Learning task-agnostic and interpretable subsequence-based representation of time series and its applications in fMRI...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time Series Analysis

Background:

• Sequential learning models excel at representation learning but are often task-specific, hindering generalization.
• Complex models produce abstract representations difficult for humans to comprehend.
• Existing methods like symbolic and recurrent learning offer limited interpretability.

Purpose of the Study:

• To develop a unified local predictive model for learning task-agnostic and interpretable time series representations.
• To enable versatile applications in temporal prediction, smoothing, and classification.
• To enhance human comprehension of spectral information within time series data.

Main Methods:

• Utilized a multi-task learning paradigm to create a unified local predictive model.
• Focused on learning subsequence-based, interpretable representations.
• Employed a proof-of-concept evaluation study comparing against conventional methods.

Main Results:

• Demonstrated empirical superiority of the proposed task-agnostic, interpretable representation over task-specific methods.
• Showcased improved performance in temporal prediction, smoothing, and classification tasks.
• Validated the model's ability to reveal ground-truth periodicity and spectral characteristics.

Conclusions:

• The proposed model generates superior, interpretable, and task-agnostic time series representations.
• These representations offer versatile applications and improve human understanding of spectral information.
• The model shows promise in functional magnetic resonance imaging (fMRI) analysis for characterizing brain activity.