๋ณธ๋ฌธ ๋ฐ”๋กœ๊ฐ€๊ธฐ

๐Ÿš“ Self Study65

Deep learning (Gradient Descent, Chain Rule, Learning Rate) Neural Network Classifier Gradient Descent ๋ฏธ๋ถ„์„ ํ†ตํ•ด์„œ Optimizeํ•˜๊ฒŒ ๋œ๋‹ค. ํ•˜๋‚˜์˜ ์Šค์นผ๋ผ ๊ฐ’์„ W์— ๋Œ€ํ•ด ๋ฏธ๋ถ„์„ ํ•˜๊ฒŒ ๋œ๋‹ค. W์˜ ๊ฐ’์„ ์ผ๋ ฌ๋กœ ์„ธ์›Œ์„œ ์ผ์ฐจ์› ๋ฒกํ„ฐ์ธ ๊ฒƒ์œผ๋กœ ์ƒ๊ฐํ•˜์ž. scalar ๊ฐ’์„ ๋‚˜์—ดํ•˜๋ฉด ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์‹์„ ์–ป์„ ์ˆ˜ ์žˆ๋‹ค. (๊ฐ€์žฅ ์˜ค๋ฅธ์ชฝ ์‹) ์ด๋Ÿฌํ•œ ๋ฒกํ„ฐ๋ฅผ Gradient ๋ฒกํ„ฐ๋ผ๊ณ  ํ•œ๋‹ค. ๊ฐ๊ฐ Element๋“ค์ด ํ•ด๋‹น Weight๋กœ ๋ฏธ๋ถ„์ด ๋œ๊ฒƒ์ด Gradient ๋ฒกํ„ฐ๋ผ๊ณ  ํ•œ๋‹ค. ์ˆ˜ํ•™์ ์œผ๋กœ Error surface์˜ error๊ฐ€ ๊ธ‰๊ฒฉํ•˜๊ฒŒ ๋ณ€ํ™”๋Š” ๋ฐฉํ–ฅ์ด๋ผ๊ณ  ๋ณผ ์ˆ˜ ์žˆ๋‹ค. W์— Gradient๋ฅผ ๋งˆ์ด๋„ˆ์Šค ํ•œ๋‹ค๋Š” ๊ฒƒ์€ ๋ฐ˜๋Œ€ ๋ฐฉํ–ฅ์œผ๋กœ ๊ฐ„๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•œ๋‹ค. ๊ฐ์†Œํ•˜๋Š” ๋ฐฉํ–ฅ์œผ๋กœ ์ด๋™ํ•˜๊ฒŒ ๋œ๋‹ค. ํ˜„์žฌ Weight์—์„œ Gradient๋ฅผ ๋นผ์ฃผ๋Š” ๊ฒƒ์ด Gradie.. 2021. 12. 23.
Deep learning(Activation Function, Forward Propagation, Single Layer Perceptron) Activation Function ์šฐ์„  Non Linearํ•œ ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋‹ค. ํ™•๋ฅ ๊ฐ’์„ ๊ณ ๋ คํ•ด๋ณด๋ฉด -1 ์—์„œ 1๊นŒ์ง€์˜ ๊ฐ’์„ ๊ฐ€์ง„๋‹ค. Measurement๋กœ ๋‚˜์˜จ ๊ฒƒ์„ ํ™•๋ฅ ๊ณผ Decision์œผ๋กœ ํ•˜๊ธฐ ์œ„ํ•ด์„œ ํ•„์š”ํ•˜๋‹ค. Forward Propagation ์ฒซ ๋ฒˆ์งธ Layer์˜ Input Vector๋Š” ๋‹จ์ˆœํžˆ ์ž…๋ ฅ์ •๋ณด๊ฐ€ ๋˜๊ณ  ๋‘ ๋ฒˆ์งธ Layer๋ถ€ํ„ฐ๋Š” ๊ฐ๊ฐ ํ•ด๋‹น Operation ํ•จ์ˆ˜๋ฅผ ๊ฑฐ์นœ ๊ฐ’๋“ค์„ ๊ฐ€์ง„๋‹ค. ๋‹ค์Œ ๋ฒˆ์˜ node๋“ค๋„ ์ด์ „์˜ ๊ฐ’๋“ค๋กœ ๋ถ€ํ„ฐ Operation์ด ์ถ•์ ๋˜์–ด ์—ฌ๋Ÿฌ Layer๋ฅผ ๊ฑฐ์ณ ์ „ํŒŒ๋˜๋Š” ๊ฒƒ์„ Propagation์ด๋ผ๊ณ  ํ•œ๋‹ค. Node์™€ Connection Weight์ด ์กด์žฌํ•œ๋‹ค๋Š” ๊ฒƒ์ด๋‹ค. InDim : Input Layer // outDim : ouput layer connect.. 2021. 12. 23.
