Machine learning for decoding listeners' attention from electroencephalography evoked by continuous speech.

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Auditory attention in multispeaker environments is crucial for communication.
• Previous methods used linear models to decode attended speech from electroencephalography (EEG).

Purpose of the Study:

• To develop and evaluate a non-linear neural network (NN) model for improved auditory attention decoding from EEG.
• To identify salient neural cues associated with attending to a specific speaker in a complex auditory scene.

Main Methods:

• Replaced linear models with a non-linear NN architecture for mapping EEG to speech envelope.
• Optimized NN parameters including EEG frequency range (1-32 Hz) and temporal segmentation.
• Applied a relevance algorithm to identify important electrode signals for attention decoding.

Main Results:

• The non-linear NN achieved a performance seven times higher than linear models.
• Identified relevant EEG activations occurring 170 ms post-stimulus at plausible physiological locations.
• These activations were specific to the attended speaker, not the unattended one.

Conclusions:

• Non-linear NNs offer a significant performance improvement over linear models for EEG-based auditory attention decoding.
• The findings suggest that non-linear NNs can provide valuable insights into the physiological mechanisms of auditory attention.
• This approach holds promise for understanding and potentially assisting individuals with hearing impairments in noisy environments.