[๐Ÿ“–๋…ผ๋ฌธ ๋ฆฌ๋ทฐ] Deep Learning for Anomaly Detection: A Review (2020)

Becky's Study Labยท2024๋…„ 5์›” 23์ผ
0

PaperReview

๋ชฉ๋ก ๋ณด๊ธฐ
19/22

์‚ฐ์—…๊ณตํ•™์› ๋Œ€ํ•™์›์˜ ๊ฝƒ์€ ์ด์ƒ ํƒ์ง€ ์•„๋‹๊นŒ ์‹ถ๋‹ค. ๊ธฐ์กด์— ๊ณต๋ถ€ํ•˜๋˜ ๊ฒƒ๋“ค์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ด์ƒํƒ์ง€๋ผ(Anomaly Detection)์ด๋ผ๋Š” ์ƒˆ๋กœ์šด ๋ถ„์•ผ๋ฅผ ๊ณต๋ถ€ํ•˜๊ณ  ์žˆ๋‹ค. ์ „๋ฐ˜์ ์ธ ๊ธฐ๋ฐ˜์„ ๋‹ค์ง€๊ณ ์ž ์ด ์œ ํŠœ๋ธŒ ์˜์ƒ์™€ ์˜์ƒ์—์„œ ์†Œ๊ฐœํ•œ ๋…ผ๋ฌธ์„ ๊ฐ€์ ธ์™€์„œ ํ•˜๋‚˜ํ•˜๋‚˜ ๋ณด๋ ค๊ณ  ํ•œ๋‹ค.

0. Abstract

  • outlier detection ๋˜๋Š” novelty detection๋ผ๊ณ ๋„ ๋ถˆ๋ฆฌ๋Š” Anomaly detection๋Š” ์ˆ˜์‹ญ ๋…„ ๋™์•ˆ ๋‹ค์–‘ํ•œ ์—ฐ๊ตฌ ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ ์ง€์†์ ์ด๊ณ  ํ™œ๋ฐœํ•œ ์—ฐ๊ตฌ ๋ถ„์•ผ์˜€์Œ
  • ์—ฌ์ „ํžˆ ๊ณ ๊ธ‰ ์ ‘๊ทผ ๋ฐฉ์‹์ด ํ•„์š”ํ•œ ๋ช‡ ๊ฐ€์ง€ ๋…ํŠนํ•œ ๋ฌธ์ œ ๋ณต์žก์„ฑ๊ณผ ๊ณผ์ œ๊ฐ€ ์žˆ์Œ
  • ์ตœ๊ทผ ๋ช‡ ๋…„ ๋™์•ˆ ๋”ฅ๋Ÿฌ๋‹์„ ํ†ตํ•œ ์ด์ƒ ํƒ์ง€๊ฐ€ ์ค‘์š”ํ•œ ๋ฐฉํ–ฅ์œผ๋กœ ๋– ์˜ค๋ฆ„.
  • 3 ๊ฐ€์ง€ ์ƒ์œ„ ์ˆ˜์ค€ ๋ฒ”์ฃผ์™€ 11๊ฐ€์ง€ ์„ธ๋ถ„ํ™”๋œ ๋ฐฉ๋ฒ• ๋ฒ”์ฃผ์˜ ๋ฐœ์ „์„ ๋‹ค๋ฃจ๋Š” ํฌ๊ด„์ ์ธ ๋ถ„๋ฅ˜๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ deep anomaly detection ์—ฐ๊ตฌ๋ฅผ ์กฐ์‚ฌํ•จ
  • key intuitions, objective functions(๋ชฉ์  ํ•จ์ˆ˜), underlying assumptions(๊ธฐ๋ณธ ๊ฐ€์ •), advantages and disadvantages(์žฅ๋‹จ์ )์„ ๊ฒ€ํ† ํ•˜๊ณ  ์•ž์„œ ์–ธ๊ธ‰ํ•œ ๊ณผ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด ๋…ผ์˜ํ•จ

1. Introduction

์ด review ๋…ผ๋ฌธ์€ ์•„๋ž˜ 5๊ฐ€์ง€๋ฅผ ์ง‘์ค‘์ ์œผ๋กœ ๋‹ค๋ฃฌ๋‹ค.

  1. Problem nature and challenges
    : ์šฐ๋ฆฌ๋Š” ์ด์ƒ ํƒ์ง€์˜ ๊ธฐ๋ณธ์ด ๋˜๋Š” ๋ช‡ ๊ฐ€์ง€ ๊ณ ์œ ํ•œ ๋ฌธ์ œ ๋ณต์žก์„ฑ๊ณผ ๊ทธ๋กœ ์ธํ•ด ๋ฐœ์ƒํ•˜๋Š” ๋Œ€๋ถ€๋ถ„์˜ ๋ฏธํ•ด๊ฒฐ ๊ณผ์ œ์— ๋Œ€ํ•ด ๋…ผ์˜ํ•จ

  2. Categorization and formulation (๋ถ„๋ฅ˜ ๋ฐ ๊ณต์‹ํ™”)
    : ์šฐ๋ฆฌ๋Š” current deep anomaly detection methods์„ 3๊ฐ€์ง€ principled frameworks(์›์น™)์— ์˜ํ•ด์„œ ๊ณต์‹ํ™”ํ–ˆ๋‹ค.
    2-1. deep learning for generic feature extraction (์ผ๋ฐ˜์ ์ธ ํŠน์ง• ์ถ”์ถœ์„ ์œ„ํ•œ ๋”ฅ ๋Ÿฌ๋‹)
    2-2. learning representations of normality (์ •์ƒ์„ฑ ํ‘œํ˜„์˜ ํ•™์Šต)
    2-3. end-to-end anomaly score learning (์—”๋“œํˆฌ์—”๋“œ ์ด์ƒ ์ ์ˆ˜ ํ•™์Šต)
    ๋˜ํ•œ 11๊ฐ€์ง€ ๋ชจ๋ธ๋ง ๊ด€์ ์— ๋”ฐ๋ผ ๋ฐฉ๋ฒ•์„ ๋ถ„๋ฅ˜ํ•˜๊ธฐ ์œ„ํ•ด ๊ณ„์ธต์  ๋ถ„๋ฅ˜๋ฅผ ์ œ์‹œํ•จ.

