Probabilistic working memory representations in human cortex guide behavior

Area of Science:

• Cognitive Neuroscience
• Computational Neuroscience
• Neuroimaging

Background:

• Working memory models propose different memory architectures.
• Information-sparse models treat memories as point estimates.
• Information-rich models encode memories as probability distributions reflecting uncertainty.

Purpose of the Study:

• To differentiate between information-sparse and information-rich working memory models.
• To investigate the representational format of individual memories in the brain.
• To link neural representations of memory to behavioral predictions.

Main Methods:

• Participants performed a motion direction recall task with iterative reports.
• Constructed probability distributions of memory from participant responses.
• Decoded neural probability distributions from functional magnetic resonance imaging (fMRI) data of the occipital and parietal cortex.

Main Results:

• The shape of behavioral memory distributions (e.g., asymmetry) on single trials mirrored neural distributions.
• Neural representations in the occipital and parietal cortex encode more than just the memorized feature.
• The variance (width, asymmetry) of neural distributions predicts memory behavior.

Conclusions:

• Findings support information-rich models of working memory.
• Neural representations of individual memories are rich and encode uncertainty.
• Memory uncertainty, reflected in neural distribution shape, is behaviorally relevant.