1.8 Overview of studies and models presented in the thesis
1.8.2 Computational models
Chapter 5 to Chapter 8 report on computational models of the empirical findings in bimodal CMTS (described in Chapter 3). We chose the behavioural results in bimodal CMTS to be our main focus for the modelling work because it encompassed the largest number of different experimental conditions investigated in the current thesis—pure trials, repetition trials, switch trials, modality-shift trials, modality-repetition trials, and asymmetric pathways (i.e.
modality-response compatibility). These different experimental conditions allowed us to explore information processes relevant to those trials.
Chapter 5 includes the bulk of the details on how our models were adapted from Gilbert and Shallice’s (2002) original task-switching model. The models in Chapter 5 were the simplest in architectural complexity, and were designed closely to the models described by Gilbert and Shallice. The model architecture involved both fixed feedforward connections and temporary priming connections between the stimulus and the task attribute. This chapter introduced different network ages (Young, Middle and Old) that were architecturally identical but different in connection weights. The models in Chapter 5 captured RT mixing costs, RT switch costs, and RT priming costs on switch trials, but failed to produce the correct error profiles. Moreover, all RT costs were substantially larger in the younger networks than in either the Middle or Old networks.
Chapter 6 explores other reactive mechanisms that might be particularly relevant to our behavioural study. The first mechanism is involuntary reactive task retrieval. Tasks may be retrieved reactively without the mediation of top-down control since both task attribute and the stimuli themselves are closely aligned in the representational space (e.g. a dog picture on a ‘dog detection’
task). This reactive mechanism is modelled by introducing additional fixed connections between the stimulus units and the corresponding task attribute units. The second mechanism is involuntary reactive response from the previous primed stimulus-response association. This reactive response was modelled by introducing priming connections between the stimulus units and the response units at the end of the trial. The models in Chapter 6 were successful at reducing RT switch costs and increasing errors in younger networks, but these effects (i.e. cost reduction and error increase) soon achieved an asymptotic level. The simulation results were still some way away from the observed behavioural data.
Chapter 7 introduces additional decay functions and additional parameters that simulated individual differences between networks. The models assume that the activity to both task relevant and irrelevant attributes starts to decay once the task cue disappears—task-relevant attribute becomes less excited, and task-irrelevant attribute becomes less inhibited. Although task attributes decayed after stimulus onset, the networks could probabilistically re-update the task attribute units during the stage of response settling. This intermittent task update is conceptually equivalent to verbal self-reminder. The models assume that the variation of update probability was a matter of individual difference, and not of between-age difference. The chapter also explores whether different ages might be more or less likely to employ a reactive task strategy or a proactive task strategy, which was defined by how the participants modulated their control by trial types. In the models, task strategies were implemented by varying the top-down signals within a network.
The addition of decay, update probability and task strategies allowed the models to generate greater RT and accuracy variations on all trial types. The
models in Chapter 7 explore how the empirical data, which contained various forms of inter-group differences and individual differences, can be captured with the existing model architecture.
In contrast to the previous models that explore task-associated effects, Chapter 8 presents two different models that explore the modality shift effect (MSE), as well as the observed asymmetry of MSE to visual and auditory targets. The two models were based on different theoretical accounts for MSE—a model based on priming effects (priming model) and a model based on carry-over effects from additional modality attribute representations. Both models were able to produce a MSE as well as asymmetric MSE patterns in the networks that simulated the pure task blocks. However, only the model with modality attribute units was able to capture the effect with statistical equivalence to the observed behavioural data in the pure blocks. None of the models were able to capture the asymmetric MSE patterns in the mixed task blocks. The model results suggested that MSE is likely to be a composite effect of multiple processes and could, therefore, change with the information processes involved in the overall task context.
Finally, Chapter 9 concludes with a general discussion and with suggestions for future research. Specifically, the discussion will focus on the empirical findings relating to between-condition mixing/switch costs as well as inter-group differences in overall accuracies. This will be followed by a discussion of the empirical findings on modality dominance and modality shift effects, and whether processes related to modality-specific representations were part of an overall task set. These behavioural findings will be explained by key mechanisms from the computational models; namely, goal activation, carry-over effects from task-associated representations, and carry-carry-over effect from
modality-associated representation. The discussion will also highlight the potential pitfall in defining a task set from the ideal external specification without probing into the internal representation of the task set. Any discrepancy between the desired and actual representation of a task set can generate very different behaviour patterns, particularly in younger participants. Lastly I will talk about modality shift effects and why these effects can be elusive and are particularly context sensitive. I will also highlight the limitation of the models in the thesis, and other additional mechanisms that are worth considering. Finally, the thesis will end with the suggestion that future developmental studies should address the developmental and individual differences in stable performance, in addition to flexible behaviours.