Self‑supervised evolutionary learning for progression-aware segmentation of neurodynamic time series during...

Area of Science:

• Neuroscience
• Cognitive Science
• Machine Learning

Background:

• Understanding cognition from neurophysiological signals is key for cognitive science and human-robot interaction.
• Discovering latent neurodynamic geometry and temporal progression is challenging due to annotation difficulties and limitations of traditional methods like Hidden Markov Models.

Purpose of the Study:

• Propose a novel representational learning method for segmenting cognitive stages from neurophysiological data.
• Shift from statistical change-point detection to self-supervised learning for discovering temporal organization directly from data.

Main Methods:

• Employed self-supervised learning with four jointly optimized objectives: temporal predictability, boundary contrast, cross-trial alignment, and sparse stage-specific feature weights.
• Utilized population-based evolutionary search to navigate the optimization landscape.
• Validated on EEG data from a road-crossing decision-making task.

Main Results:

• Achieved an order-of-magnitude improvement in boundary contrast of discovered cognitive stages.
• Demonstrated robust identification of cross-trial transferable state geometry and handling of data variability.
• Reconstructed cognitive stages were behaviorally plausible and aligned with neurophysiological underpinnings.

Conclusions:

• The proposed method effectively captures higher-order global temporal organization, moving beyond local statistical consistency.
• This approach offers a robust framework for discovering latent cognitive states from neurophysiological data.
• The findings have implications for advancing cognitive science research and wearable-enabled human-robot interaction.