A hierarchical sparse coding model predicts acoustic feature encoding in both auditory midbrain and cortex.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• The human auditory pathway processes sound through multiple stages, with neurons exhibiting diverse functions and response properties.
• A key question is whether a common encoding mechanism underlies these different stages.

Purpose of the Study:

• To investigate if sparse coding and max pooling are universal encoding principles across the human auditory pathway.
• To compare the response properties of a deep learning model with physiological recordings from the auditory system.

Main Methods:

• Trained an unsupervised deep learning model with alternating sparse coding and max pooling layers on human speech spectrograms (cochleograms).
• Evaluated the spectro-temporal receptive fields (STRFs) and feature tuning of model units.

Main Results:

• Lower model layers showed STRFs similar to inferior colliculus neurons (e.g., onset/termination, spectral motion).
• Upper layers exhibited tuning to phonetic features (plosivity, nasality), resembling auditory cortex responses.
• Higher sparseness correlated with stronger phonetic feature encoding.
• Top layer activities correlated with phoneme formant dynamics (F1, F2).

Conclusions:

• Sparse coding and max pooling appear to be fundamental, potentially universal, encoding mechanisms in the human auditory pathway.
• The deep learning model successfully replicated key response properties observed at different auditory processing levels.