๐Ÿค” end-to-end๋ž€?

๋”ฅ๋Ÿฌ๋‹์—์„œ end-to-end์˜ ์˜๋ฏธ๋Š” ์ž…๋ ฅ์—์„œ ์ถœ๋ ฅ๊นŒ์ง€ ํŒŒ์ดํ”„๋ผ์ธ ๋„คํŠธ์›Œํฌ ์—†์ด ์‹ ๊ฒฝ๋ง์œผ๋กœ ํ•œ ๋ฒˆ์— ์ฒ˜๋ฆฌํ•˜๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, ๋ณต์žกํ•œ ํŒŒ์ดํ”„๋ผ์ธ ์—†์ด ํ•˜๋‚˜์˜ ์‹ ๊ฒฝ๋ง์œผ๋กœ ์ž…๋ ฅ๋ฐ›์€ ๊ฒƒ์„ ์ถœ๋ ฅํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.

  • end-to-end ์žฅ์ 
    1) ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ๋กœ ์ž…๋ ฅ ๋ฐ›์€ ๊ฐ’์— ๋Œ€ํ•œ ์ถœ๋ ฅ ๊ฐ’์„ ์ฐพ์•„๋‚ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
    2) ์ง์ ‘ ํŒŒ์ดํ”„๋ผ์ธ ์„ค๊ณ„ํ•  ํ•„์š”๊ฐ€ ์—†์–ด์ง‘๋‹ˆ๋‹ค.
  • end-to-end ๋‹จ์ 
    1) ๋ผ๋ฒจ๋ง ๋œ ๋ฐ์ดํ„ฐ๊ฐ€ ๋งŽ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
    2) ๋ฉ”๋ชจ๋ฆฌ๊ฐ€ ๋ถ€์กฑํ•  ๊ฒฝ์šฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค.
    3) ๋ณต์žกํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ์—๋Š” ํšจ์œจ์ ์ด์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
  1. Comprehensive literature review
    : ์šฐ๋ฆฌ๋Š” ๊ธฐ๊ณ„ ํ•™์Šต, ๋ฐ์ดํ„ฐ ๋งˆ์ด๋‹, ์ปดํ“จํ„ฐ ๋น„์ „ ๋ฐ ์ธ๊ณต ์ง€๋Šฅ์„ ํฌํ•จํ•˜์—ฌ ์—ฌ๋Ÿฌ ๊ด€๋ จ ์ปค๋ฎค๋‹ˆํ‹ฐ์˜ ์ฃผ์š” ์ปจํผ๋Ÿฐ์Šค ๋ฐ ์ €๋„์—์„œ ๋‹ค์ˆ˜์˜ ๊ด€๋ จ ์—ฐ๊ตฌ๋ฅผ ๊ฒ€ํ† ํ•˜์—ฌ ์—ฐ๊ตฌ ์ง„ํ–‰ ์ƒํ™ฉ์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ๋ฌธํ—Œ ๊ฒ€ํ† ๋ฅผ ์ œ์‹œํ•จ. ์‹ฌ์ธต์ ์ธ ์†Œ๊ฐœ๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด ์šฐ๋ฆฌ๋Š” ๋ฐฉ๋ฒ•์˜ ๋ชจ๋“  ๋ฒ”์ฃผ์—์„œ ์•ž์„œ ์–ธ๊ธ‰ํ•œ ๊ณผ์ œ ์ค‘ ์ผ๋ถ€๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๊ธฐ๋ณธ ๊ฐ€์ •, ๋ชฉ์  ํ•จ์ˆ˜, ์ฃผ์š” ์ง๊ด€ ๋ฐ ๊ธฐ๋Šฅ์„ ์„ค๋ช…ํ•จ.

  2. Future opportunities
    : ์šฐ๋ฆฌ๋Š” ๊ฐ€๋Šฅํ•œ ๋ฏธ๋ž˜์˜ ๊ธฐํšŒ์™€ ๊ทธ ๊ธฐํšŒ๊ฐ€ ๊ด€๋ จ ๊ณผ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์— ๋Œ€ํ•ด ๋” ์ž์„ธํžˆ ๋…ผ์˜ํ•จ.

  3. Source codes and datasets
    : ์šฐ๋ฆฌ๋Š” ๊ฒฝํ—˜์  ๋น„๊ต ๋ฒค์น˜๋งˆํฌ๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด ๊ฑฐ์˜ ๋ชจ๋“  ๋ฒ”์ฃผ์˜ ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด ๊ณต๊ฐœ์ ์œผ๋กœ ์•ก์„ธ์Šคํ•  ์ˆ˜ ์žˆ๋Š” ์†Œ์Šค ์ฝ”๋“œ ๋ชจ์Œ๊ณผ ์‹ค์ œ ์˜ˆ์™ธ ์‚ฌํ•ญ์ด ํฌํ•จ๋œ ์ˆ˜๋งŽ์€ ์‹ค์ œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์š”์ฒญํ•จ.


