1994 Paper

A neural network is built to model a visual search task. This task involves detecting whether there is a target letter in a display containing other non-target letters. Half the displays contain a single instance of the target letter (the "present" trials), and half contain no instance of the target letter (the "absent" trials). The model makes certain predictions about reaction time (RT) in the task: (i) RT in the absent trials will increase rapidly as the number of non-target items increases; (ii) while there may be some small increase in RT in the present trials as the number of non-target items increases, this will be much smaller than in the absent trials; (iii) RT in the absent trials will be greater than in the present trials.

To test these predictions, the following experiment was carried out. Participants had to press one key if a capital E was present in a display comprising Ls and Fs and another key if there was no E present. Each participant pressed the present key with their right hand and the absent key with their left hand.

The experiment was designed such that across the trials the number of non-targets varied. Four display sizes were used, containing respectively, 5,10 ,15, or 20 letters. On present trials the target took the place of one of the non-targets. Both kinds of non-targets were equally represented in the displays. For each of the four display sizes there were 20 present trials and 20 absent trials.

On each trial a central fixation point was presented for 100 msec and then was irnrnediately replaced by a display which remained on for 150 msec. Participants were told to respond as quickly as possible. The data from this experiment are given in Table 1.

Table 1. Mean reaction times in msecs and percentage errors for different display sizes. Present trials are where the target E was present in the display and absent trials where it was not.

Display size
5 10 15 20
Present trials
RTs 429 435 454470
Absent trials
RTs 550587604675

The RTs were analysed in a 2 by 4 repeated measures analysis of variance in which trial type (present vs. absent), and display size (5,10, 15, 20) were the factors of interest. The analysis revealed a statistically significant main effect of trial type (F(1,11) = 10.85, p < .01), a statistically significant main effect of display size (F(3,33) = 3.1, p < .05) but not a significant trial type by display size interaction (F(3,33) = 1.8, n.s.) It was concluded that the predictions held up. Performance was poorer, and more affected by display size, on the absent trials.

Having obtained this support for their neural network model the researchers went on to design a further experiment. This was to test the counter-intuitive prediction that carrying out a simultaneous cognitive task would have no effect on visual search. To test this, half the participants were given a concurrent short-term memory task on every trial. They were presented with a list of six random digits and had to hold them in mind whilst the visual display was presented. When the trial was complete, the participant then had to recall the digits in the correct serial order. This was the dual-task condition. In the single task condition, participants viewed the displays and made the appropriate responses without an additional cognitive load being imposed on them.

In Experiment 2, participants were presented with the same displays as before; however, now the task was simply to press a single key when the target was present and not make any response when it was absent. RT and errors were recorded as in Experiment 1 and means for these data are presented in Table 2.

Table 2. Reaction times in msecs and percentage errors for different display sizes.

Display size
5 10 15 20
Single task
RTs 413427 460468
Dual task
RTs 347367370675
%Errors 22.537.523.724.5

The RTs were analysed in a 2 by 4 split-plot analysis of variance in which group (single task vs. dual task) was the between-subjects variable and display size (5, 10, 15, 20) was the within-subjects variable. The main effect of display size was significant (F(3,51) = 6.85, p < .01). Much to their surprise, the main effect of group was also significant (F(1,17) = 5.01, p < .05). The condition by display size interaction was not statistically significant (F(1,51 ) = 1.85, n.s.)

From this pattern of results the researchers conclude that their predictions basically hold up but that there is some additional effect of the secondary short-term memory task. Specifically, a short-term memory task prevents conscious processing of the primary search task and thus reduces interference from the non-target items. This additional effect was duly programmed into the neural network model.

[Back to Criticise-the-Experiment List] [Back to Main Page]