Representational geometry explains puzzling error distributions in behavioral tasks.

Area of Science:

• Cognitive Science
• Neuroscience
• Computational Psychology

Background:

• Understanding errors in behavioral tasks is key to cognitive research.
• Previous work assumed Gaussian error distributions for continuous variables in working memory.
• Deviations from normality were linked to complex noise sources.

Purpose of the Study:

• To reassess the assumption of Gaussian error distributions in behavioral tasks.
• To investigate the relationship between encoding manifold geometry and error distribution shape.
• To apply a new theoretical framework to visual short-term memory (VSTM) data.

Main Methods:

• Developed ideal observer models with Gaussian encoding noise.
• Analyzed the geometrical structure of the encoding manifold.
• Applied the derived theory to experimental data from visual short-term memory tasks.

Main Results:

• Error distributions are generally non-Gaussian, even with Gaussian encoding noise.
• The geometry of the encoding manifold determines error distribution shape, often resulting in flat tails for high-dimensional geometries.
• The proposed theory explains a wide range of VSTM data with only two parameters.

Conclusions:

• Challenges the conventional view of working memory mechanisms and capacity.
• Suggests the Bayesian framework effectively explains working memory, similar to perception.
• Highlights the critical, underappreciated role of representational geometry in human behavioral errors.