2. Anomaly Detection: Problem Complexities and Challenges

2.1. Major Problem Complexities

  • Unknownness : Anomalies are associated with many unknowns.(์ด์ƒ ํ˜„์ƒ์€ ๊ธฐ์กด์— ์•Œ๋ ค์ ธ ์žˆ์ง€ ์•Š์€ ๊ฐ‘์ž‘์Šค๋Ÿฌ์šด ํ–‰๋™, ๋ฐ์ดํ„ฐ ๊ตฌ์กฐ, ๋ถ„ํฌ ๋“ฑ ๋งŽ์€ ์•Œ๋ ค์ง€์ง€ ์•Š์€ ์š”์†Œ์™€ ์—ฐ๊ด€๋˜์–ด ์žˆ๋‹ค.)
  • Heterogeneous anomaly classes : Anomalies(๋ณ€์น™)์€ irregular(๋ถˆ๊ทœ์น™์ )์ด๋ฏ€๋กœ ํ•œ ๋ณ€์น™ ํด๋ž˜์Šค๋Š” ๋‹ค๋ฅธ ๋ณ€์น™ ํด๋ž˜์Šค์™€ ์™„์ „ํžˆ ๋‹ค๋ฅธ ๋น„์ •์ƒ์ ์ธ ํŠน์„ฑ์„ ๋‚˜ํƒ€๋‚ผ ์ˆ˜ ์žˆ๋‹ค.
  • Rarity and class imbalance : ์ด์ƒ ํ˜„์ƒ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ๋ฐ์ดํ„ฐ์˜ ์••๋„์ ์ธ ๋น„์œจ์„ ์ฐจ์ง€ํ•˜๋Š” ์ผ๋ฐ˜์ ์ธ ์ธ์Šคํ„ด์Šค์™€ ๋‹ฌ๋ฆฌ ๋“œ๋ฌธ ๋ฐ์ดํ„ฐ ์ธ์Šคํ„ด์Šค๋‹ค.
  • Diverse types of anomaly : ์™„์ „ํžˆ ๋‹ค๋ฅธ 3 ๊ฐ€์ง€ ์œ ํ˜•์˜ anomaly ํ˜„์ƒ์ด ์กฐ์‚ฌ๋˜์—ˆ๋‹ค.
    (a) Point anomalies
    : ๋ถ€๋ถ„์˜ ๋‹ค๋ฅธ ๊ฐœ๋ณ„ ์ธ์Šคํ„ด์Šค์— ๋น„ํ•ด ๋ณ€์น™์ ์ธ ๊ฐœ๋ณ„ ์ธ์Šคํ„ด์Šค
    (b) Conditional anomalies(Contextual anomalies)
    : ๋งฅ๋ฝ์„ ๋ณด์•˜์„ ๋•Œ, ํŠน์ • ์ƒํ™ฉ์—์„œ ๋น„์ •์ƒ ํ˜„์ƒ์„ ๋ณด์ด๋Š” ๊ฒฝ์šฐ
    (c) Group anomalies(Collective anomalies)
    : ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ ์ธ์Šคํ„ด์Šค์— ๋น„ํ•ด ์ „์ฒด์ ์œผ๋กœ ๋น„์ •์ƒ์ ์ธ ๋ฐ์ดํ„ฐ ์ธ์Šคํ„ด์Šค์˜ ํ•˜์œ„ ์ง‘ํ•ฉ, ์ง‘๋‹จ์˜ ๊ฐœ๋ณ„ point๋Š” ์ด์ƒ๊ฐ’์ด ์•„๋‹ ์ˆ˜๋„ ์žˆ์Œ.

โ—outlier ์ข…๋ฅ˜

outlier ์ข…๋ฅ˜

2.2. Main Challenges Tackled by Deep Anomaly Detection

Deep Anomaly Detection๊ฐ€ ๋ช‡ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•  ์ˆ˜ ์žˆ๋Š” ๋‹ค์Œ ๋ฌธ์ œ๋“ค์€ ๋Œ€๋ถ€๋ถ„ ํ•ด๊ฒฐ๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.

- CH1: Low anomaly detection recall rate

์ˆ˜๋…„์— ๊ฑธ์ณ ์ˆ˜๋งŽ์€ ์ด์ƒ ํƒ์ง€ ๋ฐฉ๋ฒ•์ด ๋„์ž…๋˜์—ˆ์ง€๋งŒ ํ˜„์žฌ์˜ ์ตœ์ฒจ๋‹จ ๋ฐฉ๋ฒ•(๋น„์ง€๋„ ๋ฐฉ๋ฒ•)๋˜ํ•œ ์‹ค์ œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ๋Š” ์—ฌ์ „ํžˆ ๋†’์€ ์˜คํƒ๋ฅ ์ด ๋ฐœ์ƒํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ๋‹ค. ์˜คํƒ์„ ์ค„์ด๊ณ  ๊ฒ€์ƒ‰ ์žฌํ˜„์œจ์„ ๋†’์ด๋Š” ๋ฐฉ๋ฒ•์€ ๊ฐ€์žฅ ์ค‘์š”ํ•˜๋ฉด์„œ๋„ ์–ด๋ ค์šด ๊ณผ์ œ ์ค‘ ํ•˜๋‚˜์ด๋‹ค. ํŠนํžˆ ์ด์ƒ ์ง•ํ›„๋ฅผ ๋ฐœ๊ฒฌํ•˜์ง€ ๋ชปํ•˜๋Š” ๋ฐ ๋“œ๋Š” ๋ง‰๋Œ€ํ•œ ๋น„์šฉ์ด ๋ฐœ์ƒํ•œ๋‹ค.

- CH2: Anomaly detection in high-dimensional and/or not-independent data.

