XAI

XAI Introduction

date
Jul 9, 2023
slug
xai-introduction
author
status
Public
tags
DeepLearning
XAI
summary
type
Post
thumbnail
category
XAI
updatedAt
Jul 9, 2023 02:35 PM
ย 
์ตœ๊ทผ ์ง€๋„ํ•™์Šต์— ๊ธฐ๋ฐ˜ํ•œ ํ”„๋ ˆ์ž„์›Œํฌ๊ฐ€ ํฐ ํ˜์‹ ์„ ์ด๋ฃจ๊ณ  ์žˆ๋‹ค. ๊ทธ์ค‘์—์„œ๋„ ๋”ฅ๋Ÿฌ๋‹์— ๊ธฐ๋ฐ˜ํ•œ ๊ธฐ์ˆ ๋“ค์ด ๋‹ค์–‘ํ•œ ๋ถ„์•ผ์˜ ์„ฑ๋Šฅ์„ ๋น„์•ฝ์ ์œผ๋กœ ๋ฐœ์ „์‹œํ‚ค๊ณ  ์žˆ๋‹ค.
notion image
notion image
์™ผ์ชฝ ๊ทธ๋ž˜ํ”„๋ฅผ ๋ณด๋ฉด 2012๋…„์„ ๊ธฐ์ ์œผ๋กœ ImageNet์ด๋ผ๋Š” ๋ฐ์ดํ„ฐ์…‹์˜ ์ธ์‹ ์—๋Ÿฌ์œจ์ด ๋”ฅ๋Ÿฌ๋‹์„ ์ ์šฉํ•˜๋ฉด์„œ ๋งค๋…„ ์ ˆ๋ฐ˜์œผ๋กœ ์ค„์–ด๋“œ๋Š” ํ˜์‹ ์ด ์žˆ์—ˆ๊ณ , ๊ทธ ์™ธ์—๋„ ์Œ์„ฑ์ธ์‹(speech recognition), ๊ธฐ๊ณ„๋ฒˆ์—ญ(machine translation) ๋“ฑ์—์„œ ๋†€๋ผ์šด ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๊ฐ€์ ธ์™€ ํ˜„์žฌ ๋งŽ์€ ์‚ฌ๋žŒ์˜ ์‚ถ์— ์ง์ ‘ ์˜ํ–ฅ์„ ๋ฏธ์น˜๊ณ  ์‘์šฉ ๋ถ„์•ผ์˜ ๋„๋ฆฌ ์ ์šฉ๋˜๊ณ  ์žˆ๋‹ค.

