# Ablation Studies in Artificial Neural Networks

**Authors:** Richard Meyes, Melanie Lu, Constantin Waubert de Puiseau, Tobias Meisen  
**arXiv:** [1901.08644](https://arxiv.org/abs/1901.08644)  
**Date:** 2019-01-24  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/1901.08644v1/

## Key points

- Ablation studies traditionally used in neuroscience to understand complex biological systems
- Application of ablation studies in artificial neural networks (ANNs) explored by authors Richard Meyes, Melanie Lu, Constantin Waubert de Puiseau, and Tobias Meisen
- Two ablation studies conducted on different ANNs using benchmark datasets from computer vision domain
- Specific features related to local and global structure of data selectively represented in distinct parts of the network
- Some representations found to be redundant, providing network with robustness against structural damages
- Importance of specific parts of the network for classification tasks analyzed based on weight structure of individual units
- Exploration of how damaged networks can recover from ablations through recovery training
- Ablation studies offer feasible method for investigating knowledge representations in ANNs and assessing network resilience to structural damages
- Study spans 19 pages with 23 figures, indicating comprehensive exploration of research findings
- Intention expressed by authors to publish work at prestigious AI conference

## Summary

In their paper titled "Ablation Studies in Artificial Neural Networks," authors Richard Meyes, Melanie Lu, Constantin Waubert de Puiseau, and Tobias Meisen explore the application of ablation studies in the field of artificial neural networks (ANNs). Ablation studies have traditionally been used in neuroscience to understand complex biological systems such as the Drosophila central nervous system, vertebrate brain, and human brain. The authors aim to investigate whether similar principles can be applied to ANNs to uncover the inner representations within these networks. The study involves conducting two ablation studies on different ANNs using well-known benchmark datasets from the computer vision domain. Through their experiments, the authors discover that specific features related to the local and global structure of data are selectively represented in distinct parts of the network. Interestingly, they find that some of these representations are redundant, providing the network with a level of robustness against structural damages. Furthermore, the authors analyze the importance of specific parts of the network for classification tasks based solely on the weight structure of individual units. They also explore how damaged networks can recover from ablations through recovery training. Overall, they argue that ablation studies offer a feasible method for investigating knowledge representations in ANNs and are particularly valuable for assessing a network's resilience to structural damages—a crucial aspect for future safety-critical applications. The paper spans 19 pages and includes 23 figures, indicating a comprehensive exploration of their research findings. The authors express their intention to publish this work at one of today's prestigious AI conferences. This study sheds light on how ablation studies can provide insights into understanding inner representations within artificial neural networks and emphasizes the importance of examining network robustness for future applications in critical domains.