric cue-similarity map and prospective fixation signals, features that closely resemble neural activity in the primate fronto-parietal network. Beyond reproducing known behavior and neural signatures, the model makes testable predictions about the geometry and dynamics of internal representations underpinning cue-driven prioritization, fixation preferences, their perspective memories, and prospective plans. These findings offer a computational framework for understanding visual search and a roadmap for future neurophysiological and behavioral studies.