Although much of the detailed behavioral data we will explore with the model concerns the Posner spatial cueing task, we think the more basic functional motivation for visual attention is to facilitate object recognition when multiple objects are presented simultaneously. Therefore, we will start with a quick exploration of the network's object recognition capacities as a function of the spatial distribution of two objects. This will provide an introduction to the kinds of interactions between spatial and object processing that can happen using this relatively simple model. Let's begin by viewing the events that we will present to the network.

• Click on the MultiObjs button to view the events (the one that is below the Test line that also says MultiObjs)a.

You should see 3 events. The first event has two different objects (features) present in different spatial locations. Note that the target object has slightly higher activation (i.e., it is more salient), which will result in the reliable selection of this object over the other. The next event has two of the same objects (targets) presented in different locations. Finally, the last event has the two different objects in the same spatial location. As the figure makes clear, recognizing objects when they overlap in the same location is considerably more difficult than when they appear in different locations. Although it is clearly easier to recognize objects if only copies of the same object are present as opposed to different objects, this difference is not likely to be significant for small numbers of presented objects.

Now, let's test these predictions in the model.

• Switch back to viewing Act in the NetView. Do Init and Test Trial.

This will present the first event to the Network, which will stop settling (i.e., updating the network's activations a cycle at a time) when the target unit's activation exceeds the threshold of .5 in the Output layer. You should see that the network relatively quickly focuses its spatial attention on the more active input on the right side, and the Object pathway represents that target item, causing the Output activity to get over threshold relatively quickly.

• Then Test Trial through the remaining events.

You should have seen that while the network settled relatively quickly for the first two events, it was slowed on the third event where the objects overlap in the same region of space (occasionally not so slow on the last one).

Click on the TstTrlPlot to see a plot of the settling times (number of cycles to reach threshold) for each event.

You can also set the ViewUpdt setting to Cycle instead of FastSpike and, click back on the NetView, and Test Trial back through the items, to see a cycle-by-cycle update of the network. You can use the VCR rewind buttons at the lower right of the NetView to rewind through and see exactly how the network settling played out.

To get a better sense of the overall data pattern, click back on TstTrlPlot and do Test All a few times. There is a small amount of noise so the results should be a little bit different each time, but overall quite consistent.

You should see that overall the network has more difficulty with the objects appearing in the same spatial location, where spatial attention cannot help focus on one object. The overall average cycles by condition are reported in the TstStats table -- click the etable.Table button there to answer this question:

Question 6.6: Report the Cycle:Mean values for each condition from the TstStats table.

You should have observed that spatial representations can facilitate the processing of objects by allocating attention to one object over another. The key contrast condition is when both objects lie in the same location, so that spatial attention can no longer separate them out, leaving the object pathway to try to process both objects simultaneously.

As mentioned previously, additional lesion data comes from Balint's syndrome patients, who suffered from bilateral parietal lesions. The most striking feature of these patients is that they have simultanagnosia -- the inability to recognize multiple objects presented simultaneously (see Farah, 1990 for a review). Interestingly, when such subjects were tested on the Posner task (Coslett & Saffran, 1991), they exhibited a decreased level of attentional effects (i.e., a smaller invalid-valid difference). As emphasized by Cohen et al. (1994), these data provide an important argument against the disengage explanation of parietal function offered by Posner and colleagues, which would instead predict bilateral slowing for invalid trials (i.e., difficulty disengaging). The observed pattern of data falls naturally out of the model we have been exploring.

• To simulate this condition, first set Test back to StdPosner, and then do Lesion with Locations set to LesionFull instead of Half (keep Units at LesionHalf). Do Init, Test Trial while watching the NetView, and then Test All a few times while looking at the TstTrlPlot. Then click on TstStats table.

Question 6.10: Report the results of the TstStats for the bilaterally lesioned network.

Another interesting aspect of the Posner spatial cuing task has to do with the temporal dynamics of the attentional cuing effect. To this point, we have ignored these aspects of the task by assuming that the cue activation persists to the point of target onset. This corresponds to experimental conditions when the target follows the cue after a relatively short delay (e.g., around 100 ms). However, the Posner task has also been run with longer delays between the cue and the target (e.g., 500 ms), with some interesting results. Instead of a facilitation effect for the valid trials relative to the invalid ones, the ordering of valid and invalid trials actually reverses at the long delays (Maylor, 1985). This phenomenon has been labeled inhibition of return, to denote the idea that there is something that inhibits the system from returning attention to the cued location after a sufficient delay.

Our model can be used to simulate at least the qualitative patterns of behavior on the Posner task over different delays. This is done by varying the length of cue presentation (a variable delay event could have been inserted, but residual activation would persist anyway, and varying the cue length is simpler), and turning on the sodium-gated potassium (KNa) adaptation current, which causes neurons that have been active for a while to "fatigue". Thus, if the cue activation persists for long enough, those spatial representations will become fatigued, and if attention is subsequently directed there, the network will actually be slower to respond. Also, because the spatial activations have fatigued, they no longer compete with the activation of the other location for the invalidly cued trials, eliminating the slowing.

Now, let's see this in the model.

• First, un-lesion the network by running Lesion and selecting NoLesion for Layers. Next, set Test to StdPosner. Then, click the KNaAdapt toggle to On to turn on adaptation. Next, let's choose a cue duration that, even with the accommodation channels active, still produces the original pattern of results. Set CueDur to 50. Now, do Init, Test Trial and Test All as usual.

You should observe the now-familiar pattern of a valid facilitation and an invalid slowing (although a bit weaker).

• Test All with increasing durations (change using the CueDur field) in increments of 50 from 50 to 300 or higher.

You should see that the valid-invalid difference decreases progressively with increasing duration, and ultimately, the validly cued condition can actually be a bit slower than the invalidly cued one, which is the hallmark of the inhibition of return phenomenon (Figure 8.28). The effect sizes here are fairly small because the form of adaptation here is relatively weak -- a more significant GABA-B like delayed inhibition effect (which is not currently implemented) would be needed to produce more substantial effects.

• Switch to the NetView, set ViewUpdt to Cycle and Test Trial through the running of the network with CueDur at 300 or higher.

Question 6.11: Report in detail what happens on the valid and invalid trials that produces the inhibition of return effect. It is useful to observe the activation (or lack thereof) of the various layers as the cue duration increases. While you should see changes in the ranges of durations specified, you may have to increase the cue duration even more to get the full inhibition of return effect.