Inductive Bias

Creator
Creator
Seonglae ChoSeonglae Cho
Created
Created
2023 May 9 2:9
Editor
Edited
Edited
2024 Oct 17 11:40
Refs
Refs

Assumption, how model will interpret and generalize data before learning

It is an additional assumption for contingencies to resolve errors caused by incorrect assumptions of the learning algorithm for data that has not been given. In other words, it is a bias set as a tendency for certain inference paths to be prioritized or more easily learned due to the characteristics of the data or the design of the model.
Each design choice imposes an inductive bias on the learning process. Informally, inductive biases impose a preference over ๐“•. When the inductive bias prefers a subset of ๐“• which contains a function ๐‘“ that obtains minimal generalization error, the inductive bias is useful and improves learning.
ย 

Learning Bias

model intrinsic but encodes and amplifies biases
์˜๋„๋˜์ง€ ์•Š์€ ํŽธํ–ฅ์œผ๋กœ, ํ•™์Šต ๋ฐ์ดํ„ฐ์— ์žˆ๋Š” ๋ถˆํ•„์š”ํ•˜๊ฑฐ๋‚˜ ๋ถ€์ •์ ์ธ ํŽธํ–ฅ์ด ํ•™์Šต ๊ณผ์ •์—์„œ ์ฆํญ๋˜์–ด ๋ชจ๋ธ์— ๋ฐ˜์˜๋˜๋Š” ๊ฒƒ
ย 
ย 
ย 
Inductive Bias๋ž€ ๋ฌด์—‡์ผ๊นŒ?
๋จธ์‹ ๋Ÿฌ๋‹์—์„œ Bias๋Š” ๋ฌด์Šจ ์˜๋ฏธ์ผ๊นŒ? Inductive Bias๋ผ๋Š” ์šฉ์–ด์—์„œ, Bias๋ผ๋Š” ์šฉ์–ด๋Š” ๋ฌด์—‡์„ ์˜๋ฏธํ• ๊นŒ? ๋”ฅ๋Ÿฌ๋‹์„ ๊ณต๋ถ€ํ•˜๋‹ค ๋ณด๋ฉด, Bias๊ณผ Variance๋ฅผ ํ•œ ๋ฒˆ์ฏค์€ ๋“ค์–ด๋ดค์„ ๊ฒƒ์ด๋‹ค. Bias๋Š” ํƒ€๊ฒŸ๊ณผ ์˜ˆ์ธก๊ฐ’์ด ์–ผ๋งˆ๋‚˜ ๋ฉ€๋ฆฌ ๋–จ์–ด์ ธ ์žˆ๋Š”๊ฐ€, ๊ทธ๋ฆฌ๊ณ  Variance๋Š” ์˜ˆ์ธก๊ฐ’๋“ค์ด ์–ผ๋งˆ๋‚˜ ํผ์ ธ์žˆ๋Š”๊ฐ€๋ฅผ ์˜๋ฏธํ•œ๋‹ค. ๋”ฐ๋ผ์„œ Bias๊ฐ€ ๋†’์€ ๊ฒฝ์šฐ ๋ฐ์ดํ„ฐ๋กœ๋ถ€ํ„ฐ ํƒ€๊ฒŸ๊ณผ์˜ ์—ฐ๊ด€์„ฑ์„ ์ž˜ ์ฐพ์•„๋‚ด์ง€ ๋ชปํ•˜๋Š” ๊ณผ์†Œ์ ํ•ฉ(Underfitting) ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•˜๊ณ , Variance๊ฐ€ ๋†’์€ ๊ฒฝ์šฐ์—๋Š” ๋ฐ์ดํ„ฐ์˜ ์‚ฌ์†Œํ•œ ๋…ธ์ด์ฆˆ๋‚˜ ๋žœ๋คํ•œ ๋ถ€๋ถ„๊นŒ์ง€ ๋ฏผ๊ฐํ•˜๊ฒŒ ๊ณ ๋ คํ•˜๋Š” ๊ณผ์ ํ•ฉ(Overfitting) ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•œ๋‹ค. ์—ฌ๊ธฐ์„œ, Bias๋ฅผ ํ•™์Šต ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ์ž˜๋ชป๋œ ๊ฐ€์ •์— ์˜ํ•˜์—ฌ ๋ฐœ์ƒํ•˜๋Š” ์˜ค์ฐจ๋ผ๊ณ ๋„ ํ•˜๋Š”๋ฐ, ์ž˜๋ชป๋œ ๊ฐ€์ •์— ์˜ํ•˜์—ฌ ๋ฐ์ดํ„ฐ์—์„œ ์ค‘..
Inductive Bias๋ž€ ๋ฌด์—‡์ผ๊นŒ?
ย 
ย 

Recommendations