Skip to content

AI and ML Fundamentals for AWS GenAI

Artificial Intelligence Overview

Types of AI

Narrow AI (Weak AI) - Designed for specific tasks - Cannot generalize beyond training - Examples: Image recognition, voice assistants, recommendation systems - Current state of commercial AI

General AI (Strong AI) - Human-level intelligence across all domains - Theoretical, not yet achieved - Can transfer knowledge between domains - Learn and adapt to new situations

Super AI - Surpasses human intelligence - Theoretical concept - Not currently possible

AI vs Machine Learning vs Deep Learning

┌─────────────────────────────────────────────┐
│ Artificial Intelligence                      │
│  ┌────────────────────────────────────────┐ │
│  │ Machine Learning                        │ │
│  │  ┌────────────────────────────────────┐│ │
│  │  │ Deep Learning                       ││ │
│  │  │  (Neural Networks)                  ││ │
│  │  └────────────────────────────────────┘│ │
│  └────────────────────────────────────────┘ │
└─────────────────────────────────────────────┘

Artificial Intelligence - Computer systems that mimic human intelligence - Rule-based systems, expert systems - Symbolic reasoning and knowledge representation

Machine Learning - Subset of AI - Learn from data without explicit programming - Statistical learning and pattern recognition - Improves with more data

Deep Learning - Subset of ML - Neural networks with multiple layers - Learns hierarchical representations - Requires large datasets and compute

Machine Learning Fundamentals

Types of Machine Learning

Supervised Learning - Labeled training data (input-output pairs) - Learns to map inputs to outputs - Examples: Classification, regression - Use cases: Spam detection, price prediction

Classification - Discrete output (categories) - Binary: Yes/No, Spam/Not Spam - Multi-class: Animal types, product categories - Algorithms: Decision trees, SVM, neural networks

Regression - Continuous output (numbers) - Predict values (price, temperature, sales) - Linear regression, polynomial regression - Algorithms: Linear regression, neural networks

Unsupervised Learning - Unlabeled data - Find patterns and structure - Examples: Clustering, dimensionality reduction - Use cases: Customer segmentation, anomaly detection

Clustering - Group similar data points - K-means, hierarchical clustering - Market segmentation, document grouping

Dimensionality Reduction - Reduce number of features - Preserve important information - PCA, t-SNE - Visualization, noise reduction

Reinforcement Learning - Learn through trial and error - Rewards and penalties - Agent, environment, actions, rewards - Use cases: Game playing, robotics, autonomous vehicles

Semi-Supervised Learning - Mix of labeled and unlabeled data - Leverages large unlabeled datasets - Reduces labeling cost - Use cases: Speech recognition, text classification

ML Workflow

1. Problem Definition - Define business objective - Identify success metrics - Determine ML problem type

2. Data Collection - Gather relevant data - Multiple sources (databases, APIs, web scraping) - Data quality assessment

3. Data Preparation - Data cleaning (handle missing values, outliers) - Feature engineering (create meaningful features) - Data transformation (normalization, encoding) - Train/validation/test split

4. Model Selection - Choose algorithm type - Consider problem requirements - Baseline model first

5. Model Training - Feed training data to model - Optimize parameters - Monitor training metrics (loss, accuracy) - Prevent overfitting (regularization, dropout)

6. Model Evaluation - Test on validation/test set - Metrics: Accuracy, precision, recall, F1, RMSE - Compare against baseline - Error analysis

7. Model Deployment - Deploy to production - API endpoints - Batch or real-time inference - Monitoring and maintenance

8. Model Monitoring - Track performance metrics - Detect data drift - Model degradation - Retrain as needed

Key Concepts

Features - Input variables (independent variables) - Characteristics used for prediction - Feature engineering: Creating new features

Labels - Output variables (dependent variables) - What we're trying to predict - Ground truth in training data

Training - Process of learning from data - Adjusting model parameters - Minimizing error/loss

Inference - Using trained model for predictions - Production deployment - Real-time or batch processing

Overfitting - Model too complex for data - Memorizes training data - Poor generalization - Solutions: Regularization, more data, simpler model

Underfitting - Model too simple - Cannot capture patterns - Poor performance on training and test data - Solutions: More features, complex model, more training

