# LINe: Out-of-Distribution Detection by Leveraging Important Neurons

**Authors:** Yong Hyun Ahn, Gyeong-Moon Park, Seong Tae Kim  
**arXiv:** [2303.13995](https://arxiv.org/abs/2303.13995)  
**Date:** 2023-03-24  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.13995v1/

## Key points

- Accurately quantifying uncertainty of input samples is crucial in mission-critical domains like autonomous driving and healthcare
- Models cannot express what they are not aware of, making OOD detection challenging
- Post-hoc OOD detection approaches have gained popularity as they do not require additional re-training processes
- The study proposes a novel method called LINe for post-hoc OOD detection using high-level features in deep models
- LINe focuses on analyzing the difference in model outputs between in-distribution and OOD data
- Shapley value-based pruning is used to select high-contribution neurons for predicting specific classes of input data
- Activation clipping is employed to fix values above a certain threshold into the same value, treating all class-specific features equally
- LINe outperforms state-of-the-art post hoc OOD detection methods according to experiments on popular datasets
- The study has significant implications for improving safety measures in critical domains like autonomous driving and healthcare.

## Summary

In mission-critical domains like autonomous driving and healthcare, accurately quantifying the uncertainty of input samples is crucial to prevent failure predictions on out-of-distribution (OOD) data. The challenge with OOD detection lies in the fact that models cannot express what they are not aware of. To address this issue, post-hoc OOD detection approaches have gained popularity as they do not require additional re-training processes that could potentially degrade model performance and increase training costs. In this study titled "LINe: Out-of-Distribution Detection by Leveraging Important Neurons," authors Yong Hyun Ahn, Gyeong-Moon Park, and Seong Tae Kim propose a novel method for post-hoc OOD detection from the perspective of neurons in the deep layer of a model representing high-level features. Their approach, called Leveraging Important Neurons (LINe), focuses on analyzing the difference in model outputs between in-distribution data and OOD data. To reduce the effects of noisy outputs, the authors employ Shapley value-based pruning. This technique selects only high-contribution neurons for predicting specific classes of input data while masking the rest. Additionally, activation clipping is used to fix all values above a certain threshold into the same value. This allows LINe to treat all class-specific features equally and solely consider the difference in activated feature counts between in-distribution and OOD data. The effectiveness of LINe is validated through comprehensive experiments on popular datasets such as CIFAR-10, CIFAR-100, and ImageNet. The results demonstrate that LINe outperforms state-of-the art post hoc OOD detection methods. Overall, this study introduces an innovative approach for detecting out -of - distribution samples by leveraging important neurons in deep models. By effectively quantifying uncertainty and distinguishing between in - distribution and OOD data , LINe has significant implications for improving safety measures in critical domains like autonomous driving and healthcare .