Botvinick, M., & Plaunt, D. C. (2002). Representing task context: proposals based on a connectionist model of action. Psychological Research, 66, 298-311. DOI 10.1007/s00426-002-0103-8
Representations of task context play a crucial role in shaping human behavior. While the nature of these representations remains poorly understood, existing theories share a number of basic assumptions. One of these is that task representations are discrete, independent, and non-overlapping. We present here an alternative view, according to which task representations are instead viewed as graded, distributed patterns occupying a shared, continuous representational space. In recent work, we have implemented this view in a computational model of routine sequential action. In the present article, we focus specifically on this model’s implications for understanding task representation, considering the implications of the account for two influential concepts: (1) cognitive underspecification, the idea that task representations may be imprecise or vague, especially in contexts where errors occur, and (2) information-sharing, the idea that closely related operations rely on common sets of internal representations.
Most models (e.g. Cooper and Shallice) posit discrete, independent internal representaitons; Botvinick and Plaut propose, instead, that representations grade into one another, and share multidimensional, conceptual space.
The model is a typical connectionist model, with input nodes (sensation/perception), hidden nodes (interneurons), and output nodes (motor operations on the environment). Activation in the hidden nodes can cycle back onto those nodes, making theirs a “recurrent” connectionist model. When trained to make coffee (a la Cooper and Shallice), minor perturbations to the hidden layer produced behavioral slips (correct behavioral sequences in incorrect contexts); major perturbations appeared similar to action disorganization syndrome.
First proposal: “task representations can be imprecise or underspecified.” Action slips can occur when random jitter (a simulation of degraded representation) deflects the current activity from the target activity into the space it shares with competitor actions.
Frequency of experience moderates action slips. The less frequently trained the behavior, the smaller its stake in multidimensional space. The representation for items infrequently trained is therefore more fragile — even minor perturbations may result in a transition into a more practiced (and therefore larger area-holding) routines.
Traditional models of action representations might seem to incorporate underspecification, too, in that any of a number of hierchies could be activated in a given context. However, these models would seem to hold that any hierarchy could be substituted in place of any other (originating at the same level of abstraction)–there’s no sense of the neighborhood of representations offered by the connectionist model.
Second proposal: “task representations may be ‘shared’ by multiple, structurally interrelated activities.” A connectionist model allows similar tasks to share features, and aids generalization to novel, but related actions. The model applied what it knew about adding sugar to coffee, to add cocoa mix to water.
The biggest advantage of such a model may be an an account of learning. When faced with novelty (e.g. altering “a scoop of cocoa” to become “a BIG scoop of cocoa), discrete symbolic models have to implement a categorical decision, some threshold below which the action is carried out according to the old representation and above which a new representation is composed anew. The connectionist model allows for continuous remodeling of representations, gradually differentiating new from old.
Traditional models of action representations might seem to employ information-sharing, in that supraordinate nodes can share subordinate nodes. But the models seem to posit that these subnodes are identical, the actions they represent unchanged as we shift from one context to another.
The simulation aligned with empirical evidence that showed well learned representations occupy more multidimensional space. So, well learned routines are more robust to violations; while more weakly represented routines are more likely to drift into the gravity well stronger representations. Does this square with the talk last week about highly frequent words being at the forefront of language change?
The connectionist model of action here requires fairly specific feedback. The model receives information about the intended outcome, calculates some deviation score based on this and the product of the output nodes, and back-propagates some correction factor through the network. How explicit must this feedback be? How often do we receive informative-enough negative feedback about our language learning?
Can gradualist models like this account for “ah ha” moments of insight, like those documented by Kohler? Is violation of expectation possible when brand new situations present themselves? Is there even such a thing as a “brand new situation” in a connectionist network?