Mario Belledonne¹ (), Ilker Yildirim^1,2; ¹Yale University, ²Wu Tsai Institute

Inattentional Blindness demonstrates sophisticated tradeoffs realized on-the-fly by visual cognition: When focused on a goal, we are often unaware of even salient events that involve irrelevant elements. What is the computational nature of these tradeoffs, and how do patterns of selective processing ultimately impact visual awareness? Here, we present the *Perceptual Reframing Hypothesis*, in which inattentional blindness arises from adaptive processes within visual cognition that balance the computational complexity of a perceptual “frame” (e.g., how to represent and organize objects in the scene) with the information necessary for decision-making. We realize this hypothesis with a novel computational mechanism, Multigranular Optimization (MO), which dynamically builds and manipulates increasingly more efficient frames. Every several hundred milliseconds, the MO model uses an online stream of task-relevance signals to reframe object representations — on a ladder from fine-grained individuals to coarse ensembles. Under bounded resources, the model maintains task-relevant objects as fine-grained individuals to sufficiently inform decision-making and reframes task-irrelevant elements into coarse, cheap ensemble summaries to exponentially decrease computational complexity. This process yields humanlike patterns of awareness, as reframed ensemble representations preclude the explicit encoding of visually similar, irrelevant objects by “explaining away” their corresponding sensory signals. Across several studies, the MO model not only recapitulates classical awareness patterns (e.g, Most et al., 2001) but also predicts new awareness phenomena that we confirm in a new preregistered experiment. The model also explains, for the first time, detailed, trial-level variance in humans’ awareness responses. Alternative ablation models do not recover humanlike patterns, and crucially, detailed comparisons establish that MO realizes resource-rational performance benefits that outweigh any algorithmic overhead. By revealing the algorithmic basis of inattentional blindness, this work begins to ground the striking phenomenology of awareness in precise and interpretable computational terms.

Acknowledgements: AFORS, NSF CAREER