[PS-1.15] When failures of models are informative: The case of frontostriatal mechanisms in the development of cognitive control

Chatham, C. 1 , Yerys, B. 2 & Munakata, Y. 3

1 Brown University
2 George Washington University
3 University of Colorado Boulder

Computational models are powerful tools. However, the idea that models can ?do anything? is wrong, and their failures have been informative. We present an investigation along these lines, which motivates the inclusion of frontostriatal updating mechanisms in models of the development of cognitive control. Children show remarkable limitations in cognitive control. For example, after sorting cards according to one rule (e.g., based on shape), children often perseverate, continuing to sort cards by this rule even after
being instructed to switch to a new rule (e.g., color). Existing models have simulated such limitations in cognitive control, and children's eventual successes, in terms of a gradually increasing ability to actively maintain task-relevant information (e.g., the currently relevant rule) in working memory. However, these models have not investigated the role of feedback in children's cognitive control. Here, we first present new behavioral work showing that such feedback leads some children to play an ?opposites? game -- they stick with the first rule, but learn to reverse stimulus-response mappings by putting truck cards in the flower pile, for example. We next
present simulations that demonstrate the failure of an existing neural network model to capture these failures of children. Specifically, playing the opposites game requires a robust representation of the old rule, rather than simply a weak representation of the new rule, for error-driven learning to remap the features to the responses. Simulating this pattern motivates the inclusion of updating mechanisms that gate information into working memory. These models simulate the opposites-game when these updating mechanisms are relatively inefficient, in combination with relatively robust working memory maintenance. These models make contact with a growing literature on frontostriatal function in higher level cognition, and argue for the importance of such mechanisms in understanding the development of cognitive control.