# Local Interpretable Model-agnostic Explanations of Bayesian Predictive Models via Kullback-Leibler Projections

**Authors:** Tomi Peltola  
**arXiv:** [1810.02678](https://arxiv.org/abs/1810.02678)  
**Date:** 2018-10-05  
**Summary page:** https://www.summarizepaper.com/en/paper/1810.02678v1/

## Key points

- Method introduced: Local Interpretable Model-agnostic Explanations of Bayesian Predictive Models via Kullback-Leibler Projections
- Aim of the method: Provide explanations for predictions by locally projecting predictive distribution onto a simpler and more interpretable explanation model
- Information theoretic foundation
- Balances trade-off between explanation fidelity and complexity
- Application demonstration: Explaining MNIST digit classifications performed by a [model type not specified]
- Contribution to the field: Bridges gap between complex models and interpretable explanations, enhancing trust and usability in AI systems

## Summary

In the paper titled "Local Interpretable Model-agnostic Explanations of Bayesian Predictive Models via Kullback-Leibler Projections," authored by Tomi Peltola, the method  is introduced. This method aims to provide explanations for predictions made by  by locally projecting the information in the predictive distribution onto a simpler and more interpretable explanation model. The approach combines elements from the  method with concepts derived from . One of the key strengths of the proposed KL-LIME method lies in its information theoretic foundation, which effectively balances the trade-off between explanation fidelity and complexity. By leveraging this balance, users can gain insights into how  arrive at their decisions while still maintaining a level of interpretability that is crucial for practical applications. To demonstrate the effectiveness of , the authors showcase its application in explaining MNIST digit classifications performed by a . Through this demonstration, they illustrate how KL-LIME can offer valuable insights into complex predictive models, making them more transparent and understandable to end-users. Overall, the paper contributes to the growing field of  by providing a novel approach that bridges the gap between complex  and interpretable explanations, ultimately enhancing trust and usability in AI systems.