Gen-AI-Module1

Page 1: Course Information

  • Faculty Name: Dr J Alamelu Mangai

  • Designation: Professor

  • Department: CSE

  • Subject Code & Subject Name: CSE3348 Generative AI

  • Students: School of Computer Science and Engineering


Page 2: What is GenAI?

  • Generative AI (GenAI):

    • Set of algorithms that generates new content (image, text, audio, video).

    • Content resembles training data.

  • Large Language Model (LLM):

    • Prominent type of GenAI that generates natural language text based on prompts.

    • Example: GPT (Generative Pre-trained Transformer) series.

    • ChatGPT: Renowned example of LLMs.


Page 3: Generative Models Overview

  • Module 1: Introduction to Generative models

  • Goals:

    • Identify key milestones in generative models development.

    • State the evolution and applications of generative models.


Page 4: Generative AI Characteristics

  • Key Points:

    • Generates novel content.

    • Unlike traditional predictive ML, GenAI does not analyze existing data.

    • Capable of creating content indistinguishable from human-generated content.


Page 5: Generative vs. Discriminative Modeling

  • Discriminative Modeling:

    • Linked with supervised learning.


Page 6: What is Generative Modeling?

  • Definition:

    • Describes how a dataset is generated through a probabilistic model.

    • New data can be produced by sampling from this model.


Page 7: Generative Modeling Process

  • Key Components:

    • Training Data: Examples of the entities to generate.

    • Observation: One example from the training data.

      • Example: Image features defined by pixel values.

    • Probabilistic Nature: Model should include randomness.

    • Probabilistic Distribution: Identify distribution of images in training data.


Page 8: Realistic Data Generation

  • Process:

    • Model mimics probability distribution for realistic observations.

  • Data Labeling:

    • Discriminative modeling uses labeled data, while generative modeling typically uses unlabeled data.


Page 9: Generative Modeling Examples

  • Examples:

    • StyleGan by NVIDIA: Generates hyper-realistic human face images.

    • GPT by OpenAI: Completes given introductory passages.


Page 10: OpenAI Overview

  • Company:

    • OpenAI promotes friendly AI applications and began as non-profit in 2015, becoming for-profit in 2019.

  • Achievements:

    • Gym library for training reinforcement learning algorithms.

    • Development of GPT-n models and Dall-E for image generation from text.


Page 11: Understanding Generative Models

  • Artificial Intelligence (AI):

    • Broad field of CS focused on creating intelligent agents.

  • Machine Learning (ML):

    • Subset of AI concentrating on algorithms that learn from data.

  • Deep Learning (DL):

    • Uses deep neural networks for learning complex patterns.

  • Generative Models:

    • Specific machine learning models capable of creating new data analogous to training data.


Page 12: Key Components in Model Design

  • Core Concepts in AI:

    • Deep learning, statistical learning, language modeling.


Page 13: Applications of Generative Models

  • Functionality:

    • Generate various modalities of data: text, image, music, video.

    • Synthesize new data, particularly useful when real data is scarce.


Page 14: OpenAI Models Overview

  • Model Types:

    • GPT-4o: High-intelligence model for complex tasks.

    • GPT-4o mini: Affordable model for lightweight tasks.

    • DALL-E: Generates images from text prompts.

    • Text-to-Speech (TTS): Converts text to spoken audio.


Page 15: Historical Evolution of Generative AI

  • Key Milestones:

    • 1948: Claude Shannon's paper on n-grams.

    • 1950: Turing introduced Turing Test.

    • 1952: Hodgkin and Huxley’s model of brain neurons influenced artificial neural networks.


Page 16: Advances in Learning Algorithms

  • Milestones Continued:

    • 1965: First learning algorithm for feedforward neural networks.

    • 1979: Neocognitron introduced for handwritten digit recognition.

    • 1986: Backpropagation algorithm introduced for training neural networks.


Page 17: Further Development and Innovations

  • Continued Progress:

    • 1991: Introduction of Long Short-Term Memory (LSTM) networks.

