On the Rational Boundedness of Cognitive Control: Shared Versus Separated Representations

Abstract

One of the most fundamental and striking limitations of human cognition appears to be a constraint in the number of control-dependent processes that can be executed at one time. This constraint motivates one of the most influential tenets of cognitive psychology: that cognitive control relies on a central, limited capacity processing mechanism that imposes a seriality constraint on processing. Here we provide a formally explicit challenge to this view. We argue that the causality is reversed: the constraints on control-dependent behavior reflect a rational bound that control mechanisms impose on processing, to prevent processing interference that arises if two or more tasks engage the same resource to be executed. We use both mathematical and numerical analyses of shared representations in neural network architectures to articulate the theory, and demonstrate its ability to explain a wide range of phenomena associated with control-dependent behavior. Furthermore, we argue that the need for control, arising from the shared use of the same resources by different tasks, reflects the optimization of a fundamental tradeoff intrinsic to network architectures: the increase in learning efficacy associated with the use of shared representations, versus the efficiency of parallel processing (i.e., multitasking) associated with task-dedicated representations. The theory helps frame a formally rigorous, normative approach to the tradeoff between control-dependent processing versus automaticity, and relates to a number of other fundamental principles and phenomena concerning cognitive function, and computation more generally.

Rationalizing Constraints on the Capacity for Cognitive Control

“…However, this can be mitigated by the engagement of control, by limiting the engagement of representations shared by multiple tasks to one task at a time [80,81]. From this perspective, constraints on multitasking capability reflect the engagement of control as a rational adaption to resource limitations, rather than an intrinsic limit in control mechanisms themselves [45,54]. Furthermore, these constraints can be considered a cost incurred in exchange for the benefits gained from shared representations: the repurposing of existing representations for new tasks allow these to be acquired rapidly, but at the cost of constraints on multitasking capability and dependence on control [54,82].…”

“…From this perspective, constraints on multitasking capability reflect the engagement of control as a rational adaption to resource limitations, rather than an intrinsic limit in control mechanisms themselves [45,54]. Furthermore, these constraints can be considered a cost incurred in exchange for the benefits gained from shared representations: the repurposing of existing representations for new tasks allow these to be acquired rapidly, but at the cost of constraints on multitasking capability and dependence on control [54,82]. This hypothesis is supported by a mathematical analysis of this relationship in linear networks [65,54], as well as artificial agents that optimize the tradeoff between learning efficacy and multitasking capability [68,83].…”

“…Furthermore, these constraints can be considered a cost incurred in exchange for the benefits gained from shared representations: the repurposing of existing representations for new tasks allow these to be acquired rapidly, but at the cost of constraints on multitasking capability and dependence on control [54,82]. This hypothesis is supported by a mathematical analysis of this relationship in linear networks [65,54], as well as artificial agents that optimize the tradeoff between learning efficacy and multitasking capability [68,83]. The latter work suggests that it can be optimal, under finite time horizons, for neural agents to harvest immediate rewards from tasks that are learned quickly, at the cost of having to execute them in serial.…”

“…Fig. 2D) [45,54]. The tradeoff between shared versus separated representations may also inform constraints in cognitive functions related to cognitive control, such as working memory and visual attention.…”

“…As a practical matter-both in artificial neural networks and the brain-it can be difficult to determine exactly how information is represented in a module (i.e., the specific patterns of activity associated with each representation), and similarly the extent to which two tasks share representations. To address this, the amount of representation sharing can be quantified by correlating the average pattern of activities for pairs of tasks [45,54,59]. This reconceptualization of a "shared resource" has two consequences: (1) tasks may share resources in a continuous fashion, rather than sharing or not sharing a resource; and, given the capacity for learning in neural networks, (2) the extent of sharing may change over time.…”