Learning and Automaticity: A Connectionist Explanation of the Stroop Effect

Devin McAuley and Janet Wiles
Copyright © 1998
Updated for version 2.0 by Simon Dennis 1999.

Table of Contents

1. Introduction
2. Model Architecture
3. The Stroop Simulation
4. References
5. Slides

Introduction

Suppose an individual is presented with the names of different colors, written in colored inks, and then asked to either read the word or name the color of the print. In such tasks, called Stroop tasks, it is easy to set up a situation where the color words may be written in different colored inks, which in some cases will conflict. For example, the word GREEN might be written in red ink, or the word RED might be written in green ink. An interesting empirical question is whether the processes involved in reading words and naming colors interfere. The basic Stroop effect involves a set of empirical findings, demonstrating that there is an asymmetry in the observed interference between color-word reading and color naming. When reading the names of colors, the background color of the word is relatively easy to ignore. However, in color naming, the meaning of the word tends to influence how quickly we respond with the correct color. For example, identifying a background color, such as red, is quicker when the word is RED than when the word is GREEN. In the later case, the word GREEN interferes with the process of identifying the red ink. Overall, word reading is quicker than color naming.

The Stroop effect illustrates an important aspect of selective attention: It is easy to ignore some features of the environment, but not others. One explanation for the Stroop effect is that it reflects a difference in processing speed: word reading is faster than color naming, so color naming simply does not have the opportunity to interfere with word reading. Hypothetical differences in processing speed, suggested by Stroop interference, are consistent with a distinction between two types of cognitive processes: controlled and automatic (Shiffrin & Schneider, 1977). Controlled processes are assumed to be voluntary, to require attention, and to be relatively slow, whereas automatic processes are assumed to be involuntary, to not require attention, and to be relatively fast.

Based of the processing speed explanation, the Stroop task has gained wide acceptance as a method for distinguishing between controlled and automatic processes. If process A interferes with process B, but process B does not interfere with process A, then process A is automatic and process B is controlled. An interesting challenge arises when a task such as color naming is identified as both controlled and automatic, by varying the other task involved. Color naming is identified as a controlled process when the other task is word reading, but as an automatic process when the other task is shape naming (MacLeod and Dunbar, 1988).

Cohen, Dunbar and McClelland (1990) proposed an alternative connectionist explanation of the Stroop effect, which does not distinguish between automatic and controlled processing. Instead, they proposed that automaticity is a continuum, and that Stroop interference depends on the relative degree of learning of the respective tasks, not on processing speed. According to this view, asymmetries in performance such as those observed in the Stroop task can be accounted for by differences in experience.

Model Architecture

The architecture of the Cohen, Dunbar, and McClelland (1990) model is shown in Figure 1. It is a backpropagation network with three layers. The input layer consists of six units: two task units representing the color naming and word reading tasks, two color units representing red ink and green ink, and two word units representing the words RED and GREEN. The middle (or hidden) layer consists of four units. Two of the hidden-layer units represent the color-naming pathway, while the other two units represent the word-reading pathway. The task units are used to gate (or selectively attend) to the appropriate pathway. The extent that the network is able to selectively attend to either the color naming or word-reading pathway depends on the relative strength of the weights in the to-be-ignored pathway. The output layer consists of two units: one for the response "red" and one for the response "green".

Figure 1: Network architecture of Cohen, Dunbar, and McClelland (1990) Stroop model.

While the network is a multilayer architecture the weights between the input units and hidden units and the biases of the hidden units are fixed (indicated by the pink arrows) to implment the selection mechanism.

Exercise 1: Examine the weights between the input units and the hidden units as well as the biases of the hidden units. Describe how the network is implementing the selection mechanism with reference to the sigmoid activation function.

Exercise 2: The patterns in the training set are designed to capture in a simplified form the learning that people undergo before doing the Stroop task (with the dashes corresponding to either the neutral word (noncolor word or nonword) or the neutral color (black)). Examine each of the training patterns and describe the experience that is represented by each different pattern. Why are there more multiple copies of the word naming patterns?

Exercise 3: The patterns in the test set are designed to capture the conditions within a Stroop experiment. Which patterns in the test set represent the control, congruent, and conflict conditions in the Stroop experiment? Why are the "word - -" and "color - -" conditions included in the test set?