Deep learning (Neural Network, Perceptron Neuron) Neural Network ์ƒ๋ฌผํ•™์ ์ธ ๋‡Œ์„ธํฌ๋ฅผ ๋ชจ๋ฐฉํ•ด์„œ ๋งŒ๋“  Computation Model. ์•ž ์ชฝ์—์„œ ์‹ ํ˜ธ๋ฅผ ๋ฐ›์•„์„œ Mergeํ•˜๊ณ  ์ผ์ • ์ˆ˜์น˜ ์ด์ƒ์ด๋ฉด ๋‹ค์Œ์œผ๋กœ ๋„˜๊ธฐ๋Š” ์—ญํ• ๋กœ ์•Œ๋ ค์ ธ ์žˆ๋‹ค. ์‚ฌ๋žŒ์˜ ๋‡Œ๋Š” ๋ณต์žกํ•œ ์ผ์„ ํ•  ์ˆ˜ ์žˆ์ง€๋งŒ ๋‰ด๋Ÿฐ ํ•˜๋‚˜์˜ ๋‹จ์œ„๋Š” ๊ต‰์žฅํžˆ ๋‹จ์ˆœํ•˜๋‹ค๋Š” ๊ฒƒ. ์ด๋Ÿฐ ๊ฐœ๋…์œผ๋กœ ์ปดํ“จํŒ…์„ ํ•˜๊ธฐ ์œ„ํ•œ ๋ชจ๋ธ๋ง์„ ์‹œ์ž‘ํ–ˆ๋‹ค. Connection Weight์„ ๊ณฑํ•ด์„œ ๊ฐ ๋ฐ์ดํ„ฐ๋ฅผ ๋”ํ•ด์„œ Mergeํ•ด์ฃผ๊ณ  Activation Function ์„ ํ†ตํ•ด์„œ ํ•จ์ˆ˜๊ฐ€ ๋งŒ๋“ค์–ด์ง„๋‹ค. ๋ชจ๋“  ๊ฐ’์ด ์‹œ๊ทธ๋งˆ๋กœ ์ธํ•ด ๋”ํ•ด์ง€๊ณ  (์„ธํƒ€)๋ผ๋Š” ๊ฐ’์„ ๋”ํ•˜๊ฒŒ ๋œ๋‹ค. ์ด (์„ธํƒ€)์˜ ๊ฒฝ์šฐ ์ง์„ ์˜ ๋ฐฉ์ •์‹ y = ax + b์—์„œ b๊ฐ€ ์žˆ์–ด์•ผ ๋ชจ๋“  ์ง์„ ์ด ํ‘œํ˜„๋˜๋“ฏ Bias๋ฅผ (์„ธํƒ€)๋กœ ๋”ํ•˜๊ฒŒ ๋˜๋Š” ๊ฒƒ์ด๋‹ค. ๋‰ด๋Ÿฐ ํ•˜๋‚˜๊ฐ€ ํ•˜๋Š” ๊ฒƒ์„ weighted s.. 2021. 12. 23.
