Time-series attribution maps with regularized contrastive learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Gradient-based attribution methods are crucial for explaining deep learning models.
• Current methods often lack identifiability guarantees, limiting reliable interpretation.
• Time-series data analysis requires specialized attribution techniques.

Purpose of the Study:

• To develop an interpretable deep learning method for time-series data with identifiability guarantees.
• To propose a novel attribution method, Inverted Neuron Gradient (xCEBRA), for time-series analysis.
• To theoretically and empirically validate the identifiability properties of xCEBRA.

Main Methods:

• Developed a regularized contrastive learning algorithm for time-series data.
• Introduced the Inverted Neuron Gradient (xCEBRA) attribution method.
• Conducted theoretical analysis on Jacobian matrix identification.
• Performed empirical validation on synthetic datasets and compared with existing methods.

Main Results:

• xCEBRA demonstrates theoretical properties for identifying the Jacobian matrix.
• Empirically, xCEBRA robustly approximates ground-truth attribution maps.
• Achieved significant improvements over feature ablation, Shapley values, and other gradient-based methods.
• Established the first identifiable inference method for time-series attribution maps.

Conclusions:

• xCEBRA offers a principled approach to interpretable time-series analysis in deep learning.
• This work advances the understanding of neural dynamics and decision processes within neural networks.
• Opens new research avenues for identifiable attribution in complex sequential data.