The representation of visual features and the binding problem: Feature binding in parallel networks

This thesis examines the problem of feature binding in the visual system. Separate feature dimensions, such as color and shape, must be synthesized to form integrated representations of whole objects. The binding problem is a problem that arises in parallel networks in determining which features go together, and correctly assigning features to representations of objects. The thesis research consists of fundamentally two parts: a computational analysis of the binding problem which outlines both possible solutions to the problem and the feature representations suggested by these solutions; and empirical work which examines the feature representations actually used in the visual system.;In the first part of the thesis, I argue that once the mutual constraints of possible feature representations and possible binding mechanisms are considered, there are only two broad classes of solutions to the binding problem. One solution, referred to as the serial model, uses global feature representations and a serial binding mechanism. The other class of solution is a parallel framework in which duplicate copies of feature detectors are indexed by a metric such as location, and features are implicitly bound together by virtue of their common index. The empirical work tests this analysis. Experiments 1 and 2 show that, consistent with the parallel framework, feature representations are tagged with location information. Experiments 3 and 4 suggest that location is not a special attribute for binding features together. These experiments find that features are also encoded with information about the surrounding context, perhaps including whether the context is a word or nonword. But if features are implicitly bound together through common indices such as location and context, then what is the role of focused attention? Experiment 5 suggests that a feature conjunction as a part of a larger whole is represented differently from the same feature conjunction as a whole to itself. One possible role for attention may then be in transforming features from the noncanonical level of tagged representations to an invariant canonical feature description.