Generative Adversarial Network (GAN) for Image Synthesis

This is a short yet informative presentation on one of the most popular machine learning model Generative Adversarial Network (GAN).

1. GAN Image Synthesis Date: 14 Dec,2023 Group Members Riwaz Mahat Ashim Neupane Prasis Gautam

2. Challenges and Future of GANs 2. 4. Types of GANs and Use Cases Architecture of GANs 1. 3. Brief overview of GANs How GANSs work Key concepts Generator Discriminator Overview of different types of GANs Real-world use cases of GANs Detailed look at the architecture of GANs Discussion of challenges in training GANs Future trends and research TABLE OF CONTENTS Understanding the GANs

3. WHAT EXACTLY IS GAN ? GAN, Generative Adversarial Network is a type of machine learning model comprising two neural networks: Generator and Discriminator, competing against each other to generate realistic data, enabling the creation of high quality synthetic content such as images, videos, and text. GANs leverage a game-theoretic framework where the generator learns to produce increasingly convincing data while the discriminator aims to distinguish between real and generated samples, fostering the generation of diverse and realistic outputs. HOW DOES IT WORK ?

4. UNDERSTANDING GAN KEY CONCEPTS GENERATOR DISCRIMINATOR Generator : Creates synthetic data resembling the real dataset from random noise. Discriminator : Distinguishes between real and synthetic data, improving its accuracy. Adversarial Training : Simultaneous training of generator and discriminator in a competitive manner. Loss Function : Guides training by measuring network performance. Generator : produces synthetic data from noise input. Discriminator : Distinguishes between real and synthetic data. Adversarial Process : Generator deceives discriminator and it distinguishes better. Iterative : Both networks improve until generator creates highly realistic data. Outcome : High-quality synthetic data creation. WORKING

5. Neural network layers which generates realistic data to deceive the discriminator GENERATOR Neural network layers for distinguishing real from generated data which enhances accuracy in discriminating real and fake data DISCRIMINATOR ARCHITECTURE OF GAN It follows simultaneous training where generator improves to create more convincing data and discriminator enhances discrimination abilities TRAINING PROCESS GANs evolve through adversarial training to produce high-quality, realistic synthetic data resembling the original dataset OUTCOME

6. TYPES OF GAN Vanilla GAN: This is the simplest type of GAN, composed of a generator and a discriminator. The generator captures the data distribution, while the discriminator tries to determine the probability of the input. Conditional GAN (CGAN): Here, both the generator and discriminator are provided with additional information, such as a class label or any modal data. This extra information assists the discriminator in determining the conditional probability instead of the joint probability. Deep Convolutional GAN (DCGAN): This is the first GAN where the generator used a deep convolutional network, resulting in the generation of high-resolution and quality images. CycleGAN: This GAN is designed for Image-to-Image translations, meaning one image is mapped to another image. For instance, it can convert summer images into winter images and vice versa by adding or removing features. Generative Adversarial Text to Image Synthesis: This type of GAN is used to generate images from text descriptions.

7. REAL WORLD USE CASES GANs can generate new, realistic images that are similar but specifically different from a dataset of existing photographs. This can be used for tasks like creating new designs, generating artwork, or producing realistic video game graphics. IMAGE SYNTHESIS 01 GANs can convert one type of image into another. For example, CycleGAN can convert summer images into winter images and vice versa. IMAGE-TO-IMAGE TRANSLATION 02 GANs can generate images from text descriptions. This can be used in a variety of applications, such as creating visual content from written descriptions or aiding in the design process. Text-to-Image SyNTHESIS 03

8. CHALLENGES Hindered training due to gradient issues. VANISHING GRADIENTS Lack of standardized metrics for GAN assessment. EVALUATION METRICS High sensitivity to hyperparameter values. HYPERPARAMETER SENSITIVITY Limited variety of generated outputs and techniques. MODE COLLAPSE Convergence difficulties between generator and discriminator. TRAINING INSTABLITY

9. FUTURE TRENDS AND RESEARCH OF GAN Improved Stability and Training Techniques Diversity and Realism Enhancement Interdisciplinary Applications Ethical Considerations and Regulations Hardware & Software Advancements Adversarial Learning Beyond GANs

10. CONCLUSION ANY QUESTIONS ? In simple terms, Generative Adversarial Networks (GANs) are a cool technology in artificial intelligence. They use two parts, a generator and a discriminator, to create realistic fake data. GANs have been awesome for making lifelike medias like photos, vidoes, graphics and more. They're like a creative duo where one tries to make things look real, and the other tries to figure out if they're fake. Despite their success, challenges such as training stability, mode collapse, and ethical considerations remain areas of ongoing research. Overall, GANs have opened up exciting possibilities in AI, making things like generating realistic content a lot more fun and interesting.