    • 2001-2014: Variational Autoencoder (VAE), Generative Adversarial Networks (GAN) introduced.


Page 18: Expansion of Generative Models

  • Innovations in Generative Models:

    • Diffusion Models (2015): Reverse process learning introduced.

    • 2016: WaveNet creator for speech and music generation.


Page 19: Transformers and Their Influence

  • 2017: Introduction of Transformer architecture for sequence learning.

  • 2018: Launch of Generative Pre-trained Transformer (GPT), BERT model introduced.


Page 20: Recent Developments in LLMs

  • 2019-2020: StyleGAN introduction, GPT-3 launch with 175 billion parameters, and improvements in speech recognition through wav2vec 2.0.


Page 21: Applications and Ethical Considerations

  • 2021-2024 Progress:

    • Introduction of DALL-E for image generation, increased focus on ethics and real-world application integration.


Page 22: Advantages of Generative Modeling

  • Benefits:

    • Synthetic data generation reduces costs and improves training efficiency.


Page 23: Types of Generative Models Overview

  • Introduction to Different Models:

    • Identification of various fields for generative model applications is essential for understanding.


Page 24: Generative Models Applied in Different Fields

  • Purpose: Apply real-world case studies to identify challenges and solutions in implementing generative models.


Page 25: Types of Generative Models - Text-to-text

  • Examples: Conversational agents like LLaMa 2, GPT-4, Claude, PaLM 2.

    • Applications include interaction through NLP and machine learning technologies.


Page 26: Llama 2 Overview

  • Functionality:

    • Family of pre-trained LLMs capable of various NLP tasks, available for research and commercial use in 2023.


Page 27: GPT-3 Model Details

  • Overview:

    • Large language model by OpenAI with 175 billion parameters, designed for text generation in specific tasks.


Page 28: Types of Generative Models - Text-to-Image

  • Examples: DALL-E 2, Stable Diffusion, Imagen for generating images from text captions.


Page 29: Prompt Example for Text-to-Image

  • Prompt Usage Example: Storefront illustration request for visual generation.


Page 30: Text-to-Audio Models

  • Examples: Jukebox, AudioLM, MusicGen that generate music based on text input.


Page 31: Text-to-Video Applications

  • Models: Phenaki and Emu Video for generating video content from text descriptions.


Page 32: Text and Image Transformations

  • Examples: Image captioning, neural style transfer for image manipulation.


Page 33: Inpainting Applications

  • Example: Image defect removal applications.


Page 34: Text-to-Code Generation

  • Examples: Generate programming code from textual requests using models like DALL-E 3.


Page 35: Advanced Generative Modeling Techniques

  • Various Modalities: Explore interactions for audio, video, and mathematical representations from text inputs.


Page 36: Introduction to Prompt Engineering

  • Objectives: Define prompt engineering and express its significance in optimizing generative AI models.


Page 37: Prompt Engineering Basics

  • Definition: Prompts are natural language requests leading to model responses including diverse content types.


Page 38: Importance of Iteration in Prompting

  • Iterative Process: Updating prompts based on assessments to improve outputs for complex tasks.


Page 39: Components of a Prompt

  • Required and Optional Elements: Explains different components crucial for structured prompt design.


Page 40: Example of Task Prompt

  • Prompt Illustration: Explains how to pose questions to elicit desired responses from the model.


Page 41: Instructions in Prompts

  • System Instructions: Directions that dictate style and tone responses from models.


Page 42: Role Assignment in Model Interaction

  • Benefits: Enhances response quality by attributing specific roles to models for complex queries.


Page 43: Few-Shot Examples

  • Function of Few-Shot Examples: Provide guidance for expected outcomes in model responses.


Page 44: Contextual Information Utilization

  • Defining Context: Provides necessary information for the model to assist with problem-solving effectively.


Page 45: Safety and Fallback Responses

  • Fall-back Mechanisms: Highlights scenarios where models may refuse to respond based on service policies.


