Brain-inspired model for early vocal learning and correspondence matching using free-energy optimization.

Area of Science:

• Computational Neuroscience
• Developmental Psychology
• Machine Learning

Background:

• Infants face a correspondence mismatch problem when learning to vocalize amidst diverse auditory input.
• The cortico-basal ganglia (CX-BG) system is crucial for motor learning and vocalization in humans.
• Existing models often struggle with real-time adaptation to novel acoustic environments.

Purpose of the Study:

• To propose a developmental model for infant vocal learning inspired by the CX-BG system.
• To address the challenge of acoustic matching with unfamiliar voices of varying tones and pitches.
• To introduce and evaluate the INFERNO neural network architecture for audio primitive retrieval.

Main Methods:

• Development of a novel neural architecture, INFERNO (Iterative Free-Energy Optimization of Recurrent Neural Networks).
• Utilizing free-energy minimization principles for action selection and learning in sound production.
• Implementing a CX-BG system for rapid causal linking of sound and motor primitives (millisecond timescale).

Main Results:

• The INFERNO model successfully retrieves audio primitives during vocal learning tasks.
• Demonstrated generalization capabilities across small and large audio databases.
• Exhibited robustness to noise and effective acoustic matching with previously unheard voices (varying genders and tones).

Conclusions:

• The proposed CX-BG inspired model provides a viable computational framework for understanding infant vocal learning.
• INFERNO architecture shows significant potential for real-time audio processing, adaptation, and generalization.
• The findings highlight the efficacy of free-energy minimization in neural systems for perceptual learning and motor control.