1. Limitation of supervised learning

์ด๋ฅผ ํ†ตํ•ด์„œ ์ธ๊ณต์ง€๋Šฅ AI์— ๋Œ€ํ•œ ์‚ฌํšŒ์ ์ธ ๊ด€์‹ฌ์ด ๋งค์šฐ ๋†’์•„์กŒ๊ณ , AI๊ฐ€ 4์ฐจ ์‚ฐ์—… ํ˜๋ช…์˜ ํ•ต์‹ฌ์ ์ธ ๊ธฐ์ˆ ๋กœ ์ž๋ฆฌ ์žก๊ฒŒ ๋˜์—ˆ๋‹ค. ํ•˜์ง€๋งŒ, ์ด๋ ‡๊ฒŒ ๋งค์šฐ ์ข‹์€ ์„ฑ๋Šฅ์„ ๋‚ด๋Š” ์ง€๋„ํ•™์Šต ๊ธฐ๋ฐ˜ ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์—๋„ ์ค‘์š”ํ•œ ํ•œ๊ณ„์ ์ด ์กด์žฌํ•œ๋‹ค.
ํ•œ๊ณ„์ ์€ ๋ฐ”๋กœ ๋Œ€์šฉ๋Ÿ‰ ํ•™์Šต ๋ฐ์ดํ„ฐ๋กœ๋ถ€ํ„ฐ ํ•™์Šตํ•˜๋Š” ๋ชจ๋ธ์˜ ๊ตฌ์กฐ๊ฐ€ ์ ์  ๋” ๋ณต์žกํ•ด์ง€๊ณ  ์ดํ•ดํ•˜๊ธฐ ์–ด๋ ค์›Œ์ง„๋‹ค๋Š” ์ ์ด๋‹ค.
๋‹จ์ˆœ์‹ ๊ฒฝ๋ง โ†’ ResNet/DesenNet โ†’ Transfomer
๋‹จ์ˆœ์‹ ๊ฒฝ๋ง โ†’ ResNet/DesenNet โ†’ Transfomer
์ดˆ์ฐฝ๊ธฐ ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์ธ ์™ผ์ชฝ์— ์žˆ๋Š” ๋‹จ์ˆœํ•œ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ๋กœ๋ถ€ํ„ฐ ์—ฌ๋Ÿฌ ๊ฐ€์ง€ Skip Connection์„ ๊ฐ€์ง€๊ณ  ์žˆ๋Š” ์—„์ฒญ๋‚˜๊ฒŒ ๊นŠ์€ ๊ตฌ์กฐ๋ฅผ ๊ฐ€์ง€๋Š” ResNet, DenseNet ๋“ฑ์œผ๋กœ ๋ฐœ์ „ํ•˜์˜€๊ณ , ์ตœ๊ทผ์—๋Š” ์˜ค๋ฅธ์ชฝ์— ๋ณด์‹œ๋Š” ๊ฒƒ๊ณผ ๊ฐ™์€ transformer๋ผ๋Š” attention ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ํ™œ์šฉํ•˜๋Š” ๋ชจ๋ธ๋กœ๋„ ๋ฐœ์ „ํ•˜๊ณ  ์žˆ๋‹ค.
์ด๋ ‡๊ฒŒ ์ ์  ์„ฑ๋Šฅ์ด ์ข‹์•„์ง€๋ฉด์„œ ์ ์  ๋ณต์žกํ•œ ๋ชจ๋ธ๋กœ ๋ฐœ์ „ํ•˜๋Š”๋ฐ, ๊ฒฐ๊ตญ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ์ž…๋ ฅ์„ ์ง‘์–ด๋„ฃ์œผ๋ฉด ์ถœ๋ ฅ์ด ํŠ€์–ด๋‚˜์˜ค๋Š” ํ•˜๋‚˜์˜ ๋ธ”๋ž™๋ฐ•์Šค์ฒ˜๋Ÿผ ๋™์ž‘ํ•˜๊ฒŒ ๋œ๋‹ค.
notion image
๊ทธ๋Ÿฐ๋ฐ ์ด๋ ‡๊ฒŒ ๋ณต์žกํ•œ ๊ตฌ์กฐ์˜ ๋ชจ๋ธ์„ ์‘์šฉ ๋ถ„์•ผ์— ์ ์šฉํ•  ๋•Œ, ์˜ํ™” ์ถ”์ฒœ๊ณผ ๊ฐ™์ด ๋‹จ์ˆœํžˆ ์˜ˆ์ธก ๊ฒฐ๊ณผ๋งŒ ํ•„์š”ํ•œ ๊ฒฝ์šฐ์—๋Š” ํฐ ๋ฌธ์ œ๊ฐ€ ์—†์ง€๋งŒ, ๋ชจ๋ธ์˜ ์˜ˆ์ธก ๊ฒฐ๊ณผ๊ฐ€ ์‚ฌ๋žŒ์—๊ฒŒ ์ง์ ‘ ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ๋Š” ๊ฒฝ์šฐ๋Š” ๋ฌธ์ œ๊ฐ€ ์ƒ๊ธธ ์ˆ˜ ์žˆ๋‹ค.
