Where does feature integration collide?

Feature-integration theory says the visual system registers simple features such as color, orientation, and motion automatically and in parallel across the whole field, but that combining those features into one object needs focused attention aimed at a single location. Treisman and Gelade set this out in 1980 to explain why searching for a single feature is fast and stays flat as more items are added, while searching for a conjunction of two features gets slower with each extra item. Treisman and Schmidt showed in 1982 that when attention is spread too thin, features can be miscombined into illusory conjunctions, so a red X and a green O are reported as a green X and a red O. The claim is narrow and specific: binding is the work of attention, and without it the features float free. This is the referent that the other two meanings on this map are measured against.[1][2]

Unit

A visual object bound from separate features by attention

Recovery means

Not applicable

Measured by

Visual search reaction time by set size, and illusory-conjunction rates under divided attention

The binding problem is the wider question that feature integration is only one answer to: how the brain, which handles color, shape, motion, and location in separate and distributed populations of neurons, ever assembles them into one coherent percept of a single object. Treisman's 1996 review treated attention-based binding as her own contribution but set it beside other proposals, including the idea from Singer and colleagues that neurons coding parts of the same object fire in temporal synchrony to tag them as belonging together. The problem is broader than any one theory: it spans behaviour, neurophysiology, and computational modelling, and it asks about the neural code, not just about reaction times. Feature integration is a mechanism offered for this problem, not the same thing as the problem itself. Citing feature integration when you mean the general binding problem quietly shrinks a large open question down to one account of it.[3]

Unit

The general problem of unifying distributed neural features into one percept

Recovery means

Not applicable

Measured by

Studied through single-unit recording, neural synchrony, lesion and imaging work, and models, not a single scale

Guided search is a revised account of the same search data that keeps parallel feature processing but rejects the strict split between parallel feature search and serial conjunction search. In Wolfe's 1994 model the parallel stage does more than register features: it computes an activation map that steers attention toward the most promising items first, so conjunction search is guided rather than an exhaustive item-by-item scan. This blurred the clean parallel-versus-serial line that early feature-integration theory had drawn, and Treisman herself revised the theory over the years in response to it. The phenomena are shared with feature integration, but the proposed mechanism and the boundary between stages differ. Treating the original 1980 dichotomy as the settled account overlooks that its most-tested claim was later reworked.[4]

Unit

A search process in which parallel feature maps guide attention toward likely targets

Recovery means

Not applicable

Measured by

Visual search reaction-time slopes across set sizes, modelled against a guided activation map