Zoom URL: https://mit.zoom.us/j/93867995763

Probabilistic inference has been a promising framework for characterizing many aspects of human cognition, but it also leaves open a question: how does the human mind solve inference problems given that exact inference is often computationally challenging? In this talk, I will present an initial study towards addressing this question. Drawing insights from deep learning and statistics, we take "representation fine-tuning" as a hypothesis of how the human mind achieves approximate inferences, that the mind maintains a generic yet versatile representation and in the face of a novel inference problem tunes that representation for the necessary specialized computations, in contrast to an alternative "sampling with motifs" hypothesis, that flexible inferences reflect a principled algorithmic deployment of a finite repertoire of simple and reusable inference motifs. We ground the behavioral predictions of both hypotheses in a novel linguistic task, where subjects were instructed to use as many or little words as they needed to fill in the blanks in fragmentary linguistic input, such as "____ published won ____ .". We specifically compared three fine-tuned text infilling models to a Gibbs sampling-based approximate inference algorithm, which involves the flexible combination of two simple inference motifs: masked and unidirectional language modeling. With no task-specific parameter tuning, the sampling-based algorithm achieves good performance comparable to fine-tuned counterparts ​in tasks involving complex syntactic constraints, and its completion of more open-ended fragmentary input better matches the fine-grained structural statistics in human responses. These results suggest the sampling-based approximate inference may provide a better algorithmic account of how the human mind deploys linguistic knowledge to support flexible inferences.