์ฆ‰, ์ด๋ฏธ์ง€ ์ธ์‹ ๊ธฐ์ˆ ์— ๊ธฐ๋ฐ˜ํ•œ ์ž์œจ์ฃผํ–‰ ์ž๋™์ฐจ๊ฐ€ ์‚ฌ๊ณ ๋ฅผ ๋ƒˆ์„ ๋•Œ, ์˜๋ฃŒ์˜์ƒ ๊ธฐ๋ฐ˜ ์งˆ๋ณ‘ ์ง„๋‹จ์„ ํ•˜๋Š” ๊ฒฝ์šฐ, AI ๊ธฐ๋ฐ˜ ์ธํ„ฐ๋ทฐ, AI ๊ธฐ๋ฐ˜ ๋Œ€์ถœ ์—ฌ๋ถ€ ๊ฒฐ์ • ๋“ฑ์„ ํ•˜๋Š” ๊ฒฝ์šฐ ๋‹จ์ˆœํžˆ ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์ด ๋‚ด์–ด ์ฃผ๋Š” ์˜ˆ์ธก ๊ฒฐ๊ณผ๋งŒ์„ ์‚ฌ์šฉํ•˜๊ธฐ์—๋Š” ํ•œ๊ณ„๊ฐ€ ์žˆ๊ณ , ์™œ ํ•ด๋‹น ๋ชจ๋ธ์ด ์ด๋Ÿฌํ•œ ์˜ˆ์ธก ๊ฒฐ๊ณผ๋ฅผ ๋‚ด์—ˆ๋Š”์ง€ ๊ทธ ์ด์œ ๋ฅผ ์„ค๋ช…ํ•ด์•ผ ํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋ฐ˜๋“œ์‹œ ์กด์žฌํ•œ๋‹ค.
๊ทธ๋ฆฌ๊ณ  ๊ทธ๋ ‡๊ฒŒ ์„ค๋ช…์„ ์ œ๊ณตํ•˜์ง€ ๋ชปํ•˜๋Š” ๊ฒฝ์šฐ,๋ฒ•์  ๋ฐ ์‚ฌํšŒ์ ์œผ๋กœ ํ•ด๋‹น ๋ชจ๋ธ์„ ์ ์šฉํ•˜๋Š”๋ฐ ์ œ์•ฝ์ด ์ƒ๊ธธ ์ˆ˜ ์žˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด์„œ ์ตœ๊ทผ์— ๋Œ€๋‘๋˜๊ณ  ์žˆ๋Š” AI ๋ชจ๋ธ์˜ ํŽธํ–ฅ์„ฑ ๋ฌธ์ œ๋ฅผ ๋ณด๋ฉด ์„ค๋ช… ๊ฐ€๋Šฅ ์ธ๊ณต์ง€๋Šฅ์ด ๊ผญ ํ•„์š”ํ•˜๋ฉด ์•Œ ์ˆ˜ ์žˆ๋‹ค.
notion image
๋”ฅ๋Ÿฌ๋‹ ๊ธฐ๋ฐ˜ย ๋Œ€ํ™” ๋ชจ๋ธ์ด๋‚˜ย ์ด๋ฏธ์ง€ ์ธ์‹ย ํƒœ๊น… ๋ฌธ์ œ์—์„œย ํŠน์ • ์ธ์ข…์ด๋‚˜ ์„ฑ๋ณ„์— ํŽธํ–ฅ๋œ ๊ฒฐ๊ณผ ๋ฐย ์˜ˆ์ธก์„ย ๋งŒ๋“ค์–ด๋‚ด๋Š” ๊ฒฝ์šฐ๊ฐ€ ์žˆ๋‹ค.
์˜ˆ๋ฅผ ๋“ค์–ด, Google Photo์—์„œ ์–ด๋–ค ํ‘์ธ ์—ฌ์„ฑ์„ ๊ณ ๋ฆด๋ผ๋ผ๊ณ  ์ž˜๋ชป ์ธ์‹ํ•˜๊ฑฐ๋‚˜, ๋‰ด์งˆ๋žœ๋“œ ์—ฌ๊ถŒ ์ธ์‹ ํ”„๋กœ๊ทธ๋žจ์ด ์•„์‹œ์•„๊ณ„ ๋‚จ์„ฑ์ด ๋ˆˆ์„ ๊ฐ๊ณ  ์žˆ๋‹ค๊ณ ย ์˜ค์ธ์‹ํ•˜๋Š”ย ๊ฒฝ์šฐ์— ๋ชจ๋ธ์ด ์™œ ์ธ์ข…๊ณผ ๊ด€๋ จํ•˜์—ฌ ์ž˜๋ชป๋œย ์˜ˆ์ธก ๊ฒฐ๊ณผ๋ฅผย ๋‚ด์—ˆ๊ณ , ์ด๋Ÿฌํ•œ ํŽธํ–ฅ์„ฑ(๋ฐ”์ด์–ด์Šค)์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ์ด ๋ชจ๋ธ์„ ๊ฒฐ๊ณผ๋ฅผ ์„ค๋ช…ํ•  ์ˆ˜ ์žˆ๋Š”ย ์„ค๋ช… ๊ฐ€๋Šฅย ์ธ๊ณต์ง€๋Šฅ์ดย ๊ผญ ํ•„์š”ํ•˜๋‹ค.

