Reverse-Engineering the Cortical Architecture for Controlled Semantic Cognition

Abstract

We present a ‘reverse engineering’ approach to deconstruct cognition into neurocomputational mechanisms and their underlying cortical architecture, using controlled semantic cognition as a test case. By systematically varying the structure of a computational model and assessing the functional consequences, we identified architectural properties necessary for generating the core functions of the semantic system. Semantic cognition presents a challenging test case as the brain must achieve two seemingly contradictory functions: abstracting context-invariant conceptual representations across time and modalities, whilst producing specific context-sensitive behaviours appropriate for the immediate task. These functions were best achieved in models possessing a single, deep multimodal hub with sparse connections from modality-specific inputs, and control systems acting on peripheral rather than deep network layers. These architectural features correspond well with those suggested by neural data, strongly supporting the efficacy of the reverse engineering approach, and further generating novel hypotheses about the neuroanatomy of controlled semantic cognition.

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

“…A related issue concerning representational learning is the transfer (inductive generalization) of concepts in semantic cognition (e.g. reasoning from multiple instances of birds that all birds lay eggs; Abel et al, 2015;Jackson, Rogers, & Ralph, 2019;Ralph, Jefferies, Patterson, & Rogers, 2017). Here, we have argued that shared representation across tasks facilitates inference and transfer in control-dependent processing.…”

“…Here, we have argued that shared representation across tasks facilitates inference and transfer in control-dependent processing. Similarly, in semantic cognition shared representation across stimulus modalities and contexts can achieve transfer of concepts (Jackson et al, 2019; T. T. Rogers & McClelland, 2004;Rumelhart et al, 1993). In their recent work, Jackson et al (2019) showed that the latter is facilitated in networks that allow information from different modalities to converge in the same "hub" for shared RATIONAL BOUNDEDNESS OF COGNITIVE CONTROL 167 representation.…”

“…Similarly, in semantic cognition shared representation across stimulus modalities and contexts can achieve transfer of concepts (Jackson et al, 2019; T. T. Rogers & McClelland, 2004;Rumelhart et al, 1993). In their recent work, Jackson et al (2019) showed that the latter is facilitated in networks that allow information from different modalities to converge in the same "hub" for shared RATIONAL BOUNDEDNESS OF COGNITIVE CONTROL 167 representation. Interestingly, the acquisition of semantic concepts follows a developmental trajectory which, at the level of representation learning, resembles the trajectory from controlled to automatic processing described above.…”

Rationalizing Constraints on the Capacity for Cognitive Control

“…Yet, little is known about when and why people acquire shared representations across tasks. [66]. Similarly, architectural biases toward the learning of shared representation between tasks can accelerate the sequential acquisition of these tasks [65,68].…”

Stimulus-independent neural coding of event semantics: Evidence from cross-sentence fMRI decoding

“…If these regions are important for regulating semantic activation rather than storing involved (e.g., Rogers et al, 2004;O'Connor et al, 2009;Taylor et al, 2012;Schapiro et al, 2013). Conversely, there has been little formal modelling of semantic control processes (though for recent exceptions, see Hoffman et al, 2018;Jackson et al, 2019). The lack of established theoretical models makes it difficult to make precise predictions about how neural patterns in control regions should vary and how these areas might be distinguished from those that represent knowledge.…”