CECN1 Stroop

From Computational Cognitive Neuroscience Wiki

Jump to: navigation, search

Contents

The Stroop Task

  • The project file: stroop.proj (click and Save As to download, then open in Emergent -- NOTE: requires version 4.13 or higher

Back to CECN1 Projects

Project Documentation

(note: this is a literal copy from the simulation documentation -- it contains links that will not work within the wiki)


  • To start, it is usually a good idea to do Object/Edit Dialog in the menu just above this text, which will open this documentation in a separate window that you can more easily come back to. Alternatively, you can always return by clicking on the ProjectDocs tab at the top of this middle panel.

You should see the network just as pictured in Figure 11.5 in the text.

  • Begin by exploring the connectivity using r.wt

You will notice that all of the units for red versus green are connected in the way you would expect, with the exception of the connections between the hidden and output units. Although we assume that people enter the Stroop task with more meaningful connections than the random ones we start with here (e.g., they are able to say "Red" and not "Green" when they represent red in the environment), we did not bother to preset these connections here because they become meaningful during the course of training on the task.

Training

Next, let's look at the training environment patterns.

You will see 2 tables. The top table provides the key to the bottom table, by indicating what units each box correponds to within the network (e.g. the first box in the color column corresponds to the "g" = green unit, the second box in the color column corresponds to the "r" = red unit, etc). Within the bottom table, you will see 4 rows, 2 for training the word reading pathway, and 2 for color naming (one event for each color and word). The frequency of these events is controlled by the last column in the table. You can see that the word reading events have a frequency of 3, while the color naming events are at 2. This frequency difference causes word reading to be stronger than color naming. Note that by using training to establish the strength of the different pathways, the model very naturally accounts for the MacLeod & Dunbar, 1988 training experiments.

(For further exploration, the relative frequency can be changed by modifying the arguments for .MakeTrainEnv and running the program .MakeTrainEnv.Run().)

Now, let's train the network.

You can see that even after the network learns to perform the task with no errors (black line asymptotes at 0 very early), the network continues to learn after reaching this hgh level of performance (as evidenced by the red line reflecting the average speed of response continuing down).

We next explore the differential weight strengths for the two pathways that develop as a result of training. Although the observed weight differences are not huge, they are enough to produce the behavioral effects of word reading being dominant over color naming.

View the network at .T3Tab.StroopNet, and select r.wt from the middle panel. Next, select the Interact button (the arrow button) on the far right of the right most frame. Click on the output units and note that the word reading pathway is stronger. Similarly, press on the "g" and "G" hidden units in sequence, and note that the "G" (word reading) hidden unit has stronger weights overall.

Question 11.1 (a) Report the weights for the "g" and "G" hidden units from their respective input, PFC, and output units (you need only report the "gr" output weights). (b) At which layers in the network are the differences greatest? (c) Can you explain this in terms of the error signals as they propagate through the network?

Basic Stroop Task

Now, let's test the network on the Stroop task. First, we will view the testing events.

You should see 6 rows, 3 for word reading and 3 for color naming. In the first 3 rows, you should see the control, conflict, and congruent conditions for word reading, all of which have the word reading PFC task unit clamped (the "wr" unit). All patterns have the "R" word unit active. Control does not have any active color units, conflict adds the "g" (green) color unit active, and congruent adds the "r" (red) color unit.

In rows 4-6, you should see the color naming events. You should observe a similar pattern of inputs for color naming.

Now, we can actually test the network.

You will see the response times (cycles of settling) plotted in the graph, which should resemble Figure 11.6 in the textbook. The data points are plotted at X-axis values of 0, 1, and 2, corresponding to the Control, Conflict, and Congruent conditions, respectively. The different colored lines reflect color naming and word reading, respectively. The purple labels indicate each data point.

If you compare this with the human data shown in Figure 1.3, you will see that the model reproduces all of the important characteristics of the human data as described previously: interference in the conflict condition of color naming, the imperviousness of word reading to different conditions, and the overall slowing of color naming.

Now, we can single-step through the testing events to get a better sense of what is going on.

First, be sure you are viewing the network (.T3Tab.StroopNet), and that you are viewing activations. Then, do a Test Init, Step.

Each StepTest of the process will advance one step through the three conditions of word reading in order (control, conflict, congruent) followed by the same for color naming. For the word reading conditions, you should observe that the corresponding word reading hidden unit is rapidly activated, and that this then activates the corresponding output unit, with little effect of the color pathway inputs. The critical condition of interest is the conflict color naming condition.

Question 11.2 (a) Describe what happens in the network during the conflict color naming condition, paying particular attention to the activations of the hidden units. (b) Explain how this leads to the observed slowing of response time (settling).

SOA Timing Data

Another important set of data for the model to account for are the effects of differential stimulus onset times discussed previously (Glaser and Glaser, 1982). To implement this test in the model, we simply present one stimulus for a specified number of cycles, and then add the other stimulus and measure the final response time (relative to the onset of the second stimulus). We use five different SOA (stimulus onset asynchrony) values covering a range of 20 cycles on either side of the simultaneous condition. For word reading, color starts out preceding the word by 20 cycles, then 16, 12, 8, and 4 cycles (all indicated by negative SOA), then color and word are presented simultaneously as in standard Stroop (0 SOA), and finally word precedes color by 4, 8, 12, 16, and 20 cycles (positive SOA). Similarly, for color naming, word initially precedes color (negative SOA), then word and color are presented simultaneously (0 SOA), and finally color precedes word (positive SOA). To simplify the simulation, we run only the most important conditions -- conflict and congruent.