Bias-Variance Tradeoff - Bias: Error from wrong assumptions (underfitting) - Variance: Error from sensitivity to training data (overfitting) - Goal: Balance both for good generalization

Evaluation Metrics

Classification Metrics

Confusion Matrix 

                 Predicted
               Positive  Negative
Actual Positive   TP        FN
       Negative   FP        TN

  • Accuracy: (TP + TN) / Total
  • Precision: TP / (TP + FP) - Of predicted positives, how many are correct?
  • Recall (Sensitivity): TP / (TP + FN) - Of actual positives, how many did we find?
  • F1 Score: 2 × (Precision × Recall) / (Precision + Recall)
  • Specificity: TN / (TN + FP)

When to Use - High precision: Spam detection (avoid false positives) - High recall: Disease detection (don't miss cases) - F1 score: Balanced measure

Regression Metrics - MAE (Mean Absolute Error): Average absolute difference - MSE (Mean Squared Error): Average squared difference - RMSE (Root Mean Squared Error): Square root of MSE - R² (R-squared): Proportion of variance explained

Deep Learning Fundamentals

Neural Networks

Architecture - Input layer: Receives features - Hidden layers: Process information - Output layer: Produces predictions - Neurons: Individual processing units - Weights: Connection strengths - Activation functions: Non-linear transformations

Activation Functions - ReLU (Rectified Linear Unit): max(0, x) - Most common - Sigmoid: 1/(1+e^-x) - Binary classification - Tanh: Hyperbolic tangent - Range [-1, 1] - Softmax: Multi-class probabilities

Training Process 1. Forward propagation: Calculate predictions 2. Loss calculation: Measure error 3. Backpropagation: Calculate gradients 4. Parameter update: Adjust weights

Optimization - Gradient Descent: Move toward minimum loss - Learning Rate: Step size for updates - Batch Size: Samples per update - Epochs: Complete passes through data

Convolutional Neural Networks (CNNs) - Specialized for image data - Convolutional layers: Learn spatial features - Pooling layers: Reduce dimensions - Applications: Image classification, object detection

Recurrent Neural Networks (RNNs) - Sequential data processing - Memory of previous inputs - LSTM, GRU: Handle long-term dependencies - Applications: Text, time series, speech

Transformers - Self-attention mechanism - Process sequences in parallel - Foundation for modern LLMs - Applications: NLP, language models

Transfer Learning

Concept - Use pre-trained model as starting point - Fine-tune on specific task - Requires less data and compute - Faster training

Approaches 1. Feature Extraction: Use pre-trained model as feature extractor 2. Fine-tuning: Update some or all layers 3. Domain Adaptation: Adapt to different but related domain

Benefits - Reduced training time - Less data required - Better performance with limited data - Leverage large pre-trained models

Generative AI Fundamentals

What is Generative AI?

Definition - AI that creates new content - Learns patterns from training data - Generates novel outputs similar to training data - Text, images, code, audio, video

vs Discriminative AI - Discriminative: Classify or predict (Y given X) - Generative: Create new data (X) - Both use similar underlying techniques

Types of Generative Models

Large Language Models (LLMs) - Generate human-like text - Trained on massive text corpora - Understanding and generation capabilities - Examples: GPT, Claude, LLaMA, PaLM

Diffusion Models - Image generation - Iterative denoising process - High-quality, detailed images - Examples: Stable Diffusion, DALL-E 2

Generative Adversarial Networks (GANs) - Generator creates fake data - Discriminator detects fakes - Adversarial training process - Applications: Image synthesis, style transfer

Variational Autoencoders (VAEs) - Encode data to latent space - Decode to generate new samples - Probabilistic framework - Applications: Image generation, anomaly detection

Foundation Models

Characteristics - Trained on broad data at scale - Adaptable to wide range of tasks - Large parameter count (billions) - Transfer learning capabilities

Capabilities - Few-shot learning: Learn from few examples - Zero-shot learning: Perform unseen tasks - In-context learning: Use examples in prompt - Emergent abilities: Unexpected capabilities at scale

Model Families - Text: GPT-4, Claude, PaLM, LLaMA - Vision: CLIP, DALL-E, Stable Diffusion - Multimodal: GPT-4V, Gemini, Claude 3 - Code: GitHub Copilot, CodeWhisperer, Codex

