Generative Adversarial Networks (GAN)-32

Generative Adversarial Networks (GAN

1. Generative Adversarial Networks (GANs) An Overview GALGOTIAS COLLEGE OF ENGINEERING AND TECHNOLOGY 1, Knowledge Park-II, Greater NOIDA, Uttar Pradesh, 201310 Department of Data Science

2. At the end of course, the student will be able: CO 2 : To study the concepts of deep learning Deep Learning ( BAI701) GALGOTIAS COLLEGE OF ENGINEERING AND TECHNOLOGY 1, Knowledge Park-II, Greater NOIDA, Uttar Pradesh, 201310 Department of Data Science

3. Introduction to GANs • Proposed by Ian Good fellow in 2014 • A framework for training generative models • Consists of two networks: Generator and Discriminator • Based on adversarial training

4. GAN Architecture • Generator: Produces fake data from random noise • Discriminator: Distinguishes real data from fake data • Both are trained in a minimax game setup • Objective: Improve generator to fool discriminator

5. GAN Architecture

6. Training Process 1. Generator creates fake samples 2. Discriminator evaluates samples 3. Discriminator improves classification ability 4. Generator improves to produce more realistic data 5. Process repeats until generator produces realistic data

7. Applications of GANs • Image generation (faces, art, objects) • Image-to-image translation • Text-to-image synthesis • Data augmentation • Super-resolution • Style transfer

8. Challenges in GANs • Training instability • Mode collapse • Requires large datasets • Difficult evaluation metrics

9. Conclusion • GANs are powerful generative models • Achieved state-of-the-art results in many tasks • Still face challenges in training and evaluation • Active area of research with wide applications