Poster Presentation

Poster F in Poster Session F: Thursday, August 6, 1:45 – 3:30 pm, Kimmel Center, Shorin & Rosenthal Rooms

Neural computation for prediction in noisy, unknown environments

Jacopo Epifanio¹, Laavanya Joshi Malik², Ralph G. Andrzejak¹, Tahra Eissa²; ¹Universitat Pompeu Fabra, ²University of Colorado Anschutz

Presenter: Jacopo Epifanio

A central question in neuroscience is how neural systems generate predictions about upcoming sensory input when observations are noisy and the hidden structure of the environment is unknown. We address this question in a predictive-inference task in which observers receive noisy stimuli from one of two sources (states) and must predict which source will generate the next stimulus. The task includes an unknown probability of change between trials (the hazard rate), so optimal prediction requires inferring not only the current source, but also the hazard rate. The optimal behavioral solution is captured by a hierarchical Bayesian observer model, which jointly infers these latent quantities and generates trial-wise predictions about the most likely source of the next stimulus. However, the ideal observer is a behavioral model and does not specify how the relevant computations are implemented in the brain. To address this gap, we trained recurrent neural networks (RNNs) to reproduce normative task variables derived from the Bayesian observer and to serve as candidate mechanistic models bridging behavior to neural activity. Our RNNs perform the prediction task with the same behavior and accuracy as the Bayesian model. They also show modulations in activity within individual units that jointly encode sensory evidence and latent task features, such as the hazard rate. We then consider intracranial electroencephalography (iEEG) recordings in humans during task performance and find that hippocampal activity is modulated in response to task features similarly to RNN activity, although the networks were never trained on iEEG. Thus, because RNNs' internal dynamics can be directly analyzed and systematically perturbed, they provide both: 1) a useful framework for testing how architectural and connectivity features shape inference strategy, and 2) a tool for interpreting the relationship between inference strategy and human brain recordings, suggesting a multiplexed approach to hierarchical latent inference.

Topic Area: Methods, Tools, Theory & Neural Coding