Exercise 4: Randomize the weights and biases and test the network on each pattern in the test set by clicking on the pattern and then on the Feedforward button. Record the activations of the "red" and "green" output units in the following table. These represent the performance of the network before training.

Table 1: Test conditions before learning

Task Demand	Ink Color	Word Input	"Red"	"Green"
Color naming	---	---
Color naming	Red	---
Color naming	Red	GREEN
Color naming	Red	RED
Word reading	---	---
Word reading	---	RED
Word reading	Green	RED
Word reading	Red	RED

The Stroop Simulation

Exercise 5: We will now train the network to model the day-to-day experience we have with word and color naming. It is useful to be able to watch the performance of the network as it trains. To do so, create a graph of the output error by selecting the BPLearn button and then creating a graph on the workspace. Be sure to set the graph to plot the error signal (by right clicking on the graph and updating the bound value name). Rescaling the graph to a maximum of 10 will allow you to see the entire error function.

You are now ready to train the network. Click on the BPLearn button. This will complete 40 cycles of training on patterns in the training set. One cycle of training (an epoch) is one complete pass through the training set. Record the final error value you obtained and freeze the error graph. Retest the performance of the network on the experimental conditions of the Stroop task and record the outputs in Table 2.

Table 2: Test conditions following learning

Task Demand	Ink Color	Word Input	"Red"	"Green"
Color naming	---	---
Color naming	Red	---
Color naming	Red	GREEN
Color naming	Red	RED
Word reading	---	---
Word reading	---	RED
Word reading	Green	RED
Word reading	Red	RED

Performance on the Stroop task is measured by people's response times on the word and color naming tasks. The basic Stroop effect involves four empirical findings. (1) Overall, word reading is faster than color naming. (2) Reaction times in color naming are influenced by the word used: congruent conditions facilitate color naming, whereas conflicting conditions interfere with naming. (3) However, reaction times in word reading are not affected by the color of the ink. (4) Finally, facilitation, when it occurs, is usually weaker than interference.

To account for the basic Stroop effect with their model, Cohen et al used an iterative procedure that mapped the activation values of the output units into response times for each test condition. This allowed a comparison of the model with peoples' response times.

For the purpose of this exercise, we will assume that the output unit with the highest activation value would be the response given, and that response time is proportional to one minus the activation of that unit. The following equation can be used to calculate response times (the variable A represents the activation of the winning unit).

RT = C (1 - A) + RTmin

The parameters C and RTmin are the proportionality constant and the minimum reaction time, respectively.

Exercise 6: Use this procedure to calculate "response times" for the control, congruent, and conflicting conditions, for both the color naming and word reading tasks. Try using C = 2000 and RTmin = 450. Construct and compare the graphs for your calculated response times and the empirical data shown in Table 3. On the x axis you should have the neutral, conflict and congruent conditions and on the y axis the reaction time. You should have one line representing the color naming task and one representing the word naming task.

Table 3: Human response times for the Stroop task (Cohen et al., 1990).

Task Demand	Ink Color	Word Input	Response Time (ms)
Color naming	Red	---	650
Color naming	Red	GREEN	860
Color naming	Red	RED	570
Word reading	---	RED	450
Word reading	Green	RED	460
Word reading	Red	RED	445

Exercise 7: Does the model provide a qualitative account of the four empirical findings outlined above?

Exercise 8: In what ways does the model's quantitative performance differ from human performance. How might you change the C and RTmin parameters to produce a better quantitative fit to the empirical data?

Exercise 9: Why is a quantitative fit important?

References

Cohen, J.D., Dunbar, K., and McClelland, J.L. (1990). On the Control of automatic processes: A parallel distributed processing account of the Stroop effect. Psychological Review, 97, pp. 332-361.

Macleod, C. M., and Dunbar, K. (1988). Training and Stroop-like interference: Evidence for a continuum of automaticity. Journal of Experimental Psychology: Learning, Memory, and Cognition, pp. 14, 126-135.

Shiffrin, R. M., and Schneider, W. (1977). Controlled and automatic human information processing: II. Perceptual learning, automatic attending, and a general theory. Psychological Review, 84, pp. 127-190.