Summary: A new model of attention explains how the human brain allocates limited perceptual resources to prioritize goal-relevant information in changing environments. Called “adaptive computation,” the model predicts which visual details people emphasize—such as a walk signal at an intersection rather than an eye-catching car—based on task demands.
In controlled experiments that tracked attention to multiple moving objects, the model accurately forecast where attention would shift and how difficult participants found the task. These results clarify why nonessential details can vanish from awareness when we concentrate, and how the brain flexibly optimizes attention in real time.
Key Facts:
- Adaptive Computation: Perceptual processing is allocated according to goals, so irrelevant stimuli are filtered out.
- Dynamic Attention: Attention reallocates rapidly and flexibly as visual demands change.
- Human-Like AI: The model provides a blueprint for AI that selectively ignores irrelevant stimuli in goal-directed tasks.
Source: Yale
The capacity to attend to the environment shapes what we perceive and how we act. As someone crosses a busy street, their attention shifts from glossy storefronts or a striking advertisement to the traffic lights, crosswalk signals, and nearby pedestrians.
Attention is finite and often directed by immediate goals. In new research, psychologists at Yale investigate what happens when attention is deliberately steered toward a task—such as safely navigating an intersection—rather than drawn to novel or flashy items in the surroundings.
Published in Psychological Review, the study presents a computational model of human attention that explains how the mind determines task relevance in complex, dynamic scenes and reallocates perceptual capacity accordingly.
“We have a limited number of resources with which we can see the world,” said Ilker Yildirim, assistant professor of psychology at Yale and senior author of the study. “We think of these resources as elementary computational processes; each perception we experience—like an object’s position or speed—requires some of these basic perceptual computations.”
The researchers developed “adaptive computation,” a system that ration these elementary perceptual computations so goal-relevant objects receive more processing. For example, when crossing a busy street, adaptive computation would allocate more resources to the walk signal and oncoming vehicles than to a parked luxury car.
“Our model describes a mechanism by which attention identifies what in a dynamic scene matters for the current goal and then apportions perceptual effort accordingly,” said Mario Belledonne, a graduate student at Yale and co-author of the paper.
To test the model, participants viewed eight identically colored circles on a screen. Four of those circles were briefly highlighted, and participants were instructed to track the highlighted targets while all eight circles moved randomly. Tracking multiple identical objects creates a continuously shifting pattern of attention among targets and distractors.
Researchers measured attention at sub-second resolution by asking participants to press the space bar whenever they detected a briefly flashing dot on any object. The frequency and timing of these detections reveal where attention was directed at each moment. The adaptive computation model closely matched these fine-grained patterns of attentional deployment.
In a second experiment, the team varied the number of identical distractors and the speed of motion while participants tracked four target objects. After the motion stopped, participants rated how difficult they found the task. The model explained these subjective effort ratings: trials requiring more perceptual computations were reported as more effortful.
By linking a measurable computational load to subjective difficulty, the study provides a formal signature of the mental exertion experienced during sustained attention.
“Our aim is to uncover the computational logic of perception and attention by creating algorithms that mirror human behavior and comparing their performance to people,” Yildirim said.
The model also illuminates why seemingly irrelevant items can vanish from consciousness when we focus: attention does not process everything equally but strategically preserves resources for what matters for the task.
“This work could inform AI systems that behave more like humans,” Yildirim added. “Such systems, when given a goal, might intentionally overlook irrelevant but salient stimuli so they can act flexibly and safely in real-world settings.”
The research team includes Brian Scholl, professor of psychology at Yale, and Eivinas Butkus (Columbia University), a former member of Yildirim’s lab.
Funding: Supported by a grant from the U.S. Air Force Office of Scientific Research.
About this neuroscience and attention research news
Author: Bess Connolly
Source: Yale
Contact: Bess Connolly – Yale
Image: The image is credited to Neuroscience News
Original Research: Closed access.
“Adaptive computation as a new mechanism of dynamic human attention” by Ilker Yildirim et al. DOI: 10.1037/rev0000572
Abstract
Adaptive computation as a new mechanism of dynamic human attention
Attention continually concentrates visual processing to satisfy behavioral goals. How can this be expressed computationally?
We propose adaptive computation: a dynamic mechanism that links the moment-to-moment application of perceptual computations with their influence on decision outcomes. Adaptive computation ration perceptual processing across objects on the fly, guided by a general and principled measure of task relevance.
We evaluate adaptive computation using multiple object tracking (MOT), a classic paradigm where observers follow a set of target items moving among identical distractors. The model accounts for established MOT phenomena—such as trial-level tracking accuracy and target localization error—while also capturing subsecond attentional deployment between objects and the subjective sense of effort, both measured here and not previously modeled.
Crucially, this account is domain-general and does not rely on MOT-specific heuristics. We discuss how adaptive computation could serve as a mechanistic framework for understanding the dynamic operation of many forms of visual attention.