2. Some Examples of Explainable AI

2-1) example 1

notion image
์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜ ์—ฐ๊ตฌ๋ฅผ ์œ„ํ•ด ๋ฒค์น˜๋งˆํฌ๋กœ ๋„๋ฆฌ ์“ฐ์ด๋Š” PASCAL VOC๋ผ๋Š” ๋ฐ์ดํ„ฐ์…‹์ด ์žˆ๋‹ค.
๋…์ผ์˜ ํ•œ ์—ฐ๊ตฌ์ง„์ด XAI๊ธฐ๋ฒ•์„ PASCAL VOC Data set์œผ๋กœ ํ›ˆ๋ จํ•œ ๋ถ„๋ฅ˜๊ธฐ์— ์ ์šฉํ•ด๋ณด์•˜๋‹ค.
์ฆ‰, ์–ด๋–ค ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์ด ์ด๋ฏธ์ง€๋ฅผ ๋ฐ›์•„, ์˜ˆ๋ฅผ ๋“ค์–ด ๋ง์— ๋Œ€ํ•œ ์‚ฌ์ง„์ด๋ผ๊ณ  ๋ถ„๋ฅ˜ํ•˜์˜€๋Š”๋ฐ, XAI๊ธฐ๋ฒ•์€ ๊ณผ์—ฐ ํ•ด๋‹น ๋ชจ๋ธ์ด ์ด๋ฏธ์ง€ ์–ด๋Š ๋ถ€๋ถ„์„ ๋ณด๊ณ  ์ด๋ฏธ์ง€๋ฅผ "๋ง"์ด๋ผ๊ณ  ํ–ˆ๋Š”์ง€ ๋นจ๊ฐ„์ƒ‰์œผ๋กœ ํ•˜์ด๋ผ์ดํŠธ๋ฅผ ํ•ด์„œ ๋ณด์—ฌ ์ฃผ๋Š” ๋ฐฉ์‹์ด๋‹ค.
์œ„์˜ ์‚ฌ์ง„์—์„œ ์™ผ์ชฝ์— ์žˆ๋Š” ๋‹ค์„ฏ ๊ฐœ์˜ ์ด๋ฏธ์ง€๊ฐ€ ๋ชจ๋‘ ๋ง์— ๊ด€ํ•œ ์ด๋ฏธ์ง€์ด๊ณ , ํ•ด๋‹น ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์€ ๋ชจ๋‘ ์ œ๋Œ€๋กœ ์ •๋‹ต ํด๋ž˜์Šค์ธ ๋ง์„ ์ œ๋Œ€๋กœ ์˜ˆ์ธกํ•˜์˜€๋‹ค.
๊ทธ๋Ÿฐ๋ฐ XAI์˜ ๊ธฐ๋ฒ•์€ ๋ชจ๋ธ์ด ์ด์ƒํ•˜๊ฒŒ๋„ ์ด๋ฏธ์ง€ ๋‚ด์—์„œ ๋ง์— ํ•ด๋‹นํ•˜๋Š” ๋ถ€๋ถ„์ด ์•„๋‹ˆ๋ผ ์ด๋ฏธ์ง€ ์•„๋ž˜์ชฝ์„ ์ฃผ๋กœ ๋ณด๊ณ  ์žˆ๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ฃผ์—ˆ๋‹ค.
๊ทธ๋ž˜์„œ ์—ฐ๊ตฌ์ง„๋“ค์€ ๋‹ค์‹œ ํ•ด๋‹น ๋ง ์‚ฌ์ง„๋“ค์„ ํ™•์ธ์„ ํ•ด๋ณด๋‹ˆ ์‚ฌ์ง„ ํ•˜๋‹จ๋ถ€์— ์˜ค๋ฅธ์ชฝ ๊ทธ๋ฆผ์—์„œ ๋ณด๋Š” ๊ฒƒ์ฒ˜๋Ÿผ ํ…์ŠคํŠธ๋กœ ์›Œํ„ฐ๋งˆํฌ๊ฐ€ ์žˆ๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ•˜์˜€๊ณ . ์ด ๋ฐ์ดํ„ฐ์…‹์˜ ๋ฒ„๊ทธ๊ฐ€ ์žˆ๋Š” ๊ฒƒ์„ ์•Œ๊ฒŒ ๋˜์—ˆ๋‹ค.
์ฆ‰, ๋‹ค๋ฅธ ํด๋ž˜์Šค์— ์‚ฌ์ง„๋“ค์€ ์ด๋Ÿฐ ๋ฒ„๊ทธ๊ฐ€ ์—†์—ˆ์ง€๋งŒ, ๋ฐ์ดํ„ฐ ์…‹์— ์žˆ๋Š” ๋ชจ๋“  ๋ง ๊ด€๋ จ ์‚ฌ์ง„๋“ค์—๋Š” ์ด๋Ÿฐ ํ…์ŠคํŠธ ์›Œํ„ฐ๋งˆํฌ๊ฐ€ ์žˆ์—ˆ๊ณ  ํ•™์Šตํ•œ ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ ์‹ค์ œ ๋ง ์‚ฌ์ง„์ด ์•„๋‹Œ, ํ…์ŠคํŠธ ์›Œํ„ฐ๋งˆํฌ์˜ ๊ธฐ๋ฐ˜์—์„œ ๋ง์ด๋ผ๊ณ  ๋ถ„๋ฅ˜ํ•˜๊ณ  ์žˆ์—ˆ๋˜ ๊ฒƒ์ด์—ˆ๋‹ค.
์ด ๋ฐ์ดํ„ฐ์…‹์—์„œ์˜ ์ •ํ™•๋„๋Š” ๋†’์•˜๊ฒ ์ง€๋งŒ, ์ผ๋ฐ˜์ ์ธ ์ƒํ™ฉ์— ์ ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์ œ๋Œ€๋กœ ๋œ ๋ง ๋ถ„๋ฅ˜๊ธฐ๋ผ๊ณ  ํ•  ์ˆ˜๋Š” ์—†์„ ๊ฒƒ์ด๋‹ค.
์ด์ฒ˜๋Ÿผ XAI์˜ ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉํ•˜๋ฉด ๋ชจ๋ธ์ด๋‚˜ ๋ฐ์ดํ„ฐ์…‹์ด ๊ฐ€์ง€๊ณ  ์žˆ๋Š” ์˜ค๋ฅ˜๋ฅผ ์ฐพ์•„๋‚ผ ์ˆ˜ ์žˆ๋‹ค.

2-2) example 2

notion image
๋‘ ๋ฒˆ์งธ ์˜ˆ์‹œ๋กœ, XAI ๊ธฐ๋ฒ•์ด ๋ชจ๋ธ์ด ์–ผ๋งˆ๋‚˜ ํŽธํ–ฅ๋˜์–ด ์žˆ๋Š”์ง€๋„ ์•Œ์•„๋‚ผ ์ˆ˜ ์žˆ๋‹ค.
๋ฏธ๊ตญ์—์„œ ์‹ค์ œ๋กœ COMPAS ๋ผ๋Š” ๋ฒ”์ฃ„ ์žฌ๋ฒ”๋ฅ  ์˜ˆ์ธก ์‹œ์Šคํ…œ์„ ์„ค๋ช…ํ•ด ๋ณธ ๊ฒƒ์ด๋‹ค. ์ด ์‹œ์Šคํ…œ์€ ํ˜„์žฌ ๊ต๋„์†Œ์— ์ˆ˜๊ฐ๋˜์–ด์žˆ๋Š” ์‚ฌ๋žŒ ์ค‘์—์„œ ์–ด๋–ค ์‚ฌ๋žŒ์„ ์ง‘ํ–‰์œ ์˜ˆ๋กœ ํ’€์–ด์คฌ์„ ๋•Œ ๋‹ค์‹œ ๋ฒ”์ฃ„๋ฅผ ์ €์งˆ๋Ÿฌ์„œ ๊ต๋„์†Œ๋กœ ๋Œ์•„์˜ฌ์ง€๋ฅผ ์˜ˆ์ธกํ•˜๋Š” ์‹œ์Šคํ…œ์ด๋‹ค.
๊ทธ๋Ÿฐ๋ฐ ์˜ค๋ฅธ์ชฝ ํ‘œ๋ฅผ ๋ณด์‹œ๋ฉด ์‚ฌ๋žŒ์ด ์˜ˆ์ธกํ•œ ๊ฒƒ(Humans)๊ณผ COMPAS๋ผ๊ณ  ํ•˜๋Š” ์‹œ์Šคํ…œ์ด ์˜ˆ์ธกํ•œ ๊ฒƒ์— ์ „์ฒดํ‰๊ท  ์ •ํ™•๋„๋Š” 60 ๋ช‡ ํผ์„ผํŠธ ์ •๋„๋กœ ๋น„์Šทํ•œ ๊ฒฐ๊ณผ์— ๋ฐ˜ํ•ด์„œ, ํ‘์ธ๊ณผ ๋ฐฑ์ธ์œผ๋กœ ๋‚˜๋ˆ ์„œ ์ •ํ™•๋„๋ฅผ ์„ธ๋ถ„ํ™”ํ•ด์„œ ๋ณด๊ฒŒ ๋˜๋ฉด ํ‘์ธ์— ๋Œ€ํ•œ first positive rate์™€ ๋ฐฑ์ธ์— ๋Œ€ํ•œ first negative rate๊ฐ€ COMPAS์—์„œ ์‚ฌ๋žŒ ์˜ˆ์ธก(humans)๋ณด๋‹ค ํ›จ์”ฌ ๋” ๋†’๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ๋‹ค.
์ฆ‰, XAI๋กœ ์ž๋™ํ™” ์‹œ์Šคํ…œ์˜ ํŽธํ–ฅ๋œ ๋ฌธ์ œ๋ฅผ ์ฐพ์•„๋‚ธ ๊ฒƒ์ด๋‹ค.