  • Anomalies(๋ณ€์น™ ํ˜„์ƒ)์€ ์ €์ฐจ์› ๊ณต๊ฐ„์—์„œ๋Š” ๋ช…๋ฐฑํžˆ ๋น„์ •์ƒ์ ์ธ ํŠน์„ฑ์„ ๋‚˜ํƒ€๋‚ด์ง€๋งŒ, ๊ณ ์ฐจ์› ๊ณต๊ฐ„์—์„œ๋Š” ์ˆจ๊ฒจ์ ธ ๋ˆˆ์— ๋„์ง€ ์•Š๊ฒŒ ๋œ๋‹ค. โ†’ High-dimensional anomaly detection has been a long-standing problem.
  • ์‹œ๊ฐ„์ , ๊ณต๊ฐ„์ , ๊ทธ๋ž˜ํ”„ ๊ธฐ๋ฐ˜ ๋ฐ ๊ธฐํƒ€ ์ƒํ˜ธ ์˜์กด ๊ด€๊ณ„์™€ ๊ฐ™์ด ์„œ๋กœ ์˜์กดํ•  ์ˆ˜ ์žˆ๋Š” ์ธ์Šคํ„ด์Šค์—์„œ ์ด์ƒ ํ˜„์ƒ์„ ํƒ์ง€ํ•˜๋Š” ๊ฒƒ๋„ ์–ด๋ ค์›€์ด ์žˆ์Œ.

- CH3: Data-efficient learning of normality/abnormality

  • ๋Œ€๊ทœ๋ชจ์˜ ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋œ ์ด์ƒ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•˜๋Š” ๋ฐ ๋”ฐ๋ฅธ ์–ด๋ ค์›€๊ณผ ๋น„์šฉ์œผ๋กœ ์ธํ•ด fully supervised anomaly detection๋Š” ์ •์ƒ ๋ฐ ์ด์ƒ ํด๋ž˜์Šค ๋ชจ๋‘์— ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋œ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์˜ ๊ฐ€์šฉ์„ฑ์„ ๊ฐ€์ •ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ข…์ข… ๋น„์‹ค์šฉ์ 
  • ๊ทธ๋ž˜์„œ, ์ง€๋‚œ 10๋…„๊ฐ„ unsupervised anomaly detection ์—ฐ๊ตฌ๊ฐ€ ์ง‘์ค‘๋จ. โ†’ ๊ทธ๋Ÿฌ๋‚˜,unsupervised anomaly detection์€ ์‹ค์ œ ์ด์ƒ ํ˜„์ƒ์— ๋Œ€ํ•œ ์‚ฌ์ „ ์ง€์‹์ด ์—†๊ณ , distribution of anomalies์— ๋Œ€ํ•œ ๊ฐ€์ •์— ํฌ๊ฒŒ ์˜์กดํ•จ.
  • ๋ฐ˜๋ฉด, ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋œ ์ •์ƒ ๋ฐ์ดํ„ฐ์™€ ์ผ๋ถ€ ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋œ ์ด์ƒ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•˜๋Š” ๊ฒƒ์€ ์–ด๋ ต์ง€ ์•Š์€ ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Œ. โ†’ ์‹ค์ œ๋กœ๋Š” ์‰ฝ๊ฒŒ ์ ‘๊ทผํ•  ์ˆ˜ ์žˆ๋Š” ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋œ ๋ฐ์ดํ„ฐ๋ฅผ ์ตœ๋Œ€ํ•œ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ์ด ์ข…์ข… ์ œ์•ˆ๋จ.
  • [Sol.1] Semi-supervised anomaly detection
    : ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋œ ์ •์ƒ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ฐ€์ •ํ•จ
  • [Sol.2] Weakly-supervised anomaly detection
    : anomaly label์„ ๋ถˆ์™„์ „ํ•˜๊ฒŒ ๊ฐ€์ง€๊ณ  ์žˆ๋‹ค๊ณ  ๊ฐ€์ •(anomalies์— ๋Œ€ํ•ด์„œ ๋ถ€๋ถ„์ /๋ถˆ์™„์ „, ๋ถ€์ •ํ™•ํ•œ ๋ ˆ์ด๋ธ”์„ ๊ฐ€์ง„๋‹ค๊ณ  ๊ฐ€์ •)

๐Ÿค” Semi-supervised learning ์ด๋ž€?

์ ์€ labeled data๊ฐ€ ์žˆ์œผ๋ฉด์„œ ์ถ”๊ฐ€๋กœ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋Œ€์šฉ๋Ÿ‰์˜ unlabeled data๊ฐ€ ์žˆ๋‹ค๋ฉด semi-supervised learning์„ ๊ณ ๋ คํ•  ์ˆ˜ ์žˆ๋‹ค. Semi-supervised learning (์ค€์ง€๋„ํ•™์Šต)์€ ์†Œ๋Ÿ‰์˜ labeled data์—๋Š” supervised learning์„ ์ ์šฉํ•˜๊ณ  ๋Œ€์šฉ๋Ÿ‰ unlabeled data์—๋Š” unsupervised learning์„ ์ ์šฉํ•ด ์ถ”๊ฐ€์ ์ธ ์„ฑ๋Šฅํ–ฅ์ƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋Š” ๋ฐฉ๋ฒ•๋ก ์ด๋‹ค. ์ด๋Ÿฐ ๋ฐฉ๋ฒ•๋ก ์— ๋‚ด์žฌ๋˜๋Š” ๋ฏฟ์Œ์€ label์„ ๋งž์ถ”๋Š” ๋ชจ๋ธ์—์„œ ๋ฒ—์–ด๋‚˜ ๋ฐ์ดํ„ฐ ์ž์ฒด์˜ ๋ณธ์งˆ์ ์ธ ํŠน์„ฑ์ด ๋ชจ๋ธ๋ง ๋œ๋‹ค๋ฉด ์†Œ๋Ÿ‰์˜ labeled data๋ฅผ ํ†ตํ•œ ์•ฝ๊ฐ„์˜ ๊ฐ€์ด๋“œ๋กœ ์ผ๋ฐ˜ํ™” ์„ฑ๋Šฅ์„ ๋Œ์–ด์˜ฌ๋ฆด ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ด๋‹ค. (์•„๋ž˜์˜ ๊ทธ๋ฆผ์—์„œ๋„ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋“ฏ์ด ์™ผ์ชฝ์— supervised learning๊ณผ ์˜ค๋ฅธ์ชฝ semi-supervised learning์„ ๋น„๊ตํ•ด๋ณด๋ฉด supervised learning์˜ decision boundary๋Š” ์‚ฌ์‹ค์ƒ optimalํ•˜์ง€ ์•Š๋‹ค. unlabeled data๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๋ฐ์ดํ„ฐ ์ž์ฒด์˜ ๋ถ„ํฌ๋ฅผ ์ž˜ ๋ชจ๋ธ๋งํ•˜๋ฉด ์˜ค๋ฅธ์ชฝ semi-supervised learning ๊ทธ๋ฆผ์ฒ˜๋Ÿผ ๋” ์ข‹์€ decision boundary๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ๋‹ค.)

