weight initialization 1

[EECS 498-007] Lecture 10. Training Neural Networks I

์–ด๊น€์—†์ด ์ฐพ์•„์˜ค๋Š” ~ ๊ทธ๋Ÿฐ๋ฐ ๋ฒŒ์จ ๊ธˆ์š”์ผ์ด์•ผ ๊ฑฐ์ง“๋ง ๊ฐ™์•  ์‹œ๊ฐ„์ด ์•„์ฃผ ๋งค์„ญ๊ฒŒ ํ›…ํ›… ์ง€๋‚˜๊ฐ„๋‹ค ๋‚ ์”จ๋„ ์‚ด์ง์ฟต ํ’€๋ฆฐ ๊ฒƒ ๊ฐ™์ง€? ์ดˆ์ฝ”์†ก์ด ๋จน๊ตฌ์‹ญ๋‹ค ์˜ค๋Š˜์€ ์‹ ๊ฒฝ๋ง์„ ์‹ค์ œ๋กœ ํ›ˆ๋ จํ•˜๋Š” ๋ฐฉ๋ฒ•! ์— ๋Œ€ํ•ด ์•Œ์•„๋ณผํ…Œ๋‹ค. ์„น์…˜์„ ๋‘˜๋กœ ๋‚˜๋ˆ„์–ด ์ง„ํ–‰ํ• ํ…๋ฐ, ์ด๋ฒˆ ํฌ์ŠคํŒ…์—์„œ๋„ ์ฒซ๋ฒˆ์งธ์— ํ•ด๋‹นํ•˜๋Š” One time setup์— ๊ด€ํ•ด ๋‹ค๋ฃฐ ๊ฒƒ์ด๋‹ค. ๋‹ค์Œ ํฌ์ŠคํŒ…์—์„œ๋Š” 11๊ฐ•์ธ 2, 3์— ๋Œ€ํ•ด์„œ ์จ๋ณผ๊ฒŒ์š”. [Activation Functions] ํ™œ์„ฑํ™” ํ•จ์ˆ˜๋Š” ์ธ๊ณต์‹ ๊ฒฝ๋ง์—์„œ ์ค‘์š”ํ•œ ๊ตฌ์„ฑ ์š”์†Œ๋กœ, ๋ชจ๋ธ์— nonlinearity๋ฅผ ๋ถ€์—ฌํ•˜๋Š” ์—ญํ• ์„ ํ•œ๋‹ค. ๊ธฐ๋ณธ์ ์œผ๋กœ ํ™œ์„ฑํ™”ํ•จ์ˆ˜๊ฐ€ ์ด์ „ layer์˜ ์ž…๋ ฅ๋“ค์˜ ๊ฐ€์ค‘ ํ•ฉ์— ์ž‘์šฉํ•œ ํ›„ ๊ฒฐ๊ณผ๋ฅผ ๋‹ค์Œ layer๋กœ ์ „๋‹ฌํ•ด์ฃผ๋Š”๋ฐ, ์ด๋Ÿฐ ํ™œ์„ฑํ™” ํ•จ์ˆ˜๊ฐ€ ์กด์žฌํ•˜์ง€ ์•Š์œผ๋ฉด ๋„คํŠธ์›Œํฌ๊ฐ€ ๋‹จ์ผ ์„ ํ˜• ๋ ˆ์ด์–ด๊ฐ€ ๋˜์–ด ์ฒ˜๋ฆฌ ๋Šฅ๋ ฅ์ด..

EECS 498-007 2024.01.29