Computer Science/Data Science

๋จธ์‹ ๋Ÿฌ๋‹/AI์—์„œ ์‚ฌ์šฉ๋˜๋Š” "Ground Truth" ๋œป

_cactus 2021. 3. 5. 11:19
๋ฐ˜์‘ํ˜•

๋จธ์‹ ๋Ÿฌ๋‹์— ๊ด€๋ จํ•œ ๊ธ€์„ ์ฝ๋‹ค๋ณด๋ฉด "ground-truth"๋ผ๋Š” ์šฉ์–ด๋ฅผ ๋งŽ์ด ์ ‘ํ•˜๊ฒŒ ๋œ๋‹ค.
"ground-truth"๋Š” ๊ธฐ์ƒํ•™์—์„œ ์œ ๋ž˜ํ•˜์˜€์œผ๋ฉฐ, ์–ด๋Š ํ•œ ์žฅ์†Œ์—์„œ ์ˆ˜์ง‘๋œ ์ •๋ณด๋ฅผ ์˜๋ฏธํ•˜๋Š” ์šฉ์–ด๋กœ ์‚ฌ์šฉ๋˜์—ˆ๋‹ค๊ณ  ํ•œ๋‹ค.

 

๊ทธ๋ ‡๋‹ค๋ฉด machine learning/AI์˜ ๋ฌธ๋งฅ์—์„œ ์‚ฌ์šฉ๋˜๋Š” "ground-truth"์˜ ๋œป์€ ๋ฌด์—‡์ธ๊ฐ€.

๋ฐ˜์‘ํ˜•

What is ground truth?

Ground truth isn't true. It's an ideal expected result. It might involve hand-labeling example datapoints to collect desirable answer data for training your system.
For example, a set of images might be painstakingly hand-labeled as cat or not-cat according to the opinions of whoever was in charge of the project and those cat/not-cat labels will be called "ground truth" for the project.

cited : towardsdatascience.com/in-ai-the-objective-is-subjective-4614795d179b

์ฒ˜์Œ์— ๋‚˜๋Š” "ground-truth"์ด "label"๊ณผ ๊ฐ™์€ ์˜๋ฏธ์ธ ๊ฒƒ ๊ฐ™์€๋ฐ ์™œ "label"์ด๋ผ๊ณ  ์•ˆํ•˜๊ณ  "ground-truth"๋ผ๊ณ  ํ•˜์ง€? ์ƒ๊ฐํ–ˆ์—ˆ๋‹ค.ใ…Ž
๊ฒฐ๋ก ์ ์œผ๋กœ, "ground-truth"๊ณผ "label"์€ ๊ฐ™์€ ์˜๋ฏธ๋กœ ์“ฐ์ด๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋‹ค. ๋ช…ํ™•ํ•˜๊ฒŒ ๋งํ•˜๋ฉด ๋‘˜์€ ๊ฐ™์ง€ ์•Š๋‹ค.

 

"label"์€ ์ •๋‹ต์ง€๋ผ๊ณ ๋„ ๋ถˆ๋ฆฌ๋“ฏ์ด ๋‹ต์ด ๋ช…ํ™•ํ•˜๊ฒŒ ์ •ํ•ด์ ธ ์žˆ๋Š” ๊ฐ’์ด๋‹ค.

 

"ground-truth"์€ '์šฐ๋ฆฌ๊ฐ€ ์ •ํ•œ ์ •๋‹ต', '์šฐ๋ฆฌ์˜ ๋ชจ๋ธ์ด ์šฐ๋ฆฌ๊ฐ€ ์›ํ•˜๋Š” ๋‹ต์œผ๋กœ ์˜ˆ์ธกํ•ด์ฃผ๊ธธ ๋ฐ”๋ผ๋Š” ๋‹ต'์ด๋‹ค.
์˜ˆ๋ฅผ ๋“ค์–ด, ์•„๋ž˜ ์‚ฌ์ง„์ด ๋ชจ๋ธ์„ ํ›ˆ๋ จ์‹œํ‚ฌ ๋ฐ์ดํ„ฐ๋กœ ์‚ฌ์šฉ๋œ๋‹ค๊ณ  ํ•ด๋ณด์ž. ์‚ฌ์ง„์€ ๊ณ ์–‘์ด ๋ถ„์žฅ์„ ํ•œ ์‚ฌ๋žŒ์ด๋‹ค. ์ด ์‚ฌ์ง„์— ๋Œ€ํ•œ label, ์ฆ‰ ์ •๋‹ต์€ ์—†๋‹ค. ์‚ฌ๋žŒ์ธ ๊ฒƒ๋„, ๊ณ ์–‘์ธ ๊ฒƒ๋„ ์•„๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์šฐ๋ฆฌ๋Š” ์šฐ๋ฆฌ์˜ ๋ถ„๋ฅ˜๋ชจ๋ธ์ด ์ด๋ฅผ '๊ณ ์–‘์ด'๋กœ ๋ถ„๋ฅ˜ํ•˜๊ธธ ์›ํ•œ๋‹ค. ๊ทธ๋ฆฌํ•˜์—ฌ ์ด ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ "ground-truth" ๊ฐ’์€ ์šฐ๋ฆฌ๊ฐ€ ์›ํ•˜๋Š” ๋‹ต์ธ '๊ณ ์–‘์ด'๊ฐ€ ๋œ๋‹ค.

์ถœ์ฒ˜ : Movie 'Cats'

728x90
๋ฐ˜์‘ํ˜•