Deep learning as a tool for neural data analysis: Speech classification and cross-frequency coupling in human...

Area of Science:

• Neuroscience
• Machine Learning
• Computational Neuroscience

Background:

• Understanding neural representations of complex behaviors like speech is a key neuroscience challenge.
• Linear models are limited in capturing nonlinear, hierarchical brain activity patterns.
• Higher-order brain areas require advanced analytical tools beyond linear transformations.

Purpose of the Study:

• To apply deep neural networks to predict speech from human sensorimotor cortex activity.
• To demonstrate the utility of deep networks as a neuroscience data analysis tool.
• To explore neural representations of speech production and identify relevant frequency bands.

Main Methods:

• Utilized deep networks to decode produced speech syllables from high gamma cortical surface potentials.
• Compared deep network performance against baseline linear models.
• Analyzed deep network confusions to reveal latent structures and identified novel neural frequency couplings.

Main Results:

• Deep networks achieved higher decoding prediction accuracy than linear models.
• Network confusions uncovered hierarchical structures related to speech motor control.
• Identified a novel high-gamma-to-beta coupling during speech production.
• Found high-gamma band contains most speech-relevant information, with minimal contribution from lower frequencies.

Conclusions:

• Deep networks are powerful tools for analyzing complex neural data and uncovering nonlinear representations.
• This study advances the understanding of speech production mechanisms through advanced computational analysis.
• Deep learning provides superior insights into neural information processing compared to traditional linear methods.