01_-_Generative_AI

Page 1: Introduction

• Overview of the session on Large Language Models and Generative AI.

• Instructor: Ivika Jäger, PhD

• Date: November 2024

• Institution: Mittuniversitetet, Mid Sweden University

Page 2: Agenda

• Topics to be covered:

  • Machine learning methods

  • MIUN guidelines

  • Generative AI

  • Prompt design

Page 3: Machine Learning Methods

• Introduction to different methods of machine learning.

Page 4: AI as a General Purpose Technology

• Characteristics of AI:

  • Highly adaptable and flexible

  • Lacks a user manual; requires trial and error

  • Applications based on user understanding of problems

Page 5: Understanding Machine Learning (ML)

• Definition: Machines learning from data without explicit programming.

• Key functions:

  • Recognizes patterns

  • Makes decisions

  • Improves with experience

• Essential to modern AI applications.

Page 6: Applications of Machine Learning

• Various types and applications:

  • Image Compression

  • Customer Retention

  • Discovery and Classification

• Types of ML:

  • Supervised Learning

  • Unsupervised Learning

  • Reinforcement Learning

• Fields of application:

  • Market predictions

  • Fraud detection

  • Weather forecasting

Page 7: Supervised Learning

• Uses labeled data to train the model.

• Outputs predictions based on learned mapping.

• Ideal for classification problems (image recognition, spam detection).

Page 8: Interactive Training

• Tools like Teachable Machine for practical model training.

Page 9: Unsupervised Learning

• Operates on unlabeled data to discover patterns.

• Analyzes data for clustering and anomaly detection.

• Suitable for market segmentation and fraud detection.

Page 10: K-Means Algorithm

• Steps in K-means clustering:

  1. Initialize centroid positions.

  2. Assign labels to data points.

  3. Update centroid positions.

  4. Repeat until convergence.

Page 11: Reinforcement Learning

• Agents learn through interactions and receive rewards/penalties.

• Focuses on maximizing reward in complex environments (e.g., autonomous vehicles).

Page 12: AI Learning

• Example use case in a gaming scenario (Trackmania) over days.

Page 13: Deep Learning (DL)

• Utilizes neural networks with multiple layers.

• Effective for large unstructured data and complex problems (e.g., navigation).

Page 14: Distinction Between ML and DL

• Structured vs. unstructured data processing.

• Manual feature definition vs. automated learning.

• Development of Large Language Models (LLMs) via deep learning.

Page 15: Using AI at MIUN

• Discussion on the integration of AI in educational practices.

Page 16: Considerations

• Importance of 'when' and 'how' to use Generative AI.

• Considerations:

  • Limitations and opportunities in academic settings.

  • Issues of privacy, security, and ethics.

• AI concept recognition among students and its impact on academic integrity.

• MIUN resources for students regarding academic integrity and AI use.

• Encouragement of responsible AI use while maintaining academic integrity.

• Practices include brainstorming, feedback, and tutoring.

• Hands-on approach to learning with AI.

• Emphasis on transparency and discussing outcomes.

• Risks of direct copy-pasting without engagement.

• Importance of critical interaction and proper prompting.

Page 22: Generative AI Introduction

• Contrast between Generative AI and traditional AI.

• Generative AI creates content mimicking human output.

Page 23: Functionality of Generative AI

• Trained on vast datasets to generate responses based on prompts.

Page 24: AI vs. Generative AI

• Traditional AI focuses on tasks like classification, while Generative AI synthesizes new data.

Page 25: History of Large Language Models

• Significant developments:

  • 2020: OpenAI GPT-3

  • 2017: Google Brain

  • Other notable milestones.

Page 26: Pre-Training of Models

• Learning patterns before fine-tuning for specific applications.

• Techniques:

  • Next Token Prediction

  • Masked Language Modeling

Page 27: Fine-Tuning Process

• Involves Reinforcement Learning with Human Feedback (RLHF).

• Incorporates human ratings to understand and improve model performance.

Page 28: LLM's Mission

• Prioritization of user satisfaction over accuracy.

• Risks of anthropomorphism and confirmation bias.

Page 29: Capabilities of LLMs

• Functions such as text generation, translation, data analysis, and creative assistance.

Page 30: Limitations of LLMs

• Recognized limitations include lack of comprehension and ethical concerns.

Page 31: Observation Data

• Analysis of academic articles using OpenAlex concerning specific keywords.

Page 32: Model vs. Application

• Distinction between LLMs as core technology and AI applications.

Page 33: Introduction to Prompt Engineering

• Importance of skill in guiding AI outcomes through effective prompt design.

Page 34: Prompt Engineering Attributes

• Initially relevant to software developers but valuable to all users.

Page 35: Definition

• Process of designing prompts for AI to achieve desired responses.

Page 36: Importance of Prompt Design

• Specific and clear prompts lead to more insightful AI outputs.

Page 37: Benefits of Good Prompt Design

• Consistency, accuracy, efficiency resulting from structured prompts.

Page 38: Programmatic Applications

• API automation for repeated utilization of prompts with varying outputs.

Page 39: Key Principles of Prompting

• Emphasis on clarity, specificity, precision, and iterative improvements.

Page 40: Clarity and Specificity

• Defining expectations for AI outputs, specifying required details, and framing prompts positively.

Page 41: Example Prompt

• Requesting character identification from a provided text context.

Page 42: Structure of Prompts

• Using delimiters to separate tasks, context, and constraints within the input.

Page 43: Desired Precision

• Control the length and depth of responses based on task requirements.

Page 44: Clarifying Tasks

• Instructions for summarizing text and specifying length limitations.

Page 45: Structured Output Requests

• Examples of formats for AI outputs and guidelines for when conditions are unmet.

Page 46: Example Sentiment Analysis

• Tasking AI to classify sentiments of social media comments.

Page 47: Example Result

• Sample output categorizing sentiment from social media comments.

Page 48: Few-shot Prompting

• Providing examples to guide AI style and structure.

Page 49: Example Creative Prompt

• Generating a catchphrase for a product with given constraints and context.

Page 50: Iterative Improvement

• Importance of revising prompts based on initial outputs for optimal results.

Page 51: Example Professional Email

• Tasking AI to craft a structured professional email with constraints.

Page 52: Advanced Techniques

• Techniques for guiding AI reasoning and maintaining consistent responses.

Page 53: Prompting for Improvement

• Generating creative ideas for user engagement in social media contexts.

• Options available for further engagement with AI models.

• Links to guides and platforms for effective prompt engineering and AI usage.

• Instructions for analyzing sentiment and translations of social media comments.

• Tasks involving summarizing, proofreading, and translating emails for clarity.

• Extracting key information from passages for deeper understanding.

• Measuring understanding and confidence in using LLMs post-lecture.