2-3) example 3

notion image
์ž์œจ์ฃผํ–‰ ์ž๋™์ฐจ๊ฐ€ ์ž˜๋ชป ๋™์ž‘ํ•ด์„œ ์‚ฌ๊ณ ๊ฐ€ ๋ฐœ์ƒํ–ˆ๋Š”๋ฐ, ์ด๋•Œ ์ธ์‹ ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ์™œ ๊ทธ๋Ÿฐ ์ž˜๋ชป๋œ ๋™์ž‘์„ ํ•˜๊ฒŒ ๋˜์—ˆ๋Š”์ง€ ์„ค๋ช…ํ•˜๋Š” ๊ฒƒ์€ ๋งค์šฐ ์ค‘์š”ํ•˜๋‹ค. ๋”ฐ๋ผ์„œ, ์‚ฌ๊ณ  ์ฑ…์ž„ ์†Œ์žฌ๋ฅผ ์ฐพ์„ ์ˆ˜๋„ ์žˆ๊ณ , ์˜ค๋ฅ˜๋ฅผ ๊ณ ์ณ์„œ ํ–ฅํ›„ ๋” ์•ˆ์ „ํ•œ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๋งŒ๋“ค ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค.

3-4) example 4

notion image
X-ray ์‚ฌ์ง„์„ ์ด์šฉํ•ด์„œ COVID-19๋ฅผ ์ž๋™์œผ๋กœ ์ง„๋‹จํ•˜๋Š” ์‹œ์Šคํ…œ์—์„œ ์˜์‚ฌ๋„ ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ๋„์ถœ๋œ ์˜ˆ์ธก ๊ฒฐ๊ณผ๋งŒ ๋ฌด์ž‘์ • ๋ฏฟ์„ ์ˆ˜๋Š” ์—†์„ ๊ฒƒ์ด๊ณ , ์™œ ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ๊ทธ๋Ÿฐ ์˜ˆ์ธก ๊ฒฐ๊ณผ๋ฅผ ๋„์ถœํ–ˆ๋Š”์ง€์— ๋Œ€ํ•œ ์„ค๋ช…์„ ๋ณด๊ณ  ๊ทธ AI ๋ชจ๋ธ์˜ ์˜ˆ์ธก์„ ์‹ ๋ขฐํ•  ๊ฒƒ์ธ์ง€ ์•„๋‹Œ์ง€๋ฅผ ๊ฒฐ์ •ํ•  ์ˆ˜ ์žˆ์„ ๊ฒƒ์ด๋‹ค.
ย 

