# Understanding Black-box Predictions via Influence Functions

**Authors:** Pang Wei Koh, Percy Liang  
**arXiv:** [1703.04730](https://arxiv.org/abs/1703.04730)  
**Date:** 2017-03-14  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/1703.04730v1/

## Key points

- Understanding Black-box Predictions via Influence Functions
- Pang Wei Koh and Percy Liang address the challenge of explaining predictions made by black-box machine learning models.
- They introduce influence functions, a technique rooted in robust statistics, to trace a model's decision-making process back to its training data and understand how it arrives at a specific prediction.
- Extension to Complex High-Dimensional Models
- The authors extend influence functions to complex high-dimensional black-box models operating in non-convex and non-differentiable environments by leveraging insights from second-order optimization.
- Efficient Implementation
- An efficient implementation of influence functions is presented that relies solely on gradients and Hessian-vector products.
- Versatility Demonstrated Through Experiments
- Experiments on linear models and convolutional neural networks showcase the versatility of influence functions for gaining insights into model behavior, debugging errors within datasets, and uncovering vulnerabilities that could be exploited through adversarial attacks during training.
- Significance and Practical Applications
- This research emphasizes the importance of understanding how black-box models make predictions and offers a practical approach using influence functions to enhance transparency and interpretability in machine learning systems.
- Tools for Improvement
- The study provides valuable tools for improving model performance and security against potential threats.

## Summary

In their paper "Understanding Black-box Predictions via Influence Functions," Pang Wei Koh and Percy Liang address the challenge of explaining predictions made by black-box machine learning models. They introduce influence functions, a technique rooted in robust statistics, to trace a model's decision-making process back to its training data and understand how it arrives at a specific prediction. By leveraging insights from second-order optimization, they extend influence functions to complex high-dimensional black-box models operating in non-convex and non-differentiable environments. The authors present an efficient implementation of influence functions that relies solely on gradients and Hessian-vector products. Through experiments on linear models and convolutional neural networks, they demonstrate the versatility of influence functions for gaining insights into model behavior, debugging errors within datasets, and uncovering vulnerabilities that could be exploited through adversarial attacks during training. Overall, this research emphasizes the importance of understanding how black-box models make predictions and offers a practical approach using influence functions to enhance transparency and interpretability in machine learning systems. It also provides valuable tools for improving model performance and security against potential threats. are difficult to interpret due to their complex decision-making processes. To address this issue,  are introduced as a technique rooted in robust statistics that can unravel how these models arrive at specific predictions by tracing them back through the learning process to the training data. This method is extended to modern  scenarios by leveraging insights from second-order optimization techniques. The authors provide an efficient implementation of , which only requires access to gradients and Hessian-vector products. Through experiments on linear models and convolutional neural networks, they showcase the versatility of  for various purposes such as gaining insights into model behavior, debugging errors within datasets, and uncovering vulnerabilities that could be exploited through adversarial attacks during training. This research highlights the significance of understanding how  make predictions and offers a practical approach using  to enhance transparency and interpretability in machine learning systems. It also provides valuable tools for improving model performance and security against potential threats.