Conference on Cognitive Computational Neuroscience

The distributions of stimuli in everyday life exhibit distinct biases that we may learn to predict what we observe next. Correspondingly, humans performing delayed estimation tasks are biased in favor of more commonly represented stimuli (e.g., common colors in visual tasks). These systematic biases may emerge from attractors in neural ensemble activity representing memoranda during delay periods. However, it is unclear how these heterogeneous representations are learned and how they may be represented in the brain. Here, we present a recurrently coupled neural network that can learn heterogeneous input distributions via a combination of long term potentiation and homeostatic plasticity, generating an underlying potential landscape that governs the preferred positions of a bump attractor on a ring. The resulting spatially-heterogeneous synaptic network probabilistically represents the experienced stimulus history of the environment. We validate this theory using recently published response data from a delayed estimation task in which the distribution of stimuli was heterogeneous. Our results suggest that the heterogeneous models explain the data more often than models for which connectivity is uniform throughout the duration of the experiment.