The cost of misremembering: Inferring the loss function in visual working memory

Area of Science:

• Cognitive Psychology
• Computational Neuroscience
• Decision Theory

Background:

• Visual working memory (VWM) has limited capacity, leading to imperfect memory encoding.
• Existing models often assume simple noise corruption, which may not capture the complexity of memory errors.

Purpose of the Study:

• To propose a novel theoretical framework for VWM based on implicit loss functions.
• To investigate how the costs of different memory errors shape VWM behavior.
• To determine if VWM operates optimally given specific natural task demands.

Main Methods:

• Developed a theoretical framework characterizing VWM by its implicit loss function.
• Employed inverse decision theory to estimate the empirical loss function from behavioral data.
• Compared empirical findings to predictions of an optimally efficient VWM for gaze correction tasks.

Main Results:

• Estimated an empirical loss function for VWM from delayed recall experiments.
• Demonstrated that the proposed loss function framework provides a superior account of VWM data compared to alternative models.
• Showed that VWM behavior aligns with optimal efficiency for specific natural tasks like gaze correction.

Conclusions:

• VWM error patterns are not random but reflect an underlying loss function that minimizes costly errors.
• The proposed loss function framework offers a more accurate and comprehensive model of VWM than traditional noise-based models.
• VWM appears to be an optimally efficient system, adapting its error profile to minimize expected loss in naturalistic contexts.