3. What is Explainability(์„ค๋ช…์„ ํ•œ๋‹ค๋Š” ๊ฒƒ) / Interpretability?

[1] Miller, Tim. "Explanation in artificial intelligence: Insights from the social sciences." arXiv Preprint arXiv:1706.07269. (2017)
[2] Kim, Been, Rajiv Khanna, and Oluwasanmi O. Koyejo. "Examples are not enough, learn to criticize! Criticism for interpretability." Advances in Neural Information Processing Systems (2016)
[1] Miller, Tim. "Explanation in artificial intelligence: Insights from the social sciences." arXiv Preprint arXiv:1706.07269. (2017) [2] Kim, Been, Rajiv Khanna, and Oluwasanmi O. Koyejo. "Examples are not enough, learn to criticize! Criticism for interpretability." Advances in Neural Information Processing Systems (2016)
"์„ค๋ช…์„ ํ•œ๋‹ค๋Š” ๊ฒƒ(Explainability)"์ด ๋ฌด์—‡์„ ๋œปํ•˜๋Š” ๊ฑด์ง€ ํ•œ๋ฒˆ ์ƒ๊ฐํ•ด ๋ณผ ์ˆ˜ ์žˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜, ์‚ฌ์ „์ด๋‚˜ ๋…ผ๋ฌธ๋“ค์„ ์ฐพ์•„๋ณด๋ฉด ์„ค๋ช…๊ฐ€๋Šฅ์„ฑ(explanation)์ด ์ •ํ™•ํ•˜๊ฒŒ ๋ฌด์—‡์„ ๋œปํ•˜๋Š”์ง€ ์ •์˜๋˜์–ด ์žˆ์ง€ ์•Š๊ณ , ์—ฌ๋Ÿฌ ๊ฐ€์ง€๋ฅผ ํ•จ๊ป˜ ๋œปํ•˜๊ณ  ์žˆ๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ๋‹ค.
  1. ์‚ฌ๋žŒ์ด ๊ทธ ์ด์œ ๋ฅผ ์ดํ•ดํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ด์ฃผ๋Š” ๊ฒƒ
  1. ์„ค๋ช…์„ ํ†ตํ•ด ๋ชจ๋ธ์˜ ๊ฒฐ๊ณผ๋ฅผ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ด ์ฃผ๋Š” ๊ฒƒ
  1. ์ด์œ ๋ฅผ ์„ค๋ช…ํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ด ์ฃผ๋Š” ๊ฒƒ
์ด๋ผ๊ณ  ๋…ผ๋ฌธ์—์„œ ์ •์˜ํ•œ๋‹ค.
ํ•˜์ง€๋งŒ, ์ •ํ™•ํžˆ ์–ด๋–ป๊ฒŒ ํ•˜๋Š” ๊ฒƒ์ด ์„ค๋ช… ๊ฐ€๋Šฅ์„ฑ์„ ๋†’์ด๋Š” ๊ฒƒ์ธ์ง€์— ๋Œ€ํ•œ ๊ตฌ์ฒด์ ์ธ ๋‹ต์„ ์•Œ๊ธฐ๋Š” ์–ด๋ ต๋‹ค.
Arrieta, Alejandro Barredo, et al. "Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI." Information Fusion 58 (2020): 82-115.
Arrieta, Alejandro Barredo, et al. "Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI." Information Fusion 58 (2020): 82-115.
XAI์˜ ์—ฐ๊ตฌ ๋ถ„์•ผ์—์„œ๋„ XAI๋Š” ์‚ฌ๋žŒ์ด ๋ชจ๋ธ์„ ์“ธ ๋•Œ ๊ทธ ๋™์ž‘์„ ์ดํ•ดํ•˜๊ณ  ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ด์ฃผ๋Š” ๊ธฐ๊ณ„ ํ•™์Šต ๊ธฐ์ˆ ์ด๋ผ๊ณ ๋งŒ ๋˜์–ด ์žˆ๋‹ค.

