Exploring Hierarchical Auditory Representation via a Neural Encoding Model.

Area of Science:

• Neuroscience
• Cognitive Science
• Machine Learning

Background:

• Recent functional magnetic resonance imaging (fMRI) studies use deep neural networks (DNNs) to model hierarchical auditory representations in the superior temporal gyrus (STG).
• Supervised DNNs may bias feature extraction towards discriminative attributes, potentially skewing the understanding of neural acoustic processing.
• This limitation highlights the need for alternative feature extraction methods to capture a more comprehensive view of auditory hierarchies.

Purpose of the Study:

• To investigate hierarchical neural auditory representations using an unsupervised deep learning approach.
• To explore auditory processing beyond classification-biased features.
• To identify brain regions involved in hierarchical auditory feature representation.

Main Methods:

• Employed an unsupervised deep convolutional auto-encoder (DCAE) model within an fMRI encoding framework.
• Derived hierarchical feature representations from naturalistic auditory stimuli across different categories.
• Utilized fMRI data to map neural responses to these extracted features.

Main Results:

• Hierarchical auditory feature representation extends beyond the STG.
• Neural correlates include the bilateral insula, ventral visual cortex, and thalamus.
• The unsupervised approach revealed a more distributed network for auditory processing.

Conclusions:

• The human brain's hierarchical auditory processing involves a wider network than previously identified.
• Unsupervised feature extraction provides complementary insights into neural acoustic processing.
• Findings contribute to a more nuanced understanding of auditory hierarchies in the brain.