5.3 Target processing in skeletal presentation
5.3.3 Modelling skeletal presentation
The ST2 model as published in Bowman and Wyble (2007) and described in Section 3.3 cannot simulate skeletal presentation. In the following, we will show how, by making a number of theoretically justified changes to the architecture of the model, we can replicate our experimental results with respect to skeletal presentation in both the behavioural and the EEG domain.
Step 1: Setting distractor values to zero
Manipulation In skeletal presentation, the presentation stream contains just the target and the distractor following the target. All other distractors are removed and replaced by blank intervals. In order to simulate such a stream in the ST2 model, we modify the array of values that serves as input to the model. All distractors - except the one following the target - are set to a value of zero, which corresponds to no activation.
Results The modification of the input array has a strong effect on virtual ERP traces resembling early visual processing (see Section 4.5.1 for a corresponding description of vERP methodology). For targets in RSVP, the model shows a continuous virtual ssVEP wave oscillating at the frequency of target presentation (Figure 29), hence, the model replicates the human data from Figure 28. The first item of the RSVP stream causes an increase of activation in early layers of the model. As the following items appear in rapid succession, activation in these layers does not decay back to baseline. Rather, the inhibition between items in the masking layer causes layer activation to oscillate around a certain value until the end of the RSVP stream.
In skeletal presentation, there are no distractors and hence there is no activation pre- ceding the target. When the target is presented, this creates a strong burst of activation at early layers of the ST2model. As there is no forward masking (i.e. the target representation is not inhibited by distractors preceding the target), the activation evoked by the skeletal target at early layers is higher than in regular RSVP. The distractor following the target in
.000 .010 .020 .030 .040 .050 .060 .070 -200 0 200 400 600 800 1000 1200 Activation ms equivalent RSVP Skeletal T
Figure 29 A virtual ssVEP wave for the RSVP and vERP early components for the skeletal con- dition from input and masking layer of the ST2model. ‘T’ indicates the presentation of the target
and ERPs are time locked to presentation of the target.
skeletal presentation then produces a second large burst of activation, as its activation at early layers is not constrained by backward masking. All of this activation at early layers occurs between the model equivalent of 100 and 200ms following target presentation. There is thus a qualitative match between the virtual ERP from the ST2 model (Figure 29) and the human ERP early components (P1/N1 wave) for skeletal presentation from Figure 28. To summarise, virtual ERP activation associated with early visual processing shows a distinct activation for skeletal targets and an oscillatory pattern for RSVP targets, thus qualitatively replicating the human ERP. Furthermore, the timing of the skeletal vERP activation occurs within a similar time window as the P1/N1 wave observed for skeletal targets in the human ERP.
Step 2: Moving the blaster ‘trigger’ to masking layer
Manipulation The adjustment of the input array for skeletal presentation has the desired effect on virtual ERP activation resembling early visual processing. A replication of be- havioural accuracy and the virtual P3 component, however, requires theoretically justified
Blaster circuit Input layer Masking layer Item layer Task filtered layer (TFL) Task Demand Blaster in Blaster output Stage 1 STEP 2 Blaster 'trigger' for simulation of RSVP skeletal presentation weight 0.02003 weight 0.015 Stage 2 Excitatory Inhibitory
Figure 30 Step 2 of simulating skeletal presentation. As indicated in the figure, the connection from stage one that triggers the blaster is moved from the task filtered layer to the masking layer.
changes to the architecture of the ST2 model.
As discussed in the introduction, skeletal targets appear as a visual onset on a previously blank screen, whereas in RSVP, the target has to be selected from a continuous stream of distractors. In terms of the ST2 model, we hypothesise that this influences the way in which
the blaster is triggered:
• In RSVP, the system cannot distinguish targets from distractors until they have reached the task filtered layer (TFL). In the TFL, the task demand mechanism selects targets by means of selective excitation to target nodes and inhibition to distractor nodes. • In skeletal presentation, there are no distractors preceding the target, hence, the system can assume that the first item that is ‘presented’ to the input layer is the target. Accordingly and as seen in Figure 30, we propose that in skeletal presentation the blaster is triggered as soon as activation reaches the masking layer2. Moving the connection to the
2
For the purpose of simulating the skeletal paradigm (i.e. with no distractors preceding the target), our manipulation produces the desired effect. However, our modification of the model architecture would have to be reconsidered in order to simulate a slightly different stream setup, for instance, if the target was also preceded by distractor items (e.g. a stream of the type ‘D D T D’). Under these circumstances, the distractors can potentially also fire the blaster, as task demand does not operate until the TFL and, hence, the system cannot distinguish targets from distractors at the masking layer. Note, however, that although distractors can fire the blaster in skeletal presentation, task demand at the TFL will prevent distractors from being tokenised.
blaster from TFL to the masking layer also requires a modification of the weight value of that connection (see Figure 30). This is a technical requirement that is necessary because of the way the ST2 model is implemented3.
.000 .050 .100 .150 .200 -200 0 200 400 600 800 1000 1200 Activation ms equivalent RSVP Skeletal T
Figure 31 After Step 2: Virtual P3 for the RSVP and skeletal condition. The RSVP vERP is baseline corrected to -200 to 0ms with respect to target onset to account for distractor related activity, which is absent in the skeletal RSVP condition. ‘T’ indicates the presentation of the target and ERPs are time locked to presentation of the target.
