# Influence-Directed Explanations for Deep Convolutional Networks

**Authors:** Klas Leino, Shayak Sen, Anupam Datta, Matt Fredrikson, Linyi Li  
**arXiv:** [1802.03788](https://arxiv.org/abs/1802.03788)  
**Date:** 2018-02-11  
**Summary page:** https://www.summarizepaper.com/en/paper/1802.03788v2/

## Key points

- Authors: Klas Leino, Shayak Sen, Anupam Datta, Matt Fredrikson, Linyi Li
- Novel approach: Utilizing influence-directed explanations to understand deep neural networks
- Methodology: Using influence measure grounded in axioms to interpret concepts represented by influential neurons
- Validation: Thorough evaluation on convolutional neural networks trained on ImageNet
- Identify influential concepts with generalizability across instances
- Distill the core essence of what the network has learned about a class
- Isolate individual features crucial for decision-making and differentiation between classes
- Findings: Shed light on deep neural network operations and enhance understanding of complex machine learning models

## Summary

In their paper titled "Influence-Directed Explanations for Deep Convolutional Networks," authors Klas Leino, Shayak Sen, Anupam Datta, Matt Fredrikson, and Linyi Li delve into the intricate problem of explaining a wide array of behavioral properties exhibited by deep neural networks. Their novel approach involves utilizing influence-directed explanations to peer into the network's inner workings and pinpoint neurons that hold significant sway over a specific quantity and distribution of interest. By employing an influence measure grounded in axioms, the researchers are able to provide interpretations for the concepts represented by these influential neurons. To validate the effectiveness of their methodology, the team conducts a thorough evaluation on convolutional neural networks trained on ImageNet. Through this evaluation, they showcase several key strengths of influence-directed explanations. Firstly, these explanations successfully identify influential concepts that exhibit generalizability across instances. Secondly, they demonstrate the capability to distill the core "essence" of what the network has learned about a particular class. Lastly, the approach excels in isolating individual features that play a crucial role in the network's decision-making process and its ability to differentiate between closely related classes. The findings presented in this study not only shed light on how deep neural networks operate but also highlight the potential for influence-directed explanations to enhance our understanding of complex machine learning models. With their innovative approach and compelling results, Leino et al. 's research contributes significantly to advancing interpretability in deep convolutional networks and lays a solid foundation for future investigations in this domain.