Bayesian Filtering with Multiple Internal Models: Toward a Theory of Social Intelligence

Area of Science:

• Neuroethology
• Computational Neuroscience
• Animal Behavior

Background:

• Social intelligence requires recognizing communication partners.
• Animals may use internal generative models for conspecifics.
• Learning these models involves Bayesian belief updating, posing a challenge for selecting the correct model to update.

Purpose of the Study:

• To propose a theoretical and neurobiologically plausible solution for inferring and learning sensory input generation and reproduction under multiple generative models.
• To address the problem of updating the correct internal model when encountering sensory input from a conspecific.

Main Methods:

• Utilizing active inference and post hoc (online) Bayesian model selection.
• Fitting sensory inputs under each generative model.
• Updating model parameters based on the probability of each model generating the input (model evidence).
• Implementing physiologically plausible models of birdsong production.

Main Results:

• Demonstrated the scheme using real zebra finch songs, each generated by different birds.
• Showed successful learning across generative models with varying parameters using generalized Bayesian filtering and model selection.
• Validated the approach in a biologically relevant context of birdsong communication.

Conclusions:

• The proposed scheme enables inference and learning of sensory processes under multiple generative models.
• Generalized Bayesian filtering combined with model selection facilitates learning in social contexts.
• Employing multiple internal models is advantageous for social inference with diverse sensory information sources.