Results As activation propagates through the ST2model, there is temporal delay. Hence, if the blaster is triggered from the masking layer, the blaster fires at an earlier time point than if activation has to propagate to the TFL before the blaster can be triggered. Conse- quently, the blaster’s output to item layer and TFL also occurs earlier in time.
The first consequence of this change is a shift in latency of the virtual P3 (see Sec- tion 4.5.2 for a corresponding description of vERP methodology) for skeletal compared to RSVP targets, as seen in Figure 31. In RSVP, the target reaches the TFL and triggers the blaster once the target has been identified as such by the task demand mechanism. Once the blaster becomes active, it can enhance the target for tokenisation. With the change
3
Compared to the TFL, activation values at the masking layer are higher in absolute terms. Hence, we reduce the weight values between masking layer and blaster to prevent the blaster circuit from being overcharged by the input from the masking layer.
in model architecture for skeletal presentation, the blaster is triggered earlier, thus, it is active and ready when targets reach the TFL. The consequence is an earlier tokenisation (and virtual P3) for skeletal targets.
The change in model architecture means that the blaster will now fire for all skeletal targets, regardless of their input strength. This increases the accuracy of the ST2 model at detecting skeletal targets. Whereas RSVP targets have an average simulated accuracy of 77%, the earlier blaster response caused by the modification of the ST2 model’s architecture produces a simulated skeletal accuracy of 100%. Although skeletal behavioural accuracy should indeed be above RSVP accuracy, this is not a good replication of the human be- havioural performance for detecting skeletal targets, which is below ceiling. Furthermore, the virtual P3 lacks the distinctive difference in size between skeletal and RSVP targets that is evident in the human P3 component.
A further modification of the weight value between masking layer and blaster does not have the desired effect on simulated accuracy and virtual P3 for skeletal targets. This is due to the blaster ‘trigger’ functioning in an ‘all-or-none’ fashion, hence, the weight value would have to be reduced to close to zero before there is any further effect on the target’s tokenisation process. Reducing the weight to close to zero, however, has a counterproductive effect as, in this case, the blaster can only be triggered by those targets with the highest strength values. Consequently, only a few targets are tokenised and all other targets are not ‘detected’ by the model. This reduces the simulated skeletal accuracy to below that for RSVP targets, which is obviously not a desirable replication of the human data either. Consequently, we need to perform one additional modification to the architecture of the ST2 model (as described in the following section) in order to accurately simulate skeletal presentation.
Step 3: Decreasing the strength of the task demand mechanism
Manipulation An RSVP stream consists of one or more targets embedded in a stream of distractors. In skeletal, however, the stream contains only the target and the following distractor. When an RSVP target arrives at the TFL, the task demand mechanism plays a vital role in selecting the target from simultaneously active distractors. In skeletal presen- tation, however, the target competes with only one other distractor at the TFL and hence
Blaster circuit Input layer Masking layer Item layer Task filtered layer (TFL) Task demand Blaster in Blaster output Stage 1 Stage 2 Reduce weight by 0.92% STEP 3 Excitatory Inhibitory
Figure 32 Step 3 of simulating skeletal presentation. As indicated in the figure, the weight of the connection from task demand to target nodes in the TFL is reduced by 0.92%.
there is no need for the task demand mechanism to be as strong. Conceptually, in skeletal, the focus of selection moves earlier and reducing the strength of the task filter reflects this adjustment of focus. In other words, since the system can select earlier with skeletal, its later selection mechanism (at the TFL) can be more liberal. Consequently, in our second manipulation to the architecture of the ST2 model, we reduce the weight from task demand to target nodes in the TFL by 0.92% of the original value (see Figure 32).
Results The reduction in task demand for skeletal presentation means that target nodes have less activation at the TFL. Relatively strong targets can nevertheless initiate a tokeni- sation process despite lower activation levels. Weak targets, however, fail to overcome the threshold for tokenisation and cannot proceed into stage two for working memory encoding. As seen in Figure 33, weakening task demand in skeletal presentation reduces the size of the virtual P3 for skeletal targets. As the virtual P3 for skeletal targets is now considerably smaller than for RSVP targets, this qualitatively replicates the human ERP data. The skeletal virtual P3 with weaker task demand (Figure 33) also lasts slightly longer than the skeletal virtual P3 from Figure 31. This is because the reduction in task demand reduces
.000 .050 .100 .150 .200 -200 0 200 400 600 800 1000 1200 Activation ms equivalent RSVP Skeletal T
Figure 33 After Step 3: Virtual P3 for the RSVP and skeletal condition. The RSVP vERP is baseline corrected to -200 to 0ms with respect to target onset to account for distractor related activity, which is absent in the skeletal RSVP condition. ‘T’ indicates the presentation of the target and ERPs time locked to presentation of the target.
the activation strength of targets in stage two. Less target activation in stage two slows down the tokenisation process, which prolongs the virtual P3 component.
After this second modification to the model’s architecture, weak skeletal targets have too little activation for tokenisation and are ‘missed’. The ST2 model now generates a simulated accuracy of 85% for skeletal targets, which replicates the human behavioural accuracy for skeletal targets. Simulated RSVP accuracy from the model obviously remains unchanged at 77%.