Note_Tech

All technological notes.

---

Project maintained by simonangel-fong Hosted on GitHub Pages — Theme by mattgraham

Basic Concept - AI & LLM

Back

• Basic Concept - AI \& LLM
  • AI Overview
  • ML Overview
    • Deep Learning Overview
    • Generative Artificial Intelligence(AI)
  • Foundation Models(FM)
    • Complexity of Foundation Model
    • Categorization

---

AI Overview

• Artificial intelligence (AI):
  • a branch of computer science that enables machines to simulate human behavior,
• Machine learning (ML):
  • a subset of artificial intelligence (AI) that enables machines to use data and algorithms to simulate human behavior
• Deep Learning(DL):
  • Subset of Machine learning(ML) which use neural networks to simulate human behavior
• Generative AI
  • Subset of deep-learning models that can generate highquality text, images, videos

---

ML Overview

• algorithm is trained using the historical data to make predictions on new data.

• 3 Categories:

  • Supervised Learning
    • algorithms are trained using labelled dataset, tested using test data
    • Common Algorithms
      • Classification: discrete category
      • Regression: a continuous numerical value
  • Unsupervised Learning
    • algorithm are trained using unlabeled dataset
    • finds hidden patterns and insights based on the similarities of the data
    • Common Algorithms:
      • Clustering: groups similar data points (rows) into clusters
      • Association: finds rules and relationships between variables (items) that frequently occur together
  • Reinforcement Learning
    • training an algorithm following the trial and error approach.
    • algorithm (agent) evaluates a current situation (state), takes an action, and receives feedback (reward and feedback) from the environment after each act.
    • No pre-defined datase
    • Use Case:
      • Training Industrial Robots
      • Training Self Driving Cars

---

Deep Learning Overview

• Deep Learning - focuses on building artificial neural networks that can learn from data.
• Artificial Neural networks
  • mimic the structure of the human brain
  • input layer
  • hiddent layer
  • Output layer

---

Generative Artificial Intelligence(AI)

• focuses on generating high-quality text, images, code

• Inference: Act of using model to generate text is called Inference

• 1. Prompt: Input provided to Model
• 1. Tokenizer: breaks down raw text into smaller, manageable units called tokens
     • token: the fundamental, small units of data—words, characters, or subwords—that AI models use to process and generate language.
     • token id: a unique integer assigned to a specific token (a word, part of a word, or character) from a model’s pre-defined vocabulary.
• 1. context window/working memory:
     • total amount of information (input prompt + output response) an Large Language Model (LLM) can process at one time, measured in tokens.
     • Max tokens: the upper limit on tokens (text fragments) a model can handle in one request, including both input prompt and output response
     • stop sequence user-defined string of characters, words, or tokens that signals a large language model (LLM) to cease text generation immediately upon producing it
• 1. Model/Foundation Model: GPT, Claude
• 1. detokenizer
     • the reverse process of tokenization, converting a sequence of numerical tokens (IDs) or split subwords generated by a language model back into readable, natural language text
• 1. Completion: Output of Model

---

Foundation Models(FM)

• Key Characteristics
  • Deep neural networks techniques
  • Unlabeled data - Trained on unlabeled data such as internet data - Wikipedia
  • Generalized – One FM for multiple tasks
  • Pre-trained – FMs are pre-trained models
  • Large – Exposed to vast amounts of data, enabling it to learn patterns & features from data.
  • Expensive - Expensive process that involves large amounts of computational resources (GPU’s) & Organization’s can utilize technical expertise.
• Can utilize existing FMs as the base for building task-specific models and Adapt them

---

Complexity of Foundation Model

• Large Language Models (LLM):
  • More number of Foundation Model Parameters
    • 8B to Trillions of parameters
  • Good for Complex Task + Higher response time + Costlier
• Small Language Model (SLM):
  • Lesser number of Foundation Model Parameters
    • Less than 8B parameters
  • Good for Simpler Task + Lower response time + Cheaper

---

Categorization

• Large Language Models (LLM’s):
  • Focus Area : Text Generation
  • UC: Content generation, Summarization
• Image Generation or Diffusion Model
  • Focus Area : Image Generation, Diffusion Model
  • UC : Marketing, Social Media
• Multi-Modal Model
  • Focus Area : Image Generation, Diffusion Model
  • UC : Healthcare/Physician, Insurance
• Embedding Model
  • Focus Area : Embeddings
  • UC : HR Chatbot, eCommerce

---