Active inference, selective attention, and the cocktail party problem

Holmes, Emma; Parr, Thomas; Griffiths, Timothy D.; Friston, Karl · 2021 · OpenAlex-citations

DOI: 10.1016/j.neubiorev.2021.09.038

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper introduces a generative model based on active inference to explain selective and preparatory attention within the context of the "cocktail party" listening paradigm. The authors address two primary questions: first, how computational processes underlying precision weighting explain distinct types of errors (e.g., reporting words from the competing talker versus random errors) during selective attention; and second, whether behavioral improvements in reaction times and electrophysiological ramping effects (resembling the contingent negative variation, or CNV) during preparatory attention arise from the same underlying mechanisms. The study aims to provide a quantitative, computational account of covert endogenous attention, moving beyond conceptual models to test specific hypotheses against empirical data. The researchers employed a partially observable Markov decision process (MDP) to simulate synthetic agents performing a simplified cocktail party task. In this task, agents receive simultaneous auditory inputs from left and right ears and a visual spatial cue indicating which ear to attend. The generative model defines hidden states (e.g., target words, spatial attention location) and observable outcomes (e.g., spoken words, visual cues, feedback). Attention is formalized as the precision of probabilistic mappings between hidden states and sensory observations. To investigate error mechanisms, the authors performed computational "lesions" by parametrically varying the precision of likelihood mappings for attended words, unattended words, spatial cues, and feedback. To investigate preparatory attention, they simulated agents with temporal changes in sensory precision during the interval between the cue and the target stimulus, comparing model predictions of reaction times and EEG responses to empirical data from previous studies. The results demonstrated that different precision manipulations produced distinct error profiles. Reducing precision for attended words or spatial cues primarily resulted in "absent errors" (random responses), whereas increasing the precision of unattended words led to "mix errors" (reporting a combination of target and competitor words) and "masker errors" (reporting competitor words). This finding suggests that failures in selective attention are driven by the inability to suppress precision for unattended sources rather than a lack of precision for attended sources. Regarding preparatory attention, the simulations revealed a dissociation between behavioral and electrophysiological correlates. Temporal changes in sensory precision were not required to explain faster reaction times with longer cue-target intervals. However, such temporal precision changes were necessary to reproduce the ramping ERP effects observed in human data. The model further suggested that CNV-like responses reflect subjective precision updates rather than action selection. The significance of this work lies in its ability to dissociate the computational processes underlying different aspects of attention. By showing that reaction time improvements and ERP ramping are driven by distinct mechanisms within the active inference framework, the paper challenges the assumption that these phenomena share a common origin. Furthermore, the model provides a mechanistic explanation for informational masking in cocktail party listening, attributing specific error types to the precision weighting of unattended sensory inputs. This approach offers a unified, quantitative framework for understanding how the brain optimizes precision to manage competing sensory streams and prepare for upcoming stimuli.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-25; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• attention selective divided