# Feature Encoding with AutoEncoders for Weakly-supervised Anomaly Detection

**Authors:** Yingjie Zhou, Xucheng Song, Yanru Zhang, Fanxing Liu, Ce Zhu, Lingqiao Liu  
**arXiv:** [2105.10500](https://arxiv.org/abs/2105.10500)  
**Date:** 2021-05-22  
**Summary page:** https://www.summarizepaper.com/en/paper/2105.10500v1/

## Key points

- Paper title: "Feature Encoding with AutoEncoders for Weakly-supervised Anomaly Detection"
- Addresses the problem of weakly-supervised anomaly detection
- Previous approaches used deep neural networks to map normal and abnormal samples to different regions in the feature space or fit different distributions
- Directly training these networks with discriminative loss may not be sufficient due to limited annotated anomaly samples
- Authors propose a novel strategy using an autoencoder to encode the input data
- Three factors utilized: hidden representation, reconstruction residual vector, and reconstruction error
- New representation encodes a test sample by considering its projection on the training data manifold, its direction to its projection, and its distance to its projection
- Introduce a novel network architecture that incorporates these three factors seamlessly
- Extensive experiments demonstrate superior performance compared to competitive methods
- Innovative approach for weakly-supervised anomaly detection leveraging autoencoders and incorporating multiple factors in the encoding process

## Summary

The paper titled "Feature Encoding with AutoEncoders for Weakly-supervised Anomaly Detection" addresses the problem of weakly-supervised anomaly detection. This involves learning an anomaly detector using a limited amount of labeled data and abundant unlabeled data. Previous approaches have utilized deep neural networks to map normal and abnormal samples to different regions in the feature space or fit different distributions. However, directly training these networks with discriminative loss may not be sufficient due to the limited number of annotated anomaly samples. To overcome this issue, the authors propose a novel strategy that transforms the input data into a more meaningful representation for anomaly detection. They leverage an autoencoder to encode the input data and utilize three factors: hidden representation, reconstruction residual vector, and reconstruction error. This new representation encodes a test sample by considering its projection on the training data manifold, its direction to its projection, and its distance to its projection. In addition to this encoding strategy, the authors also introduce a novel network architecture that seamlessly incorporates these three factors. Through extensive experiments, they demonstrate the benefits of their proposed strategy by achieving superior performance compared to competitive methods. Overall, this paper presents an innovative approach for weakly-supervised anomaly detection by leveraging autoencoders and incorporating multiple factors in the encoding process. The experimental results highlight the effectiveness of their strategy in detecting anomalies accurately.