Convolutional Neural Network

11 Jan 2023 in Notes / Artificialintelligence / Introductiontoai

DAY 13

• Introduction
• Goal of Generative Model

Introduction

• (참고) RNN Skipped

  GAN: Generative Adversarial Network

• Training both of generator and discriminator
• -> generates samples similar to the original samples

generator	discriminator
making something from scratch (FAKE DATA)	determines whether such data is GOOD/BAD

• Both generator and discriminator’s performance improves via competitions
  • ex: 위조지폐 생성실력과 경찰의 분류 및 검증능력이 날이 갈수록 발전하는 것 처럼

• REAL DATA required for the discriminator in order to successfully classify fake data
• Classifies ( 0 / 1 ) => Same as Binary Classification

• DIFFERENCE FROM OTHER AI MODELS: NO INPUT

• still, it DOES require any form of input for implementation
• Noise is used

Goal of Generative Model

• Find a \(p_{model}(x)\) that approximates \(p_{data}(x)\) well
  • \(p_{data}(x)\) : represents distribution of actual images
  • \(p_{model}(x)\) : of fake data, needs to be as close to \(p_{data}(x)\)
• \(Đ(X)\) = 1 (or close to): Training with real data \(X\) should be high.
• \(Đ(G(X))\) = 0 (or close to): Generated(\(G(\quad)\)) with noise data (\(z\)) should be low

\[\min_{G} \max_{D}V(D,G) = E_{x \tilde p_{data}(x)}[logD(x)] + E_{z \tilde p_{z}(z)}[log(1-D(G(z)))]\]

• Objective of D :

  • \(D\) should maximize V(D,G)
  • First Part (\(E_{x \tilde p_{data}(x)}[logD(x)]\)): Maximize when \(D(x) = 1\)
  • Second Part (\(E_{z \tilde p_{z}(z)}[log(1-D(G(z)))]\)) : Maximaize when \(D(G(z))=0\)

• Objective of G :

  • \(G\) should minimize V(D,G)
    • => Tries to make Fake Data PASS the discriminator (D)
  • First Part : NOT REQUIRED since Generator (G) will not have access to REAL DATA
  • Second Part (\(E_{z \tilde p_{z}(z)}[log(1-D(G(z)))]\)) : Minimum when \(D(G(z))=1\)
    • \[\min_{G} E_{z \tilde p_z(z)}[log(1 - D(G(z)))]\]
    • \[= \max_{G} E_{z \tilde p_z(z)}[log(D(G(z)))]\]
    • latter is used since it converges quicker (according to Statistic Experts)