A Neural Model of Task Compositionality with Natural Language Instructions

Abstract

AO_SCPLOWBSTRACTC_SCPLOWOne of humans most fundamental cognitive feats is the ability to interpret linguistic instructions in order to perform novel tasks without any explicit experience with the task. Yet, the computations that the brain might use to accomplish such a feat remains poorly understood. Here we use the latest advances in Natural Language Processing to create a neural model of generalization based on linguistic instructions. Models are trained on a set of commonly studied psychophysical tasks, and receive instructions embedded by a pre-trained language model. Our best models can perform a previously unseen task with a performance of 83% correct on average based solely on linguistic instructions (i.e. 0-shot learning). We found that language scaffolds sensorimotor representations such that activity for interrelated tasks share a common geometry with the semantic representations of instructions, allowing language to cue the proper composition of practiced skills in unseen settings. Finally, we show how this model can generate a linguistic description of a novel task it has identified using only motor feedback, which can subsequently guide a partner model to perform the task. Our models offer several experimentally testable predictions outlining how linguistic information must be represented in order to facilitate flexible and general cognition in the human brain.