Deep learning (Bayesian Theorem) Bayesian Theorem Classification ๋ฌธ์ œ์—์„œ ๋ชธ๋ฌด๊ฒŒ์™€ ํ‚ค๋ฅผ ์ฃผ๊ณ  ๋‚จ์ž์™€ ์—ฌ์ž๋ฅผ ๊ตฌ๋ถ„ํ•˜๋ผ๋Š” ๋ฌธ์ œ๋ฅผ ์ƒ๊ฐํ•ด๋ณด์ž. 100% ๋งž์ถ”์ง€๋งŒ ์–ด๋Š์ •๋„ ๋งž์ถœ ์ˆ˜ ์žˆ๋‹ค๊ณ  ๊ฐ€์ •์„ ํ•œ๋‹ค. ์˜ค๋ฉ”๊ฐ€1 = ๋‚จ์ž, ์˜ค๋ฉ”๊ฐ€2 = ์—ฌ์ž , X = ํ‚ค, ๋ชธ๋ฌด๊ฒŒ ๋“ฑ๋“ฑ์˜ ๋ฐ์ดํ„ฐ ํ•ด๋‹น Condition์„ ํ™•๋ฅ ๋กœ ๋‚˜ํƒ€๋‚ด๊ฒŒ ๋œ๋‹ค. ํ•ด๋‹น ํ™•๋ฅ ์„ ๊ณ„์‚ฐํ•˜๋ฉด ๋‚จ์ž์ธ์ง€ ์—ฌ์ž์ธ์ง€๋ฅผ ํŒ๋‹จํ•  ์ˆ˜ ์žˆ๋Š” ๊ฐ€๋Šฅ์„ฑ์ด ์žˆ๊ฒŒ ๋œ๋‹ค. 175cm์˜ 70kg๋ฅผ ๋ชจ๋‘ ๋ชจ์•„์„œ ๋‚จ์ž์™€ ์—ฌ์ž์˜ ์ˆ˜๋ฅผ ์„ธ๋Š” ๊ฒƒ์€ ๊ต‰์žฅํžˆ ์–ด๋ ต๋‹ค. ์™œ๋ƒํ•˜๋ฉด ๋ชจ๋“  ๊ฒฝ์šฐ์˜ ์ˆ˜๋ฅผ ์ง์ ‘ ์ฐพ์•„์„œ ๋‚˜์—ดํ•˜๋Š” ๊ฒƒ์€ ์–ด๋ ต๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค. ์ผ๋‹จ ๋‚จ์ž๋ฅผ ๋ชจ๋‘ ๋ชจ์•„์„œ ํ‚ค์™€ ๋ชธ๋ฌด๊ฒŒ์˜ ๋ถ„ํฌ๋ฅผ ์ฐพ๊ณ  ์—ฌ์ž๋„ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ๋ถ„ํฌ๋ฅผ ์ฐพ๋Š”๋‹ค. ์ด ๊ฒฝ์šฐ ์ „์ž์˜ ๊ฒฝ์šฐ๋ณด๋‹ค ํ›จ์”ฌ ์‰ฌ์šด ๊ฒฝ์šฐ๊ฐ€ ๋œ๋‹ค. ํ‚ค์™€ ๋ชธ๋ฌด๊ฒŒ๊ฐ€ ์ฃผ์–ด์ง„ ์ƒํƒœ์—์„œ ๋‚จ.. 2021. 12. 23.
Deep Learning(Classification, Regression, Overfitting) Classification ์ž…๋ ฅ์ด ํŠน์ •ํ•œ ์นดํ…Œ๊ณ ๋ฆฌ ์ค‘ ํ•˜๋‚˜๋‹ค๋ผ๊ณ  ์ธ์‹ํ•˜๋Š” ๊ฒƒ. ์ˆซ์ž, ์„ฑ๋ณ„, ๊ตญ์  ๋“ฑ ์ •ํ•ด์ง„ ์นดํ…Œ๊ณ ๋ฆฌ ์ค‘ ํ•˜๋‚˜๋ฅผ ๋งž์ถ”๋Š” ๋ฌธ์ œ ์ž…๋ ฅ = ํŒจํ„ด , ์ถœ๋ ฅ = Input data์˜ ์นดํ…Œ๊ณ ๋ฆฌ Regression value ์ž์ฒด๋ฅผ ๋งž์ถฐ์•ผ ํ•˜๋Š” ๋ฌธ์ œ Bankrupt prediction ์–ด๋–ค ํšŒ์‚ฌ๊ฐ€ ๋ถ€๋„๋‚  ํ™•๋ฅ ์„ ๊ณ„์‚ฐํ•˜๋Š” ๋ฌธ์ œ๊ฐ€ ํ•˜๋‚˜์˜ ์˜ˆ์‹œ Sales Prediction : ๋ช‡ ๊ฐœ๊ฐ€ ํŒ”๋ฆด ๊ฒƒ ์ธ๊ฐ€ ๋‹จ์ˆœํ•œ ์นดํ…Œ๊ณ ๋ฆฌ ๋ฌธ์ œ๊ฐ€ ์•„๋‹ˆ๋ผ Value๋ฅผ Estimate ํ•˜๋Š” ์ ‘๊ทผ์„ ํ•ด์•ผ ํ•œ๋‹ค. Input Variable์ด ์ฃผ์–ด์ง„๋‹ค. Regression Model function์„ ํ•™์Šตํ•œ๋‹ค. y = F(x; 0) 1. linear regression 2. nonlinear regression 3. logistic re.. 2021. 12. 23.