Deeply Felt Affect: The Emergence of Valence in Deep Active Inference.

Area of Science:

• Computational Neuroscience
• Affective Science
• Bayesian Inference

Background:

• Emotional valence is crucial for adaptive behavior but lacks formal theoretical and computational models.
• Existing frameworks struggle to explain the origins and functions of affective valence.

Purpose of the Study:

• To develop a principled Bayesian model of emotional valence using deep active inference.
• To elucidate the functional role of valence in optimizing action selection and confidence.
• To introduce 'affective charge' (AC) as a mechanism for updating valence states.

Main Methods:

• Utilized deep active inference, a hierarchical Bayesian inference scheme.
• Formulated valence as the expected precision of an agent's action model (subjective fitness).
• Simulated an affective agent in a T-maze task with context learning and reversal.

Main Results:

• The model demonstrates how internal valence representations optimize action selection confidence.
• Affective charge (AC) was shown to update subjective fitness, assigning sign to prediction-outcome errors.
• Simulations validated the model's ability to infer affective states and reduce uncertainty.

Conclusions:

• Active inference provides a formal and computationally tractable framework for understanding affect.
• The model links affect, action, and implicit metacognition, explaining how biological systems infer affective states.
• This work lays the foundation for future research, including fitting the model to behavioral and neural data.