Bayesian Computation through Cortical Latent Dynamics.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Cognitive Science

Background:

• Statistical regularities in the environment inform prior beliefs, crucial for optimizing behavior amidst sensory uncertainty.
• Bayesian theory provides a framework for understanding how prior beliefs influence perception, sensorimotor control, and cognition.
• The neural mechanisms by which recurrent neural interactions mediate Bayesian integration remain largely unknown.

Purpose of the Study:

• To investigate how recurrent neural interactions in the brain implement Bayesian inference.
• To understand the neural basis of how prior statistical information is incorporated into sensorimotor transformations.
• To uncover the computational principles underlying the influence of prior beliefs on behavior.

Main Methods:

• A time-interval reproduction task was employed with monkeys to study neural representations.
• Analysis of neural activity in the frontal cortex was performed during the task.
• Recurrent neural network (RNN) models were analyzed to understand computational strategies.

Main Results:

• Prior environmental statistics were found to warp neural representations in the frontal cortex.
• This warping enabled the incorporation of prior statistics into sensory-to-motor mappings, aligning with Bayesian inference.
• Analysis of RNN models identified a low-dimensional curved manifold as key to this neural warping.

Conclusions:

• Prior beliefs influence behavior by sculpting the latent dynamics of cortical neural activity.
• This neural sculpting provides a mechanism for implementing Bayesian inference in sensorimotor control.
• The findings reveal a general principle for how prior knowledge shapes neural computations.