Feature Binding

circles2.jpgSensory integration in a broad sense deals with the integration of input across modalities. That is, vision as integrated with audition, taste as integrated with smell, until a coherent representation can be established. However, it is important to note that integration within a modality must occur as well. This phenomenon is referred to as feature binding.

Spatially Selective Attention

One theory of visual feature binding is posited by Treisman (1998), and attempts to solve the binding problem. This theory can be thought of as the spatially selective attention theory of feature binding. As the name implies, features are bound together through the work of spatially selective attention, and this requires different elements in order to work. First off, there is a “master map of locations”, which can be thought of as a spatially represented map of the visual field. Secondly, there is a way in order to detect the presence of features anywhere in the visual field; these features are “flagged” in order to signify their presence. An implicit spatial layout of the features is also included in the theory, most likely to account for implicit awareness of feature locations in the absence of spatially selective attention.

When one examines a visual scene, the attentional window scans over the master map of locations. This attentional window can be scaled down such that only small portions of the visual field are under the attentional mechanism. As the attentional window scans the visual map, the features that are flagged within that window are entered into the presently active object representation. In this way, separate features that are simultaneously under the attentional window can be bound into a single representation, and feature binding is achieved (Treisman, 1998).

Studies involving patients with apparent feature binding deficits have been used to support feature binding theories. RM is one such patient whose deficits have been analyzed in terms of Treisman's theory of feature binding. When displayed with an array of objects, differing in color, shape, orientation, etc., (???), RM made a significant number of illusory conjunctions, as opposed to color intrusions or letter intrusions. An illusory conjunction is when different features of two separate objects are incorrectly bound to the opposite object. For example, if a green T and a red O are presented, an illusory conjunction would be if the subject perceived a red T and a green O. This related to the binding problem because it shows how the object shape (letter) and object color and encoded separately, and can in some instances be combined incorrectly. This type of error is opposed to color or letter intrusions, in which a letter or color not shown is perceived; these types of errors can be attributed to mechanisms involved in memory or guessing, and thus do not indicate deficits in feature binding, per se. Illusory conjunctions, on the other hand, support the claim that features are represented separately from location in the absence of spatial attention, and are then bound to objects later in the binding process.

Temporal Correlational Hypothesis

A different, yet more neurological, account of feature binding is the temporal correlational hypothesis, in which different perceptual features are bound together by means of neural synchrony. This account relies on the neurophysiological properties of cells, and their ability to synchronize activity due to the rhythmic effects of ion conductances. A brain area that has been attributed with this property is the thalamus, which takes in sensory information from the neural periphery (i.e., sensory organs), and relays it to the cortex. In doing so, different perceptual cues are bound together due to the inherent physiological nature of the thalamic and thalamic relay cells; this can, in theory, explain how different perceptual cues are bound together into a unified representation (Singer and Gray, 1995).

Upon analysis of this claim, it seems as though the integration that might be attributed to the thalamus would be of the multi-modal variety, since within-modality integration could perhaps more easily be attributed to the cortical areas in which parallel features of a unimodal input would be integrated with each other (for example, the color, motion, line orientation areas of the cortex, with respect to visual input). Since the thalamus receives driver input from many different peripheral sources, it would seem appropriate that the integration mechanisms associated with thalamic activation would be of the multi-modal variety.

Neural Networks: The Competitive Layer Model (CLM)

A different class of feature binding theories are those involving neural networks. One such network is referred to as the Competitive Layer Model (CLM), and is modeled to reflect the neural correlates of feature binding, thus remaining biologically relevant (Wersing et al, 2001). As the model is too complicated to discuss in detail here, a basic description will be offered.


The model implements the principle of topological segregation, in which competing groups of nodes represent different features. These features are selected by columnar circuits spanning the neural layers ,which are referred to as winner-take-all (WTA) circuits. There are lateral interactions within each layer which are involved in encoding the contextual information of the stimulus, and create groupings based on the extent to which the features are similar, or compatible, with one another. When different nodes in a single layer are activated at the same time, those features are bound together. The principle of topological segregation is one that can be implemented in other models besides the CLM, and can be used to augment the stability of other feature binding models (Wersing et al, 2001).

The CLM includes self-inhibitory mechanisms as well as Gestalt-based figure-ground analysis, which enables the model to pick out the most salient features of a stimulus without requiring a biologically unsound number of layers (Wersing et al, 2001).

How should we deal with these opposing theories?

The theory posited by Treisman referring to feature binding as a result of spatially selective attention, requires that the concept of “attention” is a well-defined one. The concept of “attention”, however, cannot be attributed to certain neurological facts, since “attention” itself is somewhat of an abstract theory. In this sense, Treisman’s feature binding theory is a good one insomuch as it gives a theoretical framework in which to fit data, thereby setting experimental data in a significant direction. However, since theories are somewhat flexible in the sense that data can be differently interpreted to be consistent with the theory, in order for a theory to be good it must be able to be disproved by specific characteristics of the data. Since “attention” itself is an abstractly defined term, it might be the case that it would be very difficult to disprove a theory whose central element was attention. However, the notion of expanding and shrinking receptive fields of visual neurons was referred to by Treisman (1998), and could serve as the concrete definition of “attention”, thus making the theory more disprovable and therefore valid. A major problem with theories as a guideline for validity of empirical data is that since there is so much experimental data to explain, differing theories can claim to explain different aspects of the data.

For example, a supposedly contrasting theory to Treisman’s feature binding is the aforementioned temporal correlation hypothesis, in which features are bound together in an object representation due to the synchronous firing of specific groups of neurons (Singer & Gray, 1995). However, Treisman (1998) claims that the temporal correlation hypothesis does not provide answers to the same binding problem that her theory tries to tackle. Treisman (1998) claims that neural synchrony can explain how features remain bound, but does not explain how they are originally bound together.

Thus it is clear that the binding problem is not only a real one, but a problem that has many different facets that are able to be explained by both psychological and neurological data.

Back to Multi-Modal Integration
Back to Main