4. Taxonomy of XAI Methods : ๋ถ„๋ฅ˜ ๋ฐฉ๋ฒ•

์˜ˆ์ธก ๊ฒฐ๊ณผ์— ์ค‘์š”ํ•œ ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ํŠน์ง•๋“ค์„ ์ฐพ์•„๋‚ด์–ด ๊ทธ ์ค‘์š”๋„๋ฅผ ์‹œ๊ฐํ™” ๋“ฑ์„ ํ†ตํ•ด์„œ ์„ค๋ช…ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๊ณ ์•ˆํ•˜๋Š” ์—ฐ๊ตฌ๋“ค์ด ์ตœ๊ทผ์— ๋งŽ์ด ์ œ์‹œ๋˜์—ˆ๋Š”๋ฐ, ๊ทธ๋Ÿฐ ๋ฐฉ๋ฒ•๋“ค์ด ์–ด๋–ป๊ฒŒ ๋ถ„๋ฅ˜๋˜๊ณ  ์žˆ๋Š”์ง€ ์ •๋ฆฌํ•˜๊ฒ ๋‹ค.
  • Local vs. Global โ€“ Local: Describes an individual prediction โ€“ Global: Describes entire model behavior
  • White-box vs. Black-box โ€“ White-box: Explainer can access the inside of model โ€“ Black-box: Explainer can access only the output
  • Intrinsic vs. Post-hoc โ€“ Intrinsic: Restricts the model complexity before training โ€“ Post-hoc: Applies after the ML model is trained
  • Model-specific vs. Model-agnostic โ€“ Model-specific: Some methods restricted to specific model classes (e.g., CAM requires global average pooling) โ€“ Model-agnostic: Some methods can be used for any model
XAI๋ฐฉ๋ฒ•์„ ๊ตฌ๋ถ„ํ•˜๋Š” ์ฒซ๋ฒˆ์งธ ๋ถ„๋ฅ˜๋ฐฉ๋ฒ•์€ localํ•œ ์„ค๋ช… ๋ฐฉ๋ฒ•๊ณผ globalํ•œ ์„ค๋ช… ๋ฐฉ๋ฒ•์œผ๋กœ ๋‚˜๋ˆ ๋ณผ ์ˆ˜ ์žˆ๋‹ค.
  • Local : ์ฃผ์–ด์ง„ ํŠน์ • ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ์˜ˆ์ธก ๊ฒฐ๊ณผ๋ฅผ ๊ฐœ๋ณ„์ ์œผ๋กœ ์„ค๋ช…ํ•˜๋ ค๋Š” ๋ฐฉ๋ฒ•
  • Global : ๊ฐœ๋ณ„ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋”ฐ๋กœ๋”ฐ๋กœ ์„ค๋ช…ํ•˜๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ, ์ „์ฒด ๋ฐ์ดํ„ฐ์…‹์—์„œ ๋ชจ๋ธ์˜ ์ „๋ฐ˜์ ์ธ ํ–‰๋™์„ ์„ค๋ช…ํ•˜๊ณ ์ž ํ•˜๋Š” ๋ฐฉ๋ฒ•
๋‘ ๋ฒˆ์งธ ๋ถ„๋ฅ˜๋Š” White-box ์„ค๋ช… ๋ฐฉ๋ฒ•, Black-box ์„ค๋ช… ๋ฐฉ๋ฒ•์œผ๋กœ ๊ตฌ๋ถ„ํ•˜๋Š” ๊ฒƒ์ด๋‹ค.
  • White-box : ๋ชจ๋ธ์˜ ๋‚ด๋ถ€๊ตฌ์กฐ๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ์•Œ๊ณ  ์žˆ๋Š” ์ƒํ™ฉ์—์„œ ์„ค๋ช…์„ ์‹œ๋„ํ•˜๋Š” ๋ฐฉ๋ฒ•
  • Black-box : ๋ชจ๋ธ์˜ ๋‚ด๋ถ€ ๊ตฌ์กฐ๋Š” ์ „ํ˜€ ๋ชจ๋ฅด๋Š” ์ƒํƒœ์—์„œ ๋‹จ์ˆœํžˆ ๋ชจ๋ธ์˜ ์ž…๋ ฅ๊ณผ ์ถœ๋ ฅ๋งŒ ๊ฐ€์ง€๊ณ  ์„ค๋ช…์„ ์‹œ๋„ํ•˜๋Š” ๋ฐฉ๋ฒ•
์„ธ ๋ฒˆ์งธ ๋ถ„๋ฅ˜๋Š” intrinsic(๋‚ด์žฌ์ ์ธ ์„ค๋ช… ๋ฐฉ๋ฒ•)๊ณผ Post(์‚ฌํ›„ ์ถ”๋ก )์„ ํ†ตํ•œ ์„ค๋ช… ๋ฐฉ๋ฒ•์ด๋‹ค.
  • Intrinsic : ๋ชจ๋ธ์˜ ๋ณต์žก๋„๋ฅผ ํ›ˆ๋ จํ•˜๊ธฐ ์ด์ „๋ถ€ํ„ฐ ์„ค๋ช…ํ•˜๊ธฐ ์šฉ์ดํ•˜๋„๋ก ์ œ์•ˆํ•œ ๋’ค, ํ•™์Šต์„ ์‹œ์ผœ์„œ ๊ทธ ํ›„ ํ•™์Šต๋œ ๋ชจ๋ธ์„ ๊ฐ€์ง€๊ณ  ์„ค๋ช…ํ•˜๋Š” ๋ฐฉ๋ฒ•
  • Post-hoc : ์ž„์˜์˜ ๋ชจ๋ธ์˜ ํ›ˆ๋ จ์ด ๋๋‚œ ๋’ค์— ์ด ๋ฐฉ๋ฒ•์„ ์ ์šฉํ•ด์„œ ๊ทธ ๋ชจ๋ธ์˜ ํ–‰๋™์„ ์„ค๋ช…ํ•˜๋Š” ๋ฐฉ๋ฒ•.