- CH4: Noise-resilient anomaly detection.

Many weakly/semi-supervised anomaly detection method์€ ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋œ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๊ฐ€ ๊นจ๋—ํ•˜๋‹ค๊ณ  ๊ฐ€์ • -> ์‹ค์ˆ˜๋กœ ๋ฐ˜๋Œ€ ํด๋ž˜์Šค ๋ ˆ์ด๋ธ”๋กœ ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋œ noisyํ•œ ์ธ์Šคํ„ด์Šค์— ์ทจ์•ฝํ•  ์ˆ˜ ์žˆ์Œ. -> ์ด๋Ÿฌํ•œ ๊ฒฝ์šฐ ๋น„์ง€๋„ ๋ฐฉ๋ฒ•์„ ๋Œ€์‹  ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์ง€๋งŒ ์ด๋Š” ์‹ค์ œ ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋œ ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•˜์ง€ ๋ชปํ•จ.
Noise-resilient model์€ ๋ณด๋‹ค ์ •ํ™•ํ•œ ๊ฐ์ง€๋ฅผ ์œ„ํ•ด ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋˜์ง€ ์•Š์€ ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
(์—ฌ๊ธฐ์˜ ๋…ธ์ด์ฆˆ๋Š” ๋ ˆ์ด๋ธ”์ด ์ž˜๋ชป ์ง€์ •๋œ ๋ฐ์ดํ„ฐ์ด๊ฑฐ๋‚˜ ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋˜์ง€ ์•Š์€ ๊ฒฝ์šฐ์„ ๋งํ•จ)

- CH5: Detection of complex anomalies.

๊ธฐ์กด ๋ฐฉ๋ฒ•์˜ ๋Œ€๋ถ€๋ถ„์€ point anomalies์— ๋Œ€ํ•œ ๊ฒƒ์ด๋ฉฐ conditional anomaly and group anomaly์—๋Š” ์  ์ด์ƒ๊ณผ ์ „ํ˜€ ๋‹ค๋ฅธ ํ–‰์œ„๋ฅผ ๋‚˜ํƒ€๋‚ด๊ธฐ ๋•Œ๋ฌธ์— ์‚ฌ์šฉํ•  ์ˆ˜ ์—†๋‹ค.
-> conditional/group anomalies ๊ฐœ๋…์„ ์ด์ƒ ์ธก์ •/๋ชจ๋ธ์— ํ†ตํ•ฉํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
๋˜ํ•œ ํ˜„์žฌ ๋ฐฉ๋ฒ•์€ ์ฃผ๋กœ ๋‹จ์ผ ๋ฐ์ดํ„ฐ ์†Œ์Šค์—์„œ ์ด์ƒ์„ ๊ฐ์ง€ํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘๋Š” ๋ฐ˜๋ฉด, ๋งŽ์€ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์—์„œ๋Š” -> ๋‹ค์ฐจ์› ๋ฐ์ดํ„ฐ, ๊ทธ๋ž˜ํ”„, ์ด๋ฏธ์ง€, ํ…์ŠคํŠธ ๋ฐ ์˜ค๋””์˜ค ๋ฐ์ดํ„ฐ์™€ ๊ฐ™์€ multiple heterogeneous data sources๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ด์ƒ์„ ๊ฐ์ง€ํ•ด์•ผ ํ•œ๋‹ค.

- CH6: Anomaly explanation

Anomalies๋กœ ๋ณด๊ณ ๋œ rare data ์ธ์Šคํ„ด์Šค๋Š” ์‚ฌ๊ธฐ ํƒ์ง€ ๋ฐ ๋ฒ”์ฃ„ ํƒ์ง€ ์‹œ์Šคํ…œ์—์„œ ๊ณผ์†Œ๋Œ€ํ‘œ๋œ ๊ทธ๋ฃน๊ณผ ๊ฐ™์ด ๋ฐ์ดํ„ฐ์— ํ‘œ์‹œ๋œ ์†Œ์ˆ˜ ๊ทธ๋ฃน์— ๋Œ€ํ•œ ์•Œ๊ณ ๋ฆฌ์ฆ˜ ํŽธํ–ฅ ๊ฐ€๋Šฅ์„ฑ(possible algorithmic bias)์„ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
->ํŠน์ • ๋ฐ์ดํ„ฐ ์ธ์Šคํ„ด์Šค๊ฐ€ anomaly๋กœ ์‹๋ณ„๋˜๋Š” ์ด์œ ์— ๋Œ€ํ•œ ์ง์ ‘์ ์ธ ๋‹จ์„œ๋ฅผ ์ œ๊ณตํ•˜๋Š” anomaly explanation algorithm์„ ๊ฐ–์ถ”๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
(๋Œ€๋ถ€๋ถ„์˜ ์ด์ƒ ํƒ์ง€ ์—ฐ๊ตฌ๋Š” ํƒ์ง€ ์ •ํ™•๋„์—๋งŒ ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์‹๋ณ„๋œ ์ด์ƒ์— ๋Œ€ํ•œ ์„ค๋ช…์„ ์ œ๊ณตํ•˜๋Š” ๊ธฐ๋Šฅ์„ ๋ฌด์‹œํ•ฉ๋‹ˆ๋‹ค. ํŠน์ • ํƒ์ง€ ๋ฐฉ๋ฒ•์—์„œ anomaly explanation ์„ ๋„์ถœํ•˜๋Š” ๊ฒƒ์€ ํŠนํžˆ ๋ณต์žกํ•œ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ์•„์ง ํ•ด๊ฒฐ๋˜์ง€ ์•Š์€ ๋ฌธ์ œ์ž…๋‹ˆ๋‹ค. ๋ณธ์งˆ์ ์œผ๋กœ ํ•ด์„ ๊ฐ€๋Šฅํ•œ ์ด์ƒ ํƒ์ง€ ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ๋„ ์ค‘์š”ํ•˜์ง€๋งŒ, ๋ชจ๋ธ์˜ ํ•ด์„ ๊ฐ€๋Šฅ์„ฑ๊ณผ ํšจ์œจ์„ฑ์˜ ๊ท ํ˜•์„ ์ž˜ ๋งž์ถ”๋Š” ๊ฒƒ์ด ์ฃผ์š” ๊ณผ์ œ๋กœ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค.)

