Cosyne 2024, Friday, March 1st, Poster 2-019

Authors: Yeowon Kim and Yul HR Kang

• This presentation won Cosyne Presenters Travel Grant!

Navigation requires localizing oneself in an environment. However, one’s location (“latent state”) is not directly observable, and must be inferred from noisy and partial information such as egocentric visual input and self-motion signals. Recent findings suggest that the uncertainty about the latent state from such noisy information must be considered for optimal localization, and that it is indeed considered by people during navigation and represented by place/grid fields in the brain. However, it is unclear how such probabilistic representation (i.e., representation of the uncertainty about the latent state) is acquired. Here we show that such representation naturally arises when a neural network is trained to predict the upcoming sensory input. We develop a variant of an autoencoder that receives noisy egocentric visual stimuli/self-motion signals of an agent navigating an environment, and recurrently updates its hidden state to predict the upcoming visual stimulus. Then we show its hidden state represents a handcrafted ideal observer’s belief about its location given noisy sensory inputs. The representation matched not only the optimal estimate of the location but also the estimate’s uncertainty. Also, the decoded uncertainty about the distance from the nearest wall correlated with the distance, matching that of the ideal observer and paralleling the finding about how human homing behavior depends more on closer landmarks, and how a place field’s size correlates with the distance from the nearest wall. A control network trained to reproduce the visual stimulus of the current time step failed to represent the ideal observer’s belief as reliably. Thus, our results suggest that learning to predict the upcoming noisy sensory input may be a potential mechanism for the learning of probabilistic representation in the brain, even in a natural task like spatial navigation where the relationship between the sensory input and the latent state is complex.