Chapter 6. Reactive Mechanisms (Model 2)
6.1 Model 2A: Reactive Task Retrieval
Younger children are more likely to switch mental representations reactively after detecting change signals, whereas older children and adults are more likely to utilise informative cues in advance of the stimulus onset (Chevalier et al., 2015; Doebel et al., 2017; Morton & Munakata, 2002). For example, transparent task cues have been shown to be particularly helpful for young children, as compared to arbitrary task cues (Blaye & Chevalier, 2011).
This may be due to the fact that a task can be triggered more reactively when the task relevant information is already present in the transparent (i.e. strongly associated) task cue. In our behavioural experiments, both cues and stimuli were fairly transparent and strongly associated with the task attributes (e.g.
animal categories). Thus, it is likely that both did the task cue, and the stimuli (e.g. seeing the ‘Dog’ stimulus reminded the participant to engage in the ‘Dog’
1 All parameters setting used in the simulations in Chapter 6 are listed in Table 6.1 at the end of chapter.
task) activated the task attribute. Younger participants may have relied more on this reactive task set trigger than adults.
Therefore, Model 2A explores possible mechanisms that activate task attributes other than the Top signal. In particular, Model 2A investigated whether young networks benefit from reactive activation of task attribute representations, even when their Top signal was relatively low. Specifically, we aimed to understand whether reactive task retrieval can be effective in reducing RT switch costs in younger networks.
In the next section, I will begin by describing the architecture of Model 2A in detail.
6.1.1 Model 2A: PS to TA Reactive Pathway
To model the reactive associations that exist between task representation and the percept stimulus, additional fixed feedforward connections between PS units and the associated TA units are introduced in Model 2A (e.g. DV and DA to TADOG, BA and BV to TABIRD, see Fig. 6.1). The additional connections are not expected to have an effect on pure trials since there is only one TA unit. On switch trials in the mixed networks, the PS-TA connections should be facilitative in activating the task-relevant TA unit when the target is present, especially when the Top signal is small, but it can also introduce competition when a nontarget is present on the switch trial. On repetition trials, the cost/benefit may be more limited since the task-relevant TA unit is likely to be highly excited and the task-irrelevant TA unit highly inhibited.
Unlike other fixed feedforward connections where there are developmental differences in connection weights, due to the reactive nature of the processes, the PS-TA connections are the same across all network ages.
There are additional assumptions built into Model 2A with regards to the developmental differences between networks. It is assumed that younger networks will have smaller (but sufficient) overall Top signal than older networks (TopOld=10, TopMiddle=7, TopYoung=5). Finally, in this section, there were 25 network subjects of each network age. The number of trials was the same as in Model 1 (Chapter 5), and the targets appeared with 60% probability.
Figure 6.1. Model 2A Architecture with additional Percept Stimulus to Task
Attribute feedforward connections (PS-TA). The new connection from the previous
model is highlighted in red.
6.1.2 Model 2A: Result
6.1.2.1 Reaction time and accuracy
The additional PS-TA connections in Model 2A reduced the RT on switch trials at all network ages, as compared to networks without the PS-TA connections (see Fig. 6.2). As expected, the PS-TA connections had no effect
on overall RT and accuracy on pure and repetition trials in any of the network ages. This is due to the strong existing TA activation on pure and repetition trials, so the additional input from the PS-TA reactive pathway adds little to the settling process.
On switch trials, there was a reduction of RT in Middle and Young networks, but not in Old networks. As the Top signals were smaller in Middle and Young networks, the relative contribution from PS-TA inputs was larger in these networks, as compared to Old networks. However, their faster responses on switch trials was accompanied by a reduction in accuracy on switch trials, indicating that relying on the bottom-up activation of task representation is not the optimal task approach.
Figure 6.2. Mean RT and accuracy in with and without PS-TA reactive pathways
across different network ages. Orange bars show networks without PS-TA pathways, and
yellow bars show networks with PS-TA pathways. Left to right: Old, Middle and Young
networks. Error bars represent 95% CI of means.
6.1.2.2 Between-condition RT costs
Without the PS to TA reactive connections, the developmental differences in RT switch costs were fairly substantial. The additional PS to TA connections reduced RT on switch trials in the younger networks, thereby
reducing the developmental differences in switch costs (Fig. 6.3). However, the reactive pathway did not eliminate the effect of development on RT switch costs.
Figure 6.3 Mean RT Mixing Cost and RT Switch Cost. Left panel shows costs in
networks without PS-TA connections; right panel shows costs in networks with PS-TA
connections (weight=3). Different bars represent different network ages. Error bars
represent 95% CI of means.
6.1.3. Model 2A: Discussion
Model 2A simulations explored the idea that task attribute representations can be reactively activated upon seeing the stimulus, as well as through an internal biasing mechanism such as Top signals. The reactive pathway through PS-TA connections can be both facilitative of and obstructive to task performance. The facilitative effect was observed with smaller RT on switch trials with PS-TA connections than those without, particularly when the Top signal was small. However, PS-TA input can also have a detrimental effect on accuracy. The accuracy cost was caused by the inability of the Top signal to override the reactive signal when the task-irrelevant target was present.
To model the effect of development, we made a few assumptions in terms of the Top signals of different network ages. Older networks were assumed to have more efficient top-down control, thus a stronger Top signal than younger networks. With these additional assumptions, it was found that both Middle and Young networks benefited from the additional PS-TA input on switch trials, showing faster responses on switch trials. In contrast, Old networks (i.e. larger Top input) did not display such a benefit.
Model 2A was able to reduce RT switch costs and induced errors on switch trials in Middle and Young networks. This model was therefore an improvement from Model 1 (Chapter 5). However, despite these improvements, the limitations in Model 1 remained in Model 2A; namely, the reversed developmental effect on RT switch costs and the floor effects on accuracy in all trial types.
In the next section, we will consider another reactive mechanism—
namely, reactive primed responses.