Neural network modeling of altered facial expression recognition in autism spectrum disorders based on predictive...

Area of Science:

• Computational Neuroscience
• Developmental Psychology
• Autism Spectrum Disorder Research

Background:

• The developmental mechanisms of facial emotion recognition remain unclear.
• Predictive processing theory offers a framework for understanding perception and its alterations in neurodevelopmental conditions.

Purpose of the Study:

• To propose a system-level explanation for facial emotion recognition development using predictive processing.
• To investigate how altered predictive processing contributes to impairments in autism spectrum disorder (ASD).

Main Methods:

• A hierarchical recurrent neural network (RNN) was developed to model facial emotion recognition.
• The RNN learned to predict dynamic facial expressions without explicit emotion labels, simulating developmental learning.
• The model was evaluated on recognizing novel facial expressions and simulating ASD-related neural alterations.

Main Results:

• Emotional clusters spontaneously emerged in the RNN's higher-level neurons during self-organization, without explicit emotional labels.
• The network successfully recognized unseen facial expressions by minimizing prediction errors.
• Simulating altered neural excitability in ASD led to reduced generalization and impaired emotional clustering, linked to excessive precision estimation.

Conclusions:

• Facial emotion recognition can emerge through self-organized predictive processing during development.
• Impaired facial emotion recognition in ASD can be explained by alterations in predictive processing mechanisms.
• This model provides insights into the neurophysiological basis of social-affective deficits in ASD.