Page 46: Prompting Strategies for Better Responses

  • Key Principles: Discusses strategies to optimize model outputs via structured prompts.


Page 47: Iterative Workflow

  • Essential Steps: Emphasizes the importance of definitions and testing for effective prompt engineering.


Page 48: Factors Affecting Prompt Effectiveness

  • Content and Structure: Explains how organizing prompt information influences model output quality.


Page 49: Template for Creating Prompts

  • Prompt Template: Offers an example personalized structure for prompt formulation with optional components.


Page 50: Effective Prompt Design Strategies

  • Best Practices: Shares strategies including clear instructions and contextual information for optimal response.


Page 51: Case Studies on Prompt Clarity

  • Transcript Examples: Demonstrates extraction and formatting through example instructions.


Page 52: Few-Shot Prompting Techniques

  • Importance: Emphasizes how a few-shot approach can guide model behavior and output.


Page 53: Zero-Shot vs Few-Shot

  • Comparison: Highlights differences between prompts without examples versus with examples.


Page 54: Role Assignment in Queries

  • Advantages of Role Assignment: Improving output accuracy and relevance based on assigned roles in prompts.


Page 55: Inclusion of Contextual Details

  • Contextual Understanding: Elaborates on the necessity of including relevant background information in prompts.


Page 56: System Instruction Implementation

  • Application: Introduces system instructions as a crucial element in customizing model behavior.


Page 57: Enhancing Clarity with Structure

  • Effective Layout: Advises on using layout for clarity within prompt design to aid response generation.


Page 58: Reasoning and Explanation Necessity

  • Output Specificity: Encourages the inclusion of reasoning in responses to enhance understanding.


Page 59: Breaking Down Complex Tasks

  • Prompt Segmentation: Discusses methods for managing complex tasks through breakdown into subtasks.


Page 60: Example of Sequential Task Breakdown

  • Sequential Processing: Demonstrates utilizing chain prompts for handling multifaceted processes.


Page 61: Parallel Tasks in Prompt Management

  • Simultaneous Execution: Highlights methods for analyzing multiple data sets through aggregated responses.


Page 62: Model Parameter Experimentation

  • Value Adjustments: Explains how modifying model parameters can lead to more desirable outcomes.


Page 63: Iterative Improvement for Prompting

  • Continuous Refinement: Stresses need for regular assessment in prompt design for performance enhancement.


Page 64: Overview of Large Language Models (LLMs)

  • Definition of LLMs: Explains functionality of LLMs in predicting and generating language-based content.


Page 65: Textual Comprehension in LLMs

  • Understanding Language Relationships: Discusses capabilities for high-context applications and generation.


Page 66: Embeddings Role in LLM Performance

  • Importance of Representation: Details how embeddings assist in text and image representation in learning.


Page 67: Methods for Generating Embeddings

  • Techniques Discussion: Discusses various approaches for creating effective embeddings in deep learning.


Page 68: Model Complexity and Scalability

  • Resource Intensive Models: Uncovers the challenges of training large models efficiently over extensive datasets.


Page 69: Applications of Large Models

  • Common Uses: Illustrates practical scenarios where large language models can be effectively utilized.


Page 70: Variants of Large Language Models (LLMs)

  • Types of Models: Overview of model types addressing specific challenges and needs in natural language tasks.


Page 71: Learning Paradigms of LLMs

  • Overview of Architectures: Provides insights into the design and operational mechanisms of LLMs.


Page 72: Generative Pre-trained Transformers (GPT) Overview

  • Definition and Functionality: Explains the operational principles behind the development of GPT models.


Page 73: Generative Modeling Strengths and Weaknesses

  • Analysis of GPT Models: Identifies the advantages and disadvantages of various generative approaches.


Page 74: Practical Applications of GPT Models

  • Sector/Field Use Cases: Highlights how GPT models can be employed across various domains for effective output.


Page 75: Generative Pre-trained Transformers Family

  • GPT Evolution: Tracks the advancements across different versions of the GPT models released by OpenAI.