Deep Learning Methods vs. Traditional Methods in Anomaly Detection.


3.Addressing the Challenges with Deep Anomaly Detection

3.1. Preliminaries

3.2. Categorization of Deep Anomaly Detection

ํ•ด๋‹น ์˜์—ญ์— ๋Œ€ํ•œ ์ฒ ์ €ํ•œ ์ดํ•ด๋ฅผ ์œ„ํ•ด ๋ชจ๋ธ๋ง ๊ด€์ ์—์„œ Deep Anomaly Detection์„ 3 ๊ฐ€์ง€ ์ฃผ์š” ๋ฒ”์ฃผ์™€ 11๊ฐœ์˜ ์„ธ๋ถ„ํ™”๋œ ๋ฒ”์ฃผ๋กœ ๋ถ„๋ฅ˜ํ•˜๋Š” ๊ณ„์ธต์  ๋ถ„๋ฅ˜๋ฒ•์„ ๋„์ž…ํ•ฉ๋‹ˆ๋‹ค.
(๊ฐ ๋ฐฉ๋ฒ• ๋ฒ”์ฃผ์—์„œ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ๋Š” ๊ฐ์ง€ ๋ฌธ์ œ๋„ Challenge๋กœ ์ œ์‹œ๋ฉ๋‹ˆ๋‹ค)


4. Deep Learning for Feature Extraction

  • ๋ชฉํ‘œ : ๊ณ ์ฐจ์› ๋ฐ/๋˜๋Š” ๋น„์„ ํ˜• ๋ถ„๋ฆฌ ๊ฐ€๋Šฅํ•œ ๋ฐ์ดํ„ฐ์—์„œ ์ €์ฐจ์› ํŠน์ง• ํ‘œํ˜„(feature representations)์„ ์ถ”์ถœํ•˜๋Š” ๊ฒƒ
  • ํŠน์ง• ์ถ”์ถœ(Feature Extraction)๊ณผ ์ด์ƒ์น˜ ์ ์ˆ˜ ๋งค๊ธฐ๊ธฐ(anomaly scoring)๋Š” ์„œ๋กœ ์™„์ „ํžˆ ๋ถ„๋ฆฌ๋˜๊ณ  ๋…๋ฆฝ์ 

(normally ๐ท โ‰ซ ๐พ)

  • ์ฐจ์›์ถ•์†Œ(PCA, random projection) ๊ฐ™์€ ๋ฐฉ๋ฒ•๋ก ๊ณผ ๋น„๊ตํ–ˆ์„ ๋•Œ, ๋”ฅ๋Ÿฌ๋‹ ๊ธฐ์ˆ ์€ ์˜๋ฏธ๊ฐ€ ํ’๋ถ€ํ•œ ํŠน์ง•๊ณผ ๋น„์„ ํ˜• ํŠน์ง• ๊ด€๊ณ„๋ฅผ ์ถ”์ถœํ•˜๋Š” ๋ฐ ํ›จ์”ฌ ๋” ๋‚˜์€ ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ๋‹ค.

  • assumption : ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์— ์˜ํ•ด ์ถ”์ถœ๋œ feature representation์€ ๋ณ€์น™์ ์ธ ์ƒํ™ฉ๊ณผ ์ •์ƒ์ ์ธ ์ƒํ™ฉ์„ ๊ตฌ๋ถ„ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๋Š” ์‹๋ณ„ ์ •๋ณด(discriminative information)๋ฅผ ๋ณด์กดํ•ฉ๋‹ˆ๋‹ค.

  • AlexNet, VGG ๋ฐ ResNet ๋“ฑ์„ ํ†ตํ•ด ์ €์ฐจ์› ํŠน์ง•์„ ์ถ”์ถœํ•œ๋‹ค. ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ, ๋น„๋””์˜ค ๋ฐ์ดํ„ฐ์™€ ๊ฐ™์€ ๋ณต์žกํ•œ ๊ณ ์ฐจ์› ๋ฐ์ดํ„ฐ์˜ ์ด์ƒ ํƒ์ง€์—์„œ ํƒ์ƒ‰๋œ ๋ฐฉ๋ฒ•๋ก ์ด๋‹ค.