5. Taxonomy of XAI Methods : Examples

์˜ˆ์‹œ 1) ๋จธ์‹ ๋Ÿฌ๋‹ : Linear model, Decision Tree

notion image
๋จธ์‹ ๋Ÿฌ๋‹ ๋ชจ๋ธ์—์„œ linear model์ด๋‚˜ decision tree์— ๊ธฐ๋ฐ˜ํ•œ ๋ชจ๋ธ๋“ค์˜ ๋ถ„๋ฅ˜๋Š” Global, White-box, intrinsic, model-specific ํ•œ ์„ค๋ช… ๋ฐฉ๋ฒ•์ด๋‹ค.
linear ๋ชจ๋ธ์—์„œ๋Š” ํ•™์Šตํ•˜๊ณ  ๋‚œ ๋’ค์— ์–ป์–ด์ง€๋Š” ์„ ํ˜• ๊ณ„์ˆ˜(์ˆ˜์‹์—์„œ๋Š” ๋ฒ ํƒ€)๊ฐ€ ๋ฐ”๋กœ ์ด ๋ชจ๋ธ์˜ ์„ค๋ช…์œผ๋กœ ์ง์ ‘ ์ž‘์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ๋‹ค. ์™œ๋ƒํ•˜๋ฉด ๊ฐ x ๋ณ€์ˆ˜๋“ค์ด ์„ ํ˜•์ ์œผ๋กœ ์กฐํ•ฉ๋˜์–ด ์˜ˆ์ธก๊ฐ’ y๋ฅผ ๋งŒ๋“ค์–ด๋‚ด๋Š”๋ฐ, ์ด ์„ ํ˜•๊ณ„์ˆ˜๊ฐ€ ํด์ˆ˜๋ก ๊ทธ ํŠน์ • ํŠน์ง•(x ๋ณ€์ˆ˜)๋“ค์ด ์˜ˆ์ธก์— ์ž‘์šฉํ•˜๋Š” ์ •๋„๊ฐ€ ํฌ๋‹ค๊ณ  ํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค.
๋˜ํ•œ, ์ด๋Ÿฌํ•œ ์„ค๋ช…์€ ์ฃผ์–ด์ง„ ํŠน์ • ๋ฐ์ดํ„ฐ ์„ค๋ช…์„ ํ•ด์ฃผ๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ์ „์ฒด ๋ฐ์ดํ„ฐ๋ถ€ํ„ฐ ํ•™์Šต๋œ ๋ชจ๋ธ์„ ์„ค๋ช…ํ•˜๋Š” ๊ฒƒ์ด๋ฏ€๋กœ Global ํ•œ ์„ค๋ช…์ด๋‹ค.
๋˜, ๋ชจ๋ธ์˜ ์ •ํ™•ํ•œ ๊ตฌ์กฐ๋ฅผ ์•Œ์•„์•ผ ํ•˜๋ฏ€๋กœ White-box ์„ค๋ช… ๋ฐฉ์‹์ด๋‹ค.
ํ•™์Šต๋˜๋Š” ๋ชจ๋ธ ์ž์ฒด๊ฐ€ ๊ฐ„๋‹จํ•˜๊ณ  ์ง์ ‘์ ์œผ๋กœ ์„ค๋ช…์„ ๋งŒ๋“ค์–ด๋‚ด๋Š” ๋ชจ๋ธ์ด๋ผ์„œ Intrinsic์ด๋‹ค.
์ด ๊ฒฝ์šฐ๋Š” ์„ ํ˜• ๋ชจ๋ธ์—๋งŒ ์ ์šฉ๋˜๋Š” ์„ค๋ช…์ด๊ธฐ ๋•Œ๋ฌธ์— model-specific ํ•œ ์„ค๋ช… ๋ฐฉ๋ฒ•์ด๋‹ค.
์ด์™€ ๋น„์Šทํ•œ ๋ถ„๋ฅ˜๊ธฐ์ธ desicion tree ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์—๋„ ์ ์šฉ๋œ๋‹ค. ํ•™์Šต ๊ณผ์ •์—์„œ Mean Squre Error๋ฅผ ๋งŽ์ด ๊ฐ์†Œ์‹œํ‚ค๋Š” ๊ฒƒ์„ ์ค‘์š”ํ•œ ํŠน์ง•์ด๋ผ๋Š” ์„ค๋ช…์„ ์ œ๊ณตํ•˜๊ฒŒ ๋˜๋ฏ€๋กœ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ global, white-box, intrinsic, model-specific ํ•œ ์„ค๋ช… ๋ฐฉ๋ฒ•์ด๋ผ๊ณ  ํ•  ์ˆ˜ ์žˆ๋‹ค.

์˜ˆ์‹œ 2) ๋”ฅ๋Ÿฌ๋‹ : Grad-CAM

notion image
์ฃผ๋กœ ๋”ฅ๋Ÿฌ๋‹ ๊ธฐ๋ฐ˜์˜ย ๋ถ„๋ฅ˜๋ชจ๋ธ์„ ์„ค๋ช…ํ•˜๋Š”๋ฐย ๋งŽ์ด ์“ฐ์ด๋Š” Grad-CAM ๋ฐฉ๋ฒ•์€ local, white-box, pos-hoc, model-agnostic ์„ค๋ช… ๋ฐฉ๋ฒ•์ด๋ผ๊ณ  ๋ถ„๋ฅ˜ํ•  ์ˆ˜ ์žˆ๋‹ค.
์ฆ‰, ๊ทธ๋ฆผ์— ๋ณด๋Š” ๊ฒƒ์ฒ˜๋Ÿผ ์ฃผ์–ด์ง„ ์–ด๋–ค ๊ฐœ๋ณ„ ์ด๋ฏธ์ง€๋งˆ๋‹ค ๊ทธ ์˜ˆ์ธก ๊ฒฐ๊ณผ๋ฅผ ์„ค๋ช…ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์ด๊ธฐ ๋•Œ๋ฌธ์—ย local ํ•œย ์„ค๋ช…์ด๊ณ  ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ๋ชจ๋ธ์˜ ์ •ํ™•ํ•œ ๊ตฌ์กฐ์™€ ๊ณ„์ˆ˜๋“ค์„ ๋ชจ๋‘ ์•Œ์•„์•ผ ๊ตฌํ•  ์ˆ˜ ์žˆ๋Š” ์„ค๋ช…์ด๊ธฐ ๋•Œ๋ฌธ์— White-box ์„ค๋ช… ๋ฐฉ๋ฒ•์ด๋ฉฐ,ย ๋ชจ๋ธ์ด ํ•™์Šต๋˜๊ณ ย ๋‚œ ํ›„ ์ ์šฉํ•ด์„œ ์„ค๋ช…์„ ์ œ๊ณตํ•˜๋Š” Post-hoc์„ค๋ช…ย ๋ฐฉ๋ฒ•์ด๊ณ ,ย ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธย ๊ตฌ์กฐ์™€ ์ƒ๊ด€์—†์ด ํ•ญ์ƒ ์ ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์ด๊ธฐ ๋•Œ๋ฌธ์— ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์— ๋Œ€ํ•ด์„œ๋Š” model-agnostic ํ•œ ์„ค๋ช… ๋ฐฉ๋ฒ•์ด๋‹ค.
ย 
ย