ํ‹ฐ์Šคํ† ๋ฆฌ ๋ทฐ

https://stats.stackexchange.com/questions/365444/why-is-a-0-1-loss-function-intractable

 

Why is a 0-1 loss function intractable?

In Ian Goodfellow's Deep Learning book, it is written that Sometimes, the loss function we actually care about (say, classification error) is not one that can be optimized efficiently. For exam...

stats.stackexchange.com

 

์ด ๊ธ€์„ ์ฝ๊ธฐ ์ „, ๋‹น์‹ ์ด ๋ฐฑ์ง€์— zero-one loss function์„ ์ ์„ ์ˆ˜ ์žˆ๋Š”์ง€ ์ ๊ฒ€ํ•˜๋ผ!

 

 

 

๊ฒฐ๋ก ์„ ๋งํ•˜์ž๋ฉด, zero-one loss function์ด (1) discontinous & (2) non-convex ํ•จ์ˆ˜์ด๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค!

 

zero-one loss๋ฅผ ์“ฐ๊ฒŒ ๋  ๊ฒฝ์šฐ, ์šฐ๋ฆฌ๋Š” ์ด๊ฒƒ์„ ์ตœ์†Œํ™”ํ•˜๊ธฐ ์œ„ํ•ด exponentialํ•œ ๊ฐ€๋Šฅํ•œ ๊ฒฐ๊ณผ๋“ค์„ ์‚ดํŽด๋ด์•ผ ํ•œ๋‹ค!! ์™œ๋ƒ?

๊ฐ๊ฐ์˜ (x, y) input pair์— ๋Œ€ํ•ด์„œ, ์šฐ๋ฆฌ์˜ zero-one loss๋Š” 0 ์•„๋‹˜ 1์˜ ๊ฐ’์„ ๋ฆฌํ„ดํ•  ๊ฒƒ์ด๋‹ค.

๊ทธ๋Ÿฌ๋ฉด, n๊ฐœ์˜ input์— ๋Œ€ํ•ด zero-one loss๋Š” $2^n$ ๋งŒํผ์˜ ๊ฐ€๋Šฅํ•œ ์กฐํ•ฉ์„ ๋‚ด๋ฑ‰์„ ๊ฒƒ์ด๋‹ค.

 

์—ฌ๊ธฐ์— ํ•œ์ˆ  ๋” ๋– ์„œ, zero-one loss์˜ ๊ฐ’์€ ๋‹น์‹ ์ด weight ๊ฐ’์„ ์–ด๋–ป๊ฒŒ ์กฐ์ •ํ•ด์•ผ loss๊ฐ€ 1์ด ์•„๋‹ˆ๋ผ 0์œผ๋กœ ์ค„์–ด๋“ค์ง€์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ์ฃผ์ง€ ๋ชปํ•œ๋‹ค.

 

๊ทธ๋ž˜์„œ

1. zero-one loss๋ฅผ ๋ด๋„ weight๋ฅผ ์–ด๋–ป๊ฒŒ ์ˆ˜์ •ํ•ด์•ผ ํ• ์ง€ ๋ชจ๋ฅธ๋‹ค.

2. weight๋ฅผ ์–ด๋–ป๊ฒŒ ์ˆ˜์ •ํ•ด์•ผ ํ• ์ง€ ๋ชจ๋ฅด๋‹ˆ ๊ฐ€๋Šฅํ•œ ์กฐํ•ฉ ์ „๋ถ€๋ฅผ ๋ด์•ผ ํ•œ๋‹ค.

3. ๊ฒฐ๊ตญ n๊ฐœ input์— ๋Œ€ํ•ด $2^n$๊ฐœ์˜ ๊ฐ€๋Šฅํ•œ ์กฐํ•ฉ์„ ๋ด์•ผ ํ•œ๋‹ค.

4. ๋‡Œ-์ ˆ

 

 

๋Œ“๊ธ€
๊ณต์ง€์‚ฌํ•ญ
์ตœ๊ทผ์— ์˜ฌ๋ผ์˜จ ๊ธ€
์ตœ๊ทผ์— ๋‹ฌ๋ฆฐ ๋Œ“๊ธ€
Total
Today
Yesterday
๋งํฌ
ยซ   2024/11   ยป
์ผ ์›” ํ™” ์ˆ˜ ๋ชฉ ๊ธˆ ํ† 
1 2
3 4 5 6 7 8 9
10 11 12 13 14 15 16
17 18 19 20 21 22 23
24 25 26 27 28 29 30
๊ธ€ ๋ณด๊ด€ํ•จ