Explainable Multimodal Regression with Information Decomposition
Analysis
This paper addresses the interpretability problem in multimodal regression, a common challenge in machine learning. By leveraging Partial Information Decomposition (PID) and introducing Gaussianity constraints, the authors provide a novel framework to quantify the contributions of each modality and their interactions. This is significant because it allows for a better understanding of how different data sources contribute to the final prediction, leading to more trustworthy and potentially more efficient models. The use of PID and the analytical solutions for its components are key contributions. The paper's focus on interpretability and the availability of code are also positive aspects.
Key Takeaways
- •Proposes a novel multimodal regression framework based on Partial Information Decomposition (PID).
- •Introduces Gaussianity constraints to enable analytical computation of PID terms.
- •Develops a conditional independence regularizer to isolate unique information within each modality.
- •Demonstrates improved predictive accuracy and interpretability compared to existing methods.
- •Provides a case study on brain age prediction and offers code implementation.
“The framework outperforms state-of-the-art methods in both predictive accuracy and interpretability.”