Learning representations of environmental priors in visual working memory

Experience helps us learn the structure of the environment. Such learning can be described by statistical inference models in which environmental priors adapt to new information across time and influence future estimates. For instance, we may learn that certain colors are overrepresented in the world (e.g., I will see more greens and browns if I am in a forest), leading to cognitive biases in favor of more common environmental features. Humans display systematic biases when retaining estimates of an object's features in working memory, which may reflect learned priors of object feature values. However, the neural mechanisms that support inferring the environmental prior and producing biases in working memory have not been identified.
We build on observations from human response data that show systematic biases in a delayed-estimation task of color reports to determine if humans modulate their biases based on their experience. Considering subject responses when task stimuli (colors) were drawn from heterogeneous distributions that did not necessarily correspond with reported population biases, we confirm that most subjects’ response distributions are better described by models that learn feature distributions than those that do not. We also found that a neural circuit model that infers the environmental prior via long-term potentiation and homeostatic plasticity could replicate these behavioral findings and represent experienced stimulus history. Changes to synaptic connectivity shape the persistent and collective neural activity that encodes the stimulus estimate in working memory, producing neural activity attractors are aligned to common stimulus values and mechanistically implementing probabilistic priors. This work suggests that systematic limitations in working memory reflect efficient representations of inferred environmental structure, providing new insights into how humans integrate environmental knowledge into their cognitive strategies.