Entropy, Cross-Entropy, KL Divergence

์ •๋ณด๋Ÿ‰

์ •๋ณด์ด๋ก ์—์„œ๋Š” ์ž์ฃผ ์ผ์–ด๋‚˜์ง€ ์•Š๋Š” ์‚ฌ๊ฑด์˜ ์ •๋ณด๋Ÿ‰์€ ์ž์ฃผ ๋ฐœ์ƒํ•˜๋Š” ์‚ฌ๊ฑด๋ณด๋‹ค ์ •๋ณด๋Ÿ‰์ด ๋งŽ๋‹ค๊ณ  ๊ฐ„์ฃผํ•จ

์ •๋ณด๋Ÿ‰์„ ํ™•๋ฅ ์— ๋Œ€ํ•œ ํ•จ์ˆ˜ (0~1) ๋กœ ์ •์˜ํ•œ๋‹ค๋ฉด

์‚ฌ๊ฑดA์ด ์ผ์–ด๋‚  ํ™•๋ฅ  P(A)๋กœ ์‚ฌ๊ฑด A์˜ ์ •๋ณด๋Ÿ‰ h(A)์„ ์ •์˜ํ•˜๋ฉด

h(A):=โˆ’logP(A)

์ •๋ณด๋Ÿ‰ ๋กœ๊ทธ ํ•จ์ˆ˜ ๊ทธ๋ž˜ํ”„: -log(x) X์ถ•: (0.0, 0.2, 0.4, 0.6, 0.8, 1.0) / y์ถ•: (0, 2, 4, 6, 8 ,10 ,12)

Example

  • P(A)=0.99 --> ์ •๋ณด๋Ÿ‰์€ h(A)=โˆ’logP(A)=โˆ’log0.99=0.01

  • P(B)=0.01 --> ์ •๋ณด๋Ÿ‰์€ h(B)=โˆ’logP(B)=โˆ’log0.01= 4.61

์—”ํŠธ๋กœํ”ผ(Entropy)

์ด์‚ฐํ™•๋ฅ ๋ณ€์ˆ˜(discrete random variable)์˜ ํ‰๊ท  ์ •๋ณด๋Ÿ‰, ๋ถˆํ™•์‹ค์„ฑ ์ •๋„๋ฅผ ๋‚˜ํƒ€๋ƒ„

์ด์‚ฐํ™•๋ฅ ๋ณ€์ˆ˜ X์˜ ํ‰๊ท  ์ •๋ณด๋Ÿ‰ H[X]๋Š”

H[X]=โˆ‘pilogpi,i=1ย toย NH[X]= โˆ‘p_i log p_i, i=1 ~to~ N

Example

  • P(X=0)=0.5, P(X=1)=0.5 H[X]=โˆ’(0.5log0.5+0.5log0.5)=0.69 (<-- max. entropty)

  • P(X=0)=0.8, P(X=1)=0.2, H[X]=โˆ’(0.8log0.8+0.2log0.2)=0.50

  • P(X=0)=1, P(X=1)=0, H[X]=โˆ’(1log1+0log0)=0.

KL Divergence

๋‘ ํ™•๋ฅ ๋ถ„ํฌ์˜ ๋‹ค๋ฅธ ์ •๋„๋ฅผ ์ธก์ •. Relative entropy ๋ผ๊ณ ๋„ ํ•˜๋ฉฐ ์ •์‹ ๋ช…์นญ์€ Kullbackโ€“Leibler divergence์ด๋‹ค.

KL(pโˆฃq):=โˆ’โˆ‘pilogqiโˆ’(โˆ’โˆ‘pilogpi)=โˆ’โˆ‘pilog(qipi)K L ( p | q ) := โˆ’ โˆ‘ p_ i log q_ i โˆ’ ( โˆ’ โˆ‘ p_ i log p_ i )=โˆ’ โˆ‘ p_ i log ( q_ i p _i )

Cross-entropy

์ฃผ์–ด์ง„ ํ™•๋ฅ ๋ณ€์ˆ˜ X ์— ๋Œ€ํ•ด์„œ ํ™•๋ฅ ๋ถ„ํฌ p ๋ฅผ ์ฐพ๋Š” ๋ฌธ์ œ์—์„œ ํ™•๋ฅ ๋ถ„ํฌ p ์˜ ์ •ํ™•ํ•œ ํ˜•ํƒœ๋ฅผ ๋ชจ๋ฅด๊ธฐ ๋•Œ๋ฌธ์— p ๋ฅผ ์˜ˆ์ธกํ•œ ๊ทผ์‚ฌ ๋ถ„ํฌ q ๋ฅผ ์ƒ๊ฐํ•  ๊ฒƒ์ด๋‹ค.

์ •ํ™•ํ•œ ํ™•๋ฅ ๋ถ„ํฌ๋ฅผ ์–ป๊ธฐ ์œ„ํ•ด q ์˜ parameter๋“ค์„ updateํ•˜๋ฉด์„œ q ๋ฅผ p ์— ๊ทผ์‚ฌํ•  ๊ฒƒ์ด๋‹ค

์ฆ‰. ๋‘ ๋ถ„ํฌ์˜ ์ฐจ์ด๋ฅผ ์ธก์ •ํ•˜๋Š” KL(p|q)๊ฐ€ ์ตœ์†Œ๊ฐ€ ๋˜๋Š” q๋ฅผ ์ฐพ๋Š” ๋ฌธ์ œ๊ฐ€ ๋จ

KL(p|q) ์˜ ๋‘ ๋ฒˆ์งธํ•ญ ($โˆ’โˆ‘p_i log p_i$) ๋Š” ๊ทผ์‚ฌ๋ถ„ํฌ q์— ๋ฌด๊ด€ํ•œ ํ•ญ์ด๋ฏ€๋กœ

KL Divergence๋ฅผ ์ตœ์†Œํ™” ํ•˜๋Š” ๊ฒƒ์€ ๊ฒฐ๊ตญ ์ฒซ ๋ฒˆ์งธ ํ•ญ cross-entropy๋ฅผ ์ตœ์†Œํ™”ํ•˜๋Š” q๋ฅผ ์ฐพ์•„์•ผ ํ•œ๋‹ค.

KL(pโˆฃq):=โˆ’โˆ‘pilogqiK L ( p | q ) := โˆ’ โˆ‘ p_ i log q_ i

์—ฌ๊ธฐ์„œ p =( p i ) ๋Š” ์‹ค์ œ ํ™•๋ฅ ๋ถ„ํฌ๋ฅผ ์˜๋ฏธํ•˜๊ณ  q=( q i ) ๋Š” p ๋ฅผ ๊ทผ์‚ฌํ•œ ๋ถ„ํฌ๋‹ค.

Reference

Last updated

Was this helpful?