Guided Search
Guided Search is a visual attention model developed by Jeremy M. Wolfe, Professor of Ophthalmology at Harvard Medical School . The “theory of the controlled search” is a continuous further development of the guided search model, which originally emerged as a criticism and improvement of the feature integration theory by Anne Treisman . The model enables computer simulation of human visual search and object recognition .
Are you looking for B. a certain sweater (target stimulus) in his closet that is full of other items of clothing such. B. socks, T-shirts and other sweaters (distraction stimuli), you start a visual search (consciously and unconsciously). The whole closet is searched for the (e.g. blue) sweater that has a certain color, shape and size. You can be briefly distracted by sweaters of a different color and blue socks, but keep looking until you find the blue sweater. In his Guided Search model, Wolfe tries to explain why people find this sweater relatively easy despite a world that is overloaded with stimuli.
model
Visual search with guided search
In the visual search paradigm , a certain number of stimuli are presented in a search display. The test subjects should find a target stimulus among distraction stimuli. Processes of increasing activation that emanate from the stimuli are called bottom-up (unconscious) and search criteria that influence the search, top-down-controlled (conscious), see top-down and bottom-up . The stimulus, which is very different from the other stimuli, is assigned a high level of activity. A relatively similar stimulus, on the other hand, is only assigned a low level of activity. If the situation is influenced top-down, objects that resemble the target stimulus receive an additional increase in activation. If one looks at the contents of the wardrobe with an open mind, a neon-pink raincoat will be the first thing to notice (pop-out) when we look for the blue sweater. He gets the highest bottom-up activity. But because the blue sweater is to be found, all dark sweaters are more active than light sweaters; one quickly turns away from light-colored sweaters and looks inside the darker sweaters (top-down).
Activations main map
In the GS2 model, the search display is represented in a location-based main map, i.e. H. the spatial arrangement is retained ( retinotopy ). This map shows the difference in color and orientation of the stimuli to one another. Each stimulus on the feature map thus receives a summed activation of the features (orientation and color). The stimulus with the highest activation attracts attention. However, it sometimes happens that the stimulus with the highest activation is not identified as the target stimulus. In this case, attention is drawn to the stimulus with the next highest activation. This process is repeated until the target stimulus has been discovered, all items whose activation exceeds the threshold value have already been checked, or the maximum search duration has been exceeded.
Feature maps
The activations on the main card are calculated from the individual feature cards. As with the feature integration theory , an object is first differentiated into its features (blue, long, crooked, ...). These features can be categorized (color, shape, orientation, ...), with each dimension being represented in a feature map. With these cards, too, the spatial order is retained and activation is coded. The degree of activation depends on the salience (difference from the other stimuli).
Exemplary:
- The yellow circle shows a high salience in the dimension "color". It is different from all three other stimuli. In contrast, each red circle in this dimension only differs from one other stimulus. In the "color" feature card, the yellow circle receives the highest activation, as it is most different from the other stimuli.
- In the "orientation" dimension, the respective circles do not differ from one another, so all 4 stimuli are assigned the same, low activity in the associated characteristic map.
Since an object can have a large number of features, there would theoretically be just as many feature cards. In the current model, however, only the dimensions "color" and "orientation" are taken into account.
Activation in the feature cards is caused by two processes:
- stimulus-driven bottom-up process
- person-controlled top-down process
(see top-down and bottom-up )
Categorical channels
Bottom-up and top-down processes are not mutually exclusive, but mostly work together. This can be seen by looking at another aspect of the model: the categorical channels. Your output generates the activations on the feature cards. The uniqueness of the target stimulus now ensures that the corresponding channel produces a high level of activation for the one on the feature map.
Each feature dimension has several specifically tuned channels that react selectively and optimally to a basic feature expression. For example, there could be a channel for vertical, a channel for an inclination of 45 ° and a channel for horizontal for the orientation dimension. The task of these channels is now to create an activation value for each item of the orientation dimension, by a.
- calculate the difference between the item and its neighbors
- calculate the mean value from these differences and
- make a weighting.
Steps 1 and 2 are bottom-up dependent, step 3 is top-down. The weighting of the output is necessary if the salience of the target stimulus is too low to identify it as the target object.
simulation
"Guided Search" also includes a computer simulation. It is based on the assumption that the functionality of the model can be transferred to humans if the simulation model can be used to generate data similar to that of the test subjects . The simulation can actually mimic visual search in such a way as to generate data that is very similar to human-generated data. The error rates are also similar, but the deviations are slightly larger here.
Range of explanation of the model
Furthermore, the model enables speculations regarding
- illusory links as activation hills on the main map without spatial allocation
- Effect of density as the overlap of activation hills
- different weighting of the feature maps.
However, some aspects were neglected or could not yet be simulated. For example:
- Influencing the feature maps among one another
- local differences in the visual field
- basic weighting within a dimension
- Discrepancy in error rates
The "Guided Search 4.0" model from Wolfe now exists.
literature
- JM Wolfe: Guided Search 2.0: A revised model of visual search. In: Psychonomic Bulletin and Review. 1, 1994, pp. 202-238. [1] (PDF file; 7.80 MB)
- J. Müsseler, W. Prinz: General Psychology . Spektrum Akademischer Verlag, Heidelberg 2002, ISBN 3-8274-1128-9 .
Web links
Individual evidence
- ↑ Jeremy M. Wolfe: Guided Search 4.0: Current Progress with a model of visual search. In: W. Gray (Ed.): Integrated Models of Cognitive Systems . Oxford University Press, 2007, ISBN 978-0195189193 , pp. 99-119. ( online )