A Task-Optimized Neural Network Replicates Human Auditory Behavior, Predicts Brain Responses, and Reveals a Cortical...

Area of Science:

• Auditory Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Understanding auditory cortex function is crucial for modeling neural responses to complex natural sounds.
• Existing models often lack the capacity to capture the full range of auditory processing.
• Ecological relevance is key for developing comprehensive models of sensory systems.

Purpose of the Study:

• To develop and evaluate a hierarchical neural network optimized for speech and music recognition.
• To investigate if task optimization can replicate human auditory cortex organization and performance.
• To compare the predictive power of the optimized network against traditional spectrotemporal filter models for fMRI data.

Main Methods:

• Hierarchical neural networks were trained and optimized for speech and music recognition tasks.
• Network architecture and performance were analyzed for similarities to human auditory processing.
• Functional magnetic resonance imaging (fMRI) voxel responses were predicted using the network and compared to established models.

Main Results:

• The best-performing network featured separate speech and music pathways after initial shared processing, mirroring human auditory cortex organization.
• The network achieved human-level performance on both speech and music recognition and exhibited similar error patterns.
• The optimized network significantly outperformed traditional spectrotemporal filter models in predicting fMRI responses across the auditory cortex.
• Network layers quantitatively mapped to the hierarchical representational structure of the auditory cortex.

Conclusions:

• Task optimization is a potent strategy for creating accurate computational models of sensory systems.
• Hierarchical neural networks trained on relevant tasks can effectively model auditory cortex function and organization.
• The findings suggest shared computational constraints between artificial neural networks and the human brain for auditory processing.