ํŠน์„ฑ ์ถ”์ถœ ๊ธฐ๋ฒ•

  • ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ์…‹์—์„œ ํŠน์„ฑ์„ ์ถ”์ถœํ•˜๋Š” ๊ธฐ๋ฒ•์ž…๋‹ˆ๋‹ค.
  • ํŠน์„ฑ ์ถ”์ถœ ๊ธฐ๋ฒ•์€ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์—์„œ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ๋ฅผ ์ž…๋ ฅ์œผ๋กœ ๋ฐ›์•„์„œ ํ•ฉ์„ฑ๊ณฑ์ธต(convolutional layer)์„ ํ†ตํ•ด ์ด๋ฏธ์ง€์˜ ํŠน์„ฑ์„ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ์ถ”์ถœ๋œ ํŠน์„ฑ์€ ์ตœ์ข…์ ์œผ๋กœ ์™„์ „ ์—ฐ๊ฒฐ์ธต(fully connected layer)์„ ํ†ตํ•ด ๋ถ„๋ฅ˜๋ฉ๋‹ˆ๋‹ค.
  • ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ์…‹์—์„œ ํŠน์„ฑ ์ถ”์ถœ์„ ์œ„ํ•ด์„œ๋Š”, ๋จผ์ € ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์˜ ํ•ฉ์„ฑ๊ณฑ์ธต์„ ๊ณ ์ •(freeze)์‹œํ‚ค๊ณ , ์™„์ „ ์—ฐ๊ฒฐ์ธต ๋ถ€๋ถ„๋งŒ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ์…‹์— ๋งž๊ฒŒ ๋ณ€๊ฒฝํ•˜์—ฌ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค.
  • ๋”ฐ๋ผ์„œ, ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ์…‹์—์„œ๋Š” ์ด๋ฏธ์ง€์˜ ์นดํ…Œ๊ณ ๋ฆฌ๋ฅผ ๊ฒฐ์ •ํ•˜๋Š” ๋ถ€๋ถ„์ธ ๋งˆ์ง€๋ง‰ ์™„์ „ ์—ฐ๊ฒฐ์ธต๋งŒ ํ•™์Šตํ•˜๊ณ , ๋‚˜๋จธ์ง€ ํ•ฉ์„ฑ๊ณฑ์ธต ๊ณ„์ธต๋“ค์€ ์‚ฌ์ „ ํ›ˆ๋ จ๋œ ๊ฐ€์ค‘์น˜๋ฅผ ๊ทธ๋Œ€๋กœ ์‚ฌ์šฉํ•˜์—ฌ ํ•™์Šต๋˜์ง€ ์•Š๋„๋ก ๊ณ ์ •ํ•ฉ๋‹ˆ๋‹ค.

Unmasking framework for online anomaly detection

  • key idea : ์—ฐ์†๋œ ๋‘ ๋น„๋””์˜ค ์‹œํ€€์Šค๋ฅผ ๊ตฌ๋ณ„ํ•˜๊ธฐ ์œ„ํ•ด ์ด์ง„ ๋ถ„๋ฅ˜๊ธฐ๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ํ›ˆ๋ จ์‹œํ‚ค๋ฉด์„œ, ๊ฐ ๋‹จ๊ณ„์—์„œ ๊ฐ€์žฅ ๊ตฌ๋ณ„๋ ฅ ์žˆ๋Š” ํŠน์ง•(discriminant feature)์„ ์ ์ง„์ ์œผ๋กœ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. => ๋น„๋””์˜ค์˜ ๊ฐ ์žฅ๋ฉด(Scene)์ด ๋น„์ •์ƒ์ ์ธ์ง€ ์—ฌ๋ถ€๋ฅผ ํŒ๋‹จํ•˜๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜
    (์ธ์šฉํ•œ ๋…ผ๋ฌธ์€ ๋”ฐ๋กœ ๋ฒจ๋กœ๊ทธ์— ์ •๋ฆฌํ•˜์˜€๋‹ค.)
  • ILSVRC ๋ฒค์น˜๋งˆํฌ์—์„œ ์‚ฌ์ „ ํ›ˆ๋ จ๋œ VGG ๋ชจ๋ธ (135)๋Š” ์ด๋Ÿฌํ•œ ๋ชฉ์ ์„ ์œ„ํ•ด ํ‘œํ˜„์ ์ธ ์™ธ๊ด€ ํŠน์ง•(appearance features)์„ ์ถ”์ถœํ•˜๋Š” ๋ฐ ํšจ๊ณผ์ 

Downstream Anomaly Scoring์— pre-trained model์ด ์•„๋‹Œ deep feature extraction model์„ ๋ช…์‹œ์ ์œผ๋กœ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ

  • ๋…ผ๋ฌธ์—์„œ๋Š” ํŠน์ง• ํ‘œํ˜„์„ ์ž๋™์œผ๋กœ ํ•™์Šตํ•˜๊ธฐ ์œ„ํ•ด ๋”ฅ๋Ÿฌ๋‹ ๋„คํŠธ์›Œํฌ๋ฅผ ํ™œ์šฉํ•˜๋Š” Appearance and Motion DeepNet (AMDN)์„ ์ œ์•ˆ
  • ์Šคํƒ๋œ ๋…ธ์ด์ฆˆ ์ œ๊ฑฐ ์˜คํ† ์ธ์ฝ”๋”(stacked denoising autoencoders)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์™ธ๋ชจ์™€ ๋™์ž‘ ํŠน์ง• ๋ฐ ๊ฒฐํ•ฉ๋œ ํ‘œํ˜„(์ดˆ๊ธฐ ์œตํ•ฉ)์„ ๊ฐ๊ฐ ํ•™์Šต
  • ํ•™์Šต๋œ ํ‘œํ˜„์„ ๊ธฐ๋ฐ˜์œผ๋กœ, 3๊ฐœ์˜ ์ผ๋ฅ˜ SVM(one-class SVM) ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ ์ž…๋ ฅ์˜ ์ด์ƒ ์ ์ˆ˜๋ฅผ ์˜ˆ์ธกํ•˜๋ฉฐ, ์ด๋Ÿฌํ•œ ์ ์ˆ˜๋“ค์€ ํ›„๊ธฐ ์œตํ•ฉ ์ „๋žต์„ ํ†ตํ•ด ํ†ตํ•ฉ๋˜์–ด ์ตœ์ข… ์ด์ƒ ๊ฐ์ง€๋ฅผ ์ˆ˜ํ–‰

Linear one-class SVM + DBN

๋…ผ๋ฌธ์—์„œ One-class Support Vector Machines์€ Deep Belief net(DBN)์—์„œ ์ƒ์„ฑ๋œ ๊ณ ์ฐจ์› ํ…Œ์ด๋ธ” ํ˜•์‹ ๋ฐ์ดํ„ฐ์˜ ์ €์ฐจ์› ํ‘œํ˜„์— ๋Œ€ํ•œ ์ด์ƒ ํƒ์ง€๋ฅผ ํ™œ์„ฑํ™”ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ.

