(수정중) [Paper Review] NeuTraL AD

Neural Transformation Learning for Anomaly Detection Beyond Images

link to abstract: https://arxiv.org/abs/2103.16440

 

1. Introduction

Deep 비정상 탐지를 위한 neural transformation 학습이며 자기지도 학습 방법이다.

Key Idea: Transformation을 거친 데이터를 semantic space로 임베딩 시킨다. Transform 된 데이터는 원본 데이터와 비슷함과 동시에 각 transformation들은 매우 달라 특이점으로 구분 가능하다.

이 모델의 2개의 구성 요소: 1) 정해진 갯수의 학습 가능한 transformation들, 그리고 2) encoder 모델.

 

2. Related Works

2-1. 

• Deep AD (e.g. deep AE variants, deep one-class classification, deep generative models, outlier exposure):
• self-supervised learning:
• contrastive representation learning:
• learning data augmentation schemes:

NeuTraL AD

• neural transformation learning for anomaly detection
• deep anomaly detection method based on contrastive learning for general data types
• components: a set of learnable transformations and an encoder. both jointly trained on a deterministic contrastive loss (DCL)
• purposes of objective:
• learnable data transformations:
• deterministic contrastive loss (DCL):