Live Session 0622

less than 1 minute read

Underprediction

\(t_{i}\): target label(i-th sample)
\(y_{i}\): modelโ€™s prediction

underprediction: \(t_i > y_i\)
overprediction: \(t_i < y_i\)

loss function์„ ๋งŒ๋“ค๋•Œ,
\(t_i > y_i \rightarrow (t_i - y_i)^2\)
\(t_i < y_i \rightarrow \frac{1}{2}(t_i - y_i)^2\)

max๋ฅผ ํ™œ์šฉํ•ด์„œ ๊ฐ€๋Šฅํ•˜๋‹ค.
t_i > y_i ์ธ ๊ฒฝ์šฐ 0๋ณด๋‹ค ํฌ๊ฒŒ ๋˜๊ณ  ๋ฐ˜๋Œ€๋Š” 0์ด ๋œ๋‹ค.

\[\frac{1}{N}\sum_{i=1}^{N}\left( max\{0, t_i-y_i\}^2 + \frac{1}{2}max\{0, y_i - t_i\}^2\right)\]

hinge loss

overprediction: \(\frac{1}{2}max\{0, y_i - t_i\}^2\)์€ 0.5์˜ ๊ฐ€์ค‘์น˜๊ฐ€ ์ฃผ์–ด์กŒ๋‹ค.

0.5์˜ ๊ฐ’์€ hyperparameter๋กœ ๋ณด๊ณ  ์ฐพ์•„์•ผ ํ•˜์ง€๋งŒ ๋งŽ์€ ๊ฒฝ์šฐ์—๋Š”โ€ฆ

0.5๋กœ ์„ค์ •ํ•œ ์ด์œ ๋Š” ์—…๊ณ„์—์„œ underprediction์ด 2๋ฐฐ๋กœ ์ค‘์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์— practical ๊ด€์ ์—์„œ ์„ ํƒํ•œ ๊ฒƒ์ด๋‹ค.

Leave a comment