โ—์žฅ๋‹จ์ 

  • ์žฅ์  :
    (i) state-of-the-art(pre-trained) deep model๊ณผ ๊ธฐ์„ฑ ์ด์ƒ ํƒ์ง€๊ธฐ๋ฅผ ์‰ฝ๊ฒŒ ์‚ฌ์šฉ
    (ii) Deep feature extraction์€ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ์„ ํ˜• ๋ฐฉ๋ฒ•๋ณด๋‹ค ๋” ๊ฐ•๋ ฅํ•œ ์ฐจ์› ์ถ•์†Œ๋ฅผ ์ œ๊ณต
    (iii) ์‹ฌ์ธต ๋ชจ๋ธ ๋ฐ ํƒ์ง€ ๋ฐฉ๋ฒ•์ด ๊ณต๊ฐœ์ ์œผ๋กœ ์ œ๊ณต๋˜๋ฏ€๋กœ ๊ตฌํ˜„์ด ์‰ฌ์›€

  • ๋‹จ์  :
    (i) ์™„์ „ํžˆ ๋ถ„๋ฆฌ๋œ feature extraction ๊ณผ anomaly scoring์€ ์ข…์ข… ์ตœ์ ์ด ์•„๋‹Œ ์ด์ƒ ์ ์ˆ˜(suboptimal anomaly scores)๋กœ ์ด์–ด์ง.
    (ii) Pre-trained deep model์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํŠน์ • ์œ ํ˜•์˜ ๋ฐ์ดํ„ฐ๋กœ ์ œํ•œ


5. Learning Feature Representations of Normality

5.1.Generic Normality Feature Learning

5.1.1.Autoencoders

[์˜คํ† ์ธ์ฝ”๋” ๋ชฉํ‘œ] : ๋ฐ์ดํ„ฐ ์ธ์Šคํ„ด์Šค๋ฅผ ์ž˜ ์žฌ๊ตฌ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ์ €์ฐจ์› ํŠน์ง• ํ‘œํ˜„ ๊ณต๊ฐ„์„ ํ•™์Šตํ•˜๋Š” ๊ฒƒ

  • ๋ฐ์ดํ„ฐ ์••์ถ•์ด๋‚˜ ์ฐจ์› ์ถ•์†Œ๋ฅผ ์œ„ํ•ด ๋„๋ฆฌ ์‚ฌ์šฉ๋จ
  • ์ด ๊ธฐ์ˆ ์„ ์ด์ƒ ํƒ์ง€์— ์‚ฌ์šฉํ•˜๋Š” ์ด์œ ๋Š” ํ•™์Šต๋œ ํŠน์ง• ํ‘œํ˜„(feature representation)์ด ์žฌ๊ตฌ์„ฑ ์˜ค๋ฅ˜๋ฅผ ์ตœ์†Œํ™”ํ•˜๊ธฐ ์œ„ํ•ด(minimize reconstruction errors) ๋ฐ์ดํ„ฐ์˜ ์ค‘์š”ํ•œ ๊ทœ์น™์„ฑ(important regularities of the data)์„ ํ•™์Šตํ•˜๋„๋ก ๊ฐ•์ œ๋˜๊ธฐ ๋•Œ๋ฌธ
  • ์ด์ƒ์น˜๋Š” ์žฌ๊ตฌ์„ฑํ•˜๊ธฐ ์–ด๋ ค์›Œ => ํฐ ์žฌ๊ตฌ์„ฑ ์˜ค๋ฅ˜๋ฅผ ๊ฐ€์ง€๊ฒŒ ๋จ

[์˜คํ† ์ธ์ฝ”๋” ๊ฐ€์ •] : ์ •์ƒ ์ธ์Šคํ„ด์Šค(Normal instance)๋Š” ์••์ถ•๋œ ๊ณต๊ฐ„(compressed space)์—์„œ ์ด์ƒ์น˜๋ณด๋‹ค ๋” ์ž˜ ์žฌ๊ตฌ์„ฑ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค

(์˜คํ† ์ธ์ฝ”๋”์— ๋Œ€ํ•œ ์„ค๋ช…์€ ์—ฌ๊ธฐ ๋ธ”๋กœ๊ทธ์„ ๋“ค์–ด๊ฐ€์„œ ํ•œ ๋ฒˆ ๋ณด๊ณ ์˜ค๋Š”๊ฒŒ ์ข‹๋‹ค.)

5.2.Anomaly Measure-dependent Feature Learning

๐Ÿ”– Reference
๋ณธ ๋…ผ๋ฌธ
๋”ฅ๋Ÿฌ๋‹์—์„œ์˜ end-to-end
semi-supervised learning
[pytorch] ์ „์ด ํ•™์Šต - ํŠน์„ฑ ์ถ”์ถœ ๊ธฐ๋ฒ• (Feature Extraction)
[Paper Review] Unmasking the abnormal events in video
[๋…ผ๋ฌธ๋ฆฌ๋ทฐ] Restricted Boltzmann Machine(RBM)์™€ Deep Belief Network(DBN

profile
๋ฐฐ์šฐ๊ณ  ๊ณต๋ถ€ํ•˜๊ณ  ๊ธฐ๋กํ•˜๋Š” ๊ฒƒ์„ ๋ฉˆ์ถ”์ง€ ์•Š๋Š”๋‹ค.

0๊ฐœ์˜ ๋Œ“๊ธ€