Model Training Approaches

Pre-training - Train on large, unlabeled dataset - Learn general language understanding - Self-supervised learning - Predict next word, mask filling

Fine-tuning - Adapt to specific task - Smaller, labeled dataset - Update model parameters - Task-specific performance improvement

Prompt Engineering - Design effective inputs - No model retraining - Few-shot examples - Task instructions

Retrieval Augmented Generation (RAG) - Combine model with external knowledge - Retrieve relevant information - Generate response with context - Up-to-date and accurate information

Natural Language Processing (NLP)

NLP Tasks

Text Classification - Assign categories to text - Sentiment analysis - Spam detection - Topic classification

Named Entity Recognition (NER) - Identify entities (person, organization, location) - Information extraction - Knowledge graph construction

Question Answering - Answer questions from text - Extractive: Select answer span - Generative: Generate answer

Text Summarization - Create concise summary - Extractive: Select key sentences - Abstractive: Generate new text

Machine Translation - Translate between languages - Neural machine translation - Attention mechanisms

Text Generation - Create coherent text - Story generation - Content creation - Code generation

Text Representation

Bag of Words - Count word occurrences - Ignore word order - Simple but loses context

TF-IDF - Term frequency-inverse document frequency - Weight words by importance - Less common words higher weight

Word Embeddings - Dense vector representations - Capture semantic meaning - Word2Vec, GloVe - Similar words have similar vectors

Contextual Embeddings - Embeddings depend on context - BERT, ELMo - Same word, different meanings - State of the art

Computer Vision

CV Tasks

Image Classification - Assign category to image - Single label or multi-label - Pre-trained models available

Object Detection - Locate and classify objects - Bounding boxes - YOLO, R-CNN, Faster R-CNN

Image Segmentation - Pixel-level classification - Semantic: Class per pixel - Instance: Separate object instances

Face Recognition - Identify individuals - Face detection + recognition - Security, authentication

Optical Character Recognition (OCR) - Extract text from images - Document digitization - Amazon Textract

Image Generation - Create new images - Text-to-image - Style transfer - Image editing

AWS AI/ML Services Overview

AI Services (Pre-trained)

  • Amazon Rekognition: Computer vision
  • Amazon Textract: Document analysis
  • Amazon Comprehend: NLP
  • Amazon Polly: Text-to-speech
  • Amazon Transcribe: Speech-to-text
  • Amazon Translate: Language translation
  • Amazon Lex: Conversational AI
  • Amazon Kendra: Intelligent search

ML Platform

  • Amazon SageMaker: End-to-end ML
  • SageMaker Studio: IDE for ML
  • SageMaker Autopilot: AutoML
  • SageMaker JumpStart: Pre-built models

Generative AI

  • Amazon Bedrock: Foundation models
  • Amazon Q: AI assistant
  • Amazon CodeWhisperer: Code generation

Best Practices

Model Development

  1. Start simple: Baseline model first
  2. Iterate quickly: Fast experimentation
  3. Measure everything: Track metrics
  4. Version control: Models and data
  5. Document assumptions: Technical debt management

Data Management

  1. Quality over quantity: Clean, relevant data
  2. Data versioning: Track changes
  3. Bias detection: Ensure fairness
  4. Privacy: Protect sensitive information
  5. Continuous collection: Keep data fresh

Production Deployment

  1. Monitoring: Track performance
  2. A/B testing: Compare models
  3. Gradual rollout: Minimize risk
  4. Rollback capability: Quick recovery
  5. Cost optimization: Right-size resources

Exam Tips

Key Concepts

  • Supervised vs unsupervised learning
  • Classification vs regression
  • Overfitting vs underfitting
  • Precision vs recall tradeoffs
  • Transfer learning benefits
  • Foundation model capabilities
  • RAG for current information
  • Prompt engineering basics

AWS Service Selection

  • Pre-trained AI services: Quick start, no ML expertise
  • SageMaker: Custom ML models
  • Bedrock: Generative AI, foundation models
  • Match service to use case and skill level

Common Patterns

  • Image analysis → Rekognition
  • Document extraction → Textract
  • Text analysis → Comprehend
  • Chatbots → Lex + Bedrock
  • Custom ML → SageMaker
  • Code generation → CodeWhisperer