llien30 / paper_summary

Write the summaries of papers I've read.
7 stars 0 forks source link

Memorizing Normality to Detect Anomaly: Memory-augmented Deep Autoencoder for Unsupervised Anomaly Detection #99

Open llien30 opened 3 years ago

llien30 commented 3 years ago

INFO

Author

Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, Anton van den Hengel

Affiliation

Conference or Year

ICCV2019

Link

Abstract

提案手法(MemAE)

Screen Shot 2021-01-19 at 16 04 37

Proposed Method

Screen Shot 2021-01-19 at 16 12 01

Memory-module

Hard Shrinkaga for Sparse Addressing

それでもまだ、reconstructionできてしまう異常があるので(wがdenseだと、細かな情報を持ちうるから)、wをsparseにすることで対応する (λ : shrinkage threshold)

Screen Shot 2021-01-19 at 16 21 46 Screen Shot 2021-01-19 at 16 24 24

Objective Function

Screen Shot 2021-01-19 at 16 25 58

再構成誤差

Screen Shot 2021-01-19 at 16 26 05

wのsparse性をなくすロス

Screen Shot 2021-01-19 at 16 26 17

論文内ではα = 0.0002

Loss = (xと\hat{x}のL2 norm) + 0.0002 * (wのentropy loss)

Evaluation

Image data

Screen Shot 2021-01-19 at 16 33 13

Video data

Screen Shot 2021-01-19 at 16 32 55

KDDCUP

Screen Shot 2021-01-19 at 16 32 04

Contribution

Discussion, Future Work

Comment

Date

2021/01/19