Deep Learning Foundations for Gen AI

Description:

This training provides an in-depth introduction to Generative AI, exploring the foundational concepts, key techniques, and applications of generative models in Deep learning. Students will gain a practical understanding of how to develop and apply models such as Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and other advanced generative approaches. The course will also cover the ethical considerations, challenges, and real-world use cases of generative AI in various domains, including natural language processing, image generation, and more.

Duration: 2 Days

Course Code: BDT395

Learning Objectives:

By the end of this course, participants will:

• Understand the core principles of Generative AI and its various models.
• Design, implement, and train generative models like GANs and VAEs.
• Apply generative models to solve real-world problems in image generation, text generation, and more.
• Navigate ethical and regulatory considerations in deploying generative AI models.
• Stay updated with emerging trends and technologies in the field of Generative AI.

• Basic understanding of machine learning concepts (supervised and unsupervised learning). Familiarity with deep learning frameworks like TensorFlow or PyTorch. Basic programming skills in Python.

• Software developers, engineers, or machine learning practitioners with basic Python knowledge who want to learn how to build and implement generative AI models.

  Familiarity with machine learning concepts (like neural networks) is helpful but not required.

Course Outline:

Module 1: Introduction to Generative AI

• 1 What is Generative AI?
  • Definition and Overview of Generative Models
  • Difference between Generative and Discriminative Models
  • Applications of Generative AI
• 2 Evolution of Generative AI
  • Historical Context and Key Milestones
  • Development of GANs, VAEs, and Transformer-based models
  • Generative AI in the Context of Deep Learning
• 3 Overview of Machine Learning Models
  • Supervised, Unsupervised, and Reinforcement Learning
  • Key Concepts in ML that are Related to Generative AI

Module 2: Key Techniques in Generative AI

• 1 Generative Adversarial Networks (GANs)
  • Architecture of GANs: Generator vs. Discriminator
  • Training Process and Challenges in GANs
  • Types of GANs: DCGAN, CycleGAN, StyleGAN, and more
  • Applications of GANs in Image Generation, Style Transfer, and Super Resolution
• 2 Variational Autoencoders (VAEs)
  • The VAE Architecture: Encoder, Decoder, and Latent Space
  • Understanding the Variational Inference and ELBO
  • Applications in Image Generation and Data Reconstruction
• 3 Other Generative Models
  • Autoregressive Models (e.g., PixelCNN, WaveNet)
  • Flow-based Models (e.g., RealNVP, Glow)
  • Diffusion Models and their Recent Popularity

Module 3: Deep Dive into GANs

• 1 Advanced GANs Architecture
  • Conditional GANs (cGANs)
  • Wasserstein GANs (WGANs) and WGAN-GP
  • Progressive Growing of GANs
  • Self-Attention GANs (SAGAN)
• 2 Training GANs
  • Challenges in GAN Training: Mode Collapse, Non-Convergence
  • Techniques to Improve GAN Stability
  • Evaluating GAN Performance: Inception Score, FID Score
• 3 Applications of GANs
  • Image Generation and Synthesis
  • Style Transfer and Deepfakes
  • Data Augmentation for Training ML Models
  • Text-to-Image Generation

Module 4: Variational Autoencoders (VAEs)

• 1 Understanding Latent Variables in VAEs
  • The Role of Latent Variables in Data Generation
  • Reconstructing Data via VAEs
  • ELBO and its Role in Training VAEs
• 2 Advanced Topics in VAEs
  • Conditional VAEs (CVAE)
  • VAE for Generating Complex Data (Images, Text, etc.)
  • Disentangled VAEs and their Applications
• 3 Applications of VAEs
  • Image and Text Generation
  • Anomaly Detection
  • Latent Variable Modeling in Healthcare and Finance

Module 5: Transformers and Attention Mechanisms in Generative AI

• 1 The Transformer Architecture
  • Attention Mechanism and Self-Attention
  • Transformer-based Models: BERT, GPT, T5, and more
• 2 Generative Transformers
  • GPT Series: Architecture and Training of Large Language Models
  • Text Generation with Transformer Models
  • Fine-Tuning Transformers for Specific Generative Tasks
• 3 Applications of Transformer-based Generative Models
  • Text Generation, Summarization, Translation
  • Music Composition and Code Generation
  • Multimodal Generative Models (e.g., CLIP, DALL·E)

Module 6: Practical Implementation of Generative AI

• 1 Tools and Frameworks for Generative AI
  • Popular Libraries: TensorFlow, PyTorch, Keras
  • Specialized Libraries for GANs and VAEs (e.g., PyTorch-GAN, TensorFlow-GAN)
  • Cloud-based tools for large-scale generative model training (e.g., Google Colab, AWS, GCP)
• 2 Building and Training GANs
  • Step-by-step guide to building a GAN for Image Generation
  • Data Preprocessing and Augmentation Techniques
  • Training GANs and Fine-Tuning Hyperparameters
• 3 Building and Training VAEs
  • Step-by-step guide to building a VAE for Data Reconstruction
  • Hyperparameter Tuning in VAEs
  • Applications of VAEs in Healthcare and Biomedicine

Module 7: Ethical and Practical Challenges in Generative AI

• 1 Ethical Considerations in Generative AI
  • Impact of Generative AI on Society
  • Bias and Fairness Issues in Generative Models
  • Addressing Deepfakes, Misinformation, and Fake Media
• 2 Responsible Use of Generative AI
  • Ensuring Transparency and Accountability
  • Guardrails and Regulatory Frameworks
  • Privacy and Security Concerns in AI Models
• 3 Addressing the Environmental Impact of Generative AI
  • Computational Resources and Energy Consumption
  • Techniques for Improving Efficiency in Training Large Models

Module 8: Applications of Generative AI

• 1 Generative AI in Image and Video Synthesis
  • Applications in Film, Animation, and Art Generation
  • GANs in Face and Style Transfer
  • Video Synthesis and Deepfake Detection
• 2 Natural Language Generation
  • Text-to-Image Models like DALL·E
  • Story Generation and Content Creation with GPT-based Models
  • Chatbots and Conversational AI using Generative Models
• 3 Music and Audio Generation
  • Generative Models for Music Composition (e.g., OpenAI Jukedeck)
  • Text-to-Speech and Voice Synthesis
  • Speech-to-Text and Text-to-Speech with Generative Models
• 4 Healthcare and Scientific Applications
  • Data Synthesis in Medical Research
  • Drug Discovery and Molecular Generation using GANs
  • Predictive Modeling and Data Augmentation

Training material provided: Yes (Digital format)