The graph in the right panel should display the response time as a function of SOA on the X axis. You should see something like Figure 11.7 in the text when the test is complete -- because it is not possible to label the lines in the simulation, you should rely on the figure to decode what is going on. The two solid lines represent the color naming conflict and congruent conditions. The two dotted lines represent the word reading conflict and congruent conditions -- these lines basically fall on top of one another and so are not labeled separately, but the conflict condition line is slightly higher at the earliest SOA's.

By comparing the simulation data with the human data shown in Figure 11.4, you can see that the model's performance shows both commonalities and contrasts with the behavioral data. We first consider the commonalities. The model simulates several important features of the behavioral data. Most importantly, the model shows that word reading is relatively impervious to color conditions (conflict vs. congruent), even when the colors precede the word inputs, as indicated by the similarity of the two dotted lines in the graph. Thus, the dominant effect in the model is a strength-based competition -- the word reading pathway is sufficiently strong that even when it comes on later, it is relatively unaffected by competition from the weaker color naming pathway.

Another important feature of the human data captured by the model is the elimination of the interference effect of words on color naming when the color precedes the word by a relatively long time (right hand side of the graph). Thus, if the color pathway is given enough time to build up activation, it can drive the response without being affected by the word.

There are two important differences between the model and the human data, however. One difference is that processing is relatively slowed across all conditions as the two inputs get closer to being presented simultaneously. This is particularly evident in the two word reading conditions and in the congruent color naming condition, in the upward slope from -20 to 0 SOA, followed by a downward slope from 0 to 20. This effect can be attributed to the effects of competition -- when inputs are presented together, they compete with one another and thus slow processing. This may be an artifact of the kWTA form of competition, as it was not found in the original Cohen, Dunbar, and McClelland (1990) model.

Another difference, which was present in that model, is the increasingly large interference effect for earlier word SOA's on color naming in the model, but not in people. It appears that people are somehow able to reduce the word activation if it appears sufficiently early, thereby minimizing its interfering effects. Cohen, Dunbar, and McClelland (1990) suggested that people might be habituating to the word when it is presented early, reducing its influence. However, this explanation appears unlikely given that the effects of the early word presentation are minimal even when the word is presented only 100 msec early, allowing little time for habituation. Further, this model and other models still fail to replicate the minimal effects of early word presentation even when habituation (accommodation) is added to the models.

An alternative possibility is that the minimal effects of early word presentation reflect a strategic use of perceptual (spatially mediated?) attentional mechanisms (like those explored in chapter 8) that can be engaged after identifying the stimulus as a word. According to this account, once the word has been identified as such, it can be actively ignored, reducing its impact. Such mechanisms would not work when both stimuli are presented together because there would not be enough time to isolate the word without also processing its color.

Effects of Frontal Damage

Now that we have seen that the model accounts for several important aspects of the normal data, we can assess the importance of the prefrontal (PFC) task units in the model by weakening their contribution to biasing the posterior processing pathways (i.e., the hidden layer units in the model). The strength of this contribution can be manipulated using a weight scaling "wt_scale" parameter for the connections from the PFC to the Hidden layer. This parameter is shown as thte first row of the overall control panel, with a default value of .8 (the right value, in the "rel" field). Because the model is relatively sensitive, we only need to reduce this value to .75 to see an effect. Note that this reduction in the impact of the PFC units is functionally equivalent to the gain manipulation performed by Cohen & Servan-Schreiber (1992).

  • Reduce the PFC to Hidden wt_scale.rel parameter in the control panel from 0.8 to 0.75, and then view .T3Tab.TrialOutputData, and do Test: Init, Run.

You should see that the model is now much slower for the conflict color naming condition, but not for any of the other conditions. This is exactly the same pattern of data observed in frontal and schizophrenic patient populations Cohen & Servan-Schreiber (1992). Thus, we can see that the top-down activation coming from the PFC task units is specifically important for the controlled-processing necessary to overcome the prepotent word reading response. Note that to fit the model to the actual patient response times, one must adjust for overall slowing effects that are not present in the model (see chapter 8 for a discussion of how to compare model and patient data).

Although we have shown that reducing the PFC gain can produce the characteristic behavior of frontal patients and schizophrenics, it is still possible that other manipulations could cause this same pattern of behavior without specifically affecting the PFC. In other words, the observed behavior may not be particularly diagnostic of PFC deficits. For example, one typical side effect of neurological damage is that overall processing is slower -- what if this overall slowing had a differential impact on the color naming conflict condition? To test this possibility in the model, let's reduce the dt_vm parameter in the control panel, which determines the overall rate of settling in the model.

  • Restore the PFC to Hidden wt_scale.rel field to 0.8, and then reduce the TestUnit dt.vm from 0.01 to 0.008. Make sure you are still viewing .T3Tab.TrialOutputData, and do Test: Init, Run.

Question 11.3 Compare the results of this overall slowing manipulation to the PFC gain manipulation performed previously. Does slowing also produce the characteristic behavior seen in frontal and schizophrenic patients?

Retrieved from "http://grey.colorado.edu/CompCogNeuro/index.php/CECN1_Stroop"
Personal tools