Feature Selection Methods for Zero-Shot Learning of Neural Activity.

Area of Science:

• Neuroscience
• Machine Learning
• Data Science

Background:

• High dimensionality in neuroimaging data presents challenges for predictive modeling.
• Zero-shot prediction models map neural signals to semantic attributes, enabling classification of unseen stimuli.
• Feature selection significantly impacts model performance but lacks systematic study in this context.

Purpose of the Study:

• To systematically investigate feature selection methods for zero-shot learning in human neuroimaging.
• To compare the tradeoffs of different feature selection techniques on functional Magnetic Resonance Imaging (fMRI) and Electrocorticography (ECoG) data.
• To identify feature selection strategies that optimize accuracy, robustness, and efficiency.

Main Methods:

• Compared correlation-based feature stability with other feature selection techniques.
• Utilized comparable datasets from fMRI and ECoG.
• Evaluated zero-shot prediction accuracy, feature set size, and spatial/spectral patterns.

Main Results:

• Most feature selection methods yielded similar zero-shot prediction accuracies and feature patterns.
• A novel feature/attribute correlation approach achieved comparable accuracies using significantly fewer features.
• This suggests potential for more parsimonious and efficient prediction models.

Conclusions:

• Feature selection is a critical determinant of success in neuroimaging-based zero-shot learning.
• A novel feature/attribute correlation method offers a promising avenue for developing simpler yet effective zero-shot prediction models.
• Future research should explore this approach further for broader applications in brain-computer interfaces and cognitive neuroscience.