AWS Certified AI Practitioner (AIF-C01) certification study notes, this guide will help you with quick revision before the exam. it can use as study notes for your preparation.
Accuracy: A common evaluation metric for classification models that measures how many predictions were correct out of all predictions made. It works well when classes are balanced.
Algorithm: A defined set of steps or rules that a machine learning model follows to learn patterns from data and make predictions.
Artificial Intelligence (AI): The broader field of creating systems capable of performing tasks that typically require human intelligence, such as reasoning, perception, and language understanding.
Machine Learning (ML): A subset of AI where systems automatically learn patterns from data without being explicitly programmed for each rule.
Deep Learning: A branch of ML that uses multi-layer neural networks to process complex data like images, audio, and text.
Generative AI: AI systems designed to create new content such as text, images, videos, music, or code instead of just predicting outcomes.
Foundation Model: Very large models trained on massive and diverse datasets that can be adapted to many downstream tasks through prompting or fine-tuning.
Large Language Model (LLM): A type of foundation model trained on large volumes of text to understand and generate human-like language.
Small Language Model (SLM): A compact language model optimized for efficiency, often used on edge devices where compute and memory are limited.
Multi-modal Model: A model capable of understanding and processing multiple data types (text, image, audio) together.
Neural Network: A computational model made of interconnected layers that transform input data into predictions, inspired by the human brain.
Learning Types
Supervised Learning: The model is trained using labeled data, learning a direct mapping between inputs and correct outputs.
Unsupervised Learning: The model analyzes unlabeled data to find hidden structures or patterns without predefined answers.
Semi-Supervised Learning: Uses a small amount of labeled data along with a large amount of unlabeled data to improve learning efficiency.
Self-Supervised Learning: A technique where the model creates its own labels from raw data, reducing the need for human annotation.
Reinforcement Learning (RL): The model (agent) learns by interacting with an environment and receiving rewards or penalties.
Reinforcement Learning from Human Feedback (RLHF): A refinement of RL where humans rank or score model outputs to better align them with human expectations.
Model Training and Optimization
Pre-trained Model: A model already trained on large, general-purpose datasets and reused for specific tasks.
Fine-tuning: Further training a pre-trained model on a smaller, task-specific dataset to improve performance.
Instruction-Based Fine-Tuning: Uses labeled prompt–response pairs to train models to follow instructions better.
Continued Pre-training: Trains a model further using domain-specific unlabeled data to gain deeper subject knowledge.
Transfer Learning: Reusing knowledge learned from one task and applying it to a related task.
Hyperparameters: Configurable values that control how a model learns, such as learning rate and batch size.
Learning Rate: Controls how much the model updates its weights during each training step.
Batch Size: Number of training samples processed together before updating model weights.
Epoch: One complete iteration through the entire training dataset.
Regularization: Techniques used to reduce overfitting by limiting model complexity.
Model Evaluation Metrics
Precision: Measures how many positive predictions were actually correct; important when false positives are costly.
Recall: Measures how many actual positive cases were correctly identified; important when missing positives is risky.
F1 Score: Combines precision and recall into a single metric, useful for imbalanced datasets.
Confusion Matrix: A table that shows true positives, false positives, true negatives, and false negatives.
MAE (Mean Absolute Error): Measures average absolute difference between predicted and actual values.
RMSE (Root Mean Squared Error): Similar to MAE but punishes bigger errors more harshly.
R² (R-Squared): Shows how much of the variation in results your model can explain - closer to 1 is better.
Perplexity: Measures how surprised a language model is by text - lower is better because it means more confident predictions.
BLEU (Bilingual Evaluation Understudy): An automated metric used to evaluate the quality of text that has been machine-translated from one language to another by comparing it to high-quality human translations.
ROUGE (Recall-Oriented Understudy for Gisting Evaluation): Measures how good a summary is by comparing it to human-written summaries.
Model Behavior and Risks
Overfitting: When a model memorizes the training data too well and can’t handle new examples it hasn’t seen.
Underfitting: When a model is too simple and can’t learn the patterns in the data properly.
Bias: A systematic error in an AI system that results in unfair or discriminatory outcomes, often stemming from unrepresentative training data.
Fairness: The ethical goal of ensuring that an AI model’s predictions do not create discriminatory or unjust outcomes for different demographic groups.
Hallucination: A phenomenon where a generative AI model produces text that is plausible and well-written but is factually incorrect, invented, or nonsensical.
Explainability: Being able to explain why a model made a certain decision in simple terms people can understand.
Interpretability: How easily humans can understand a model’s internal decision process.
Opaque Model (Black-box): A complex model, such as a deep neural network, whose internal decision-making process is not easily understood by humans.
Transparent Model (White-box): A model, such as a decision tree, whose internal decision-making process is easily understood by humans.
Nondeterminism: A characteristic of generative AI models where asking the same prompt multiple times can produce different, non-identical responses.
Toxicity: Harmful, offensive, or inappropriate content in AI outputs.
Jailbreaking: Attempts to bypass AI safety mechanisms.
Prompt Injection: Malicious instructions embedded in prompts to override system behavior.
Prompt Poisoning: Injecting harmful data into prompts or training data to influence outputs.
IP Infringement: Risk of generating outputs that resemble copyrighted material.
Prompting Techniques
Prompt Engineering: Designing effective prompts to guide model outputs.
Zero-shot Prompting: Asking the model to perform a task without examples.
Single-shot Prompting: Providing one example in the prompt.
Few-shot Prompting: Providing multiple examples to guide responses.
System Prompt: Initial instructions that set how the AI should behave throughout the entire conversation.
Data Concepts
Structured Data: Organized data in neat rows and columns like spreadsheets or databases.
Unstructured Data: Data that does not have a predefined organizational structure, such as the text of an email, an audio file, a video, or an image.
Labeled Data: Data where each example is tagged with the correct answer or “label.” It is the required input for supervised learning.
Unlabeled Data: Raw data with no predefined answers or labels. It is the required input for unsupervised learning.
Feature Engineering: The process of selecting, transforming, and creating input variables (“features”) from raw data to improve the performance of a machine learning model.
Exploratory Data Analysis (EDA): The initial process of analyzing a dataset to understand its main characteristics, discover patterns, and spot anomalies.
Data Drift: When the real-world data starts looking different from what the model was trained on, making it less accurate.
Model Drift: When a model becomes less accurate over time because the world has changed since it was trained.
Data Lineage: The practice of tracking the origin, movement, and transformation of data throughout its lifecycle. Crucial for reproducibility and auditing.
Data Residency: The requirement that certain data must be physically stored within a specific geographical location to comply with local laws and regulations.
Data Retention: A policy that defines how long data must be kept for legal, regulatory, or business reasons.
Encryption: The process of encoding data to protect it from unauthorized access. Can be “at rest” (when stored) or “in transit” (when moving over a network).
Inference and Deployment
Inference: Using a trained model to make predictions.
Real-time Inference: Low-latency predictions for live applications.
Batch Inference: Run predictions on a huge pile of data all at once when you don’t need instant results.
Asynchronous Inference: Run predictions that take a long time on big files, and get the results back when they’re ready.
Serverless Inference: Only pay when predictions are running, scales to zero when idle - great for unpredictable traffic.
Edge Inference: Running AI models directly on devices like phones or cameras instead of sending data to the cloud..
Latency: How long it takes to get a response back after you send a request.
Retrieval and Vector Search
Embedding: Converting text or images into numbers that capture their meaning, so similar things have similar numbers.
Vector Database: Special database built for storing embeddings and finding similar items super fast.
Vector Search (Similarity Search): Finding items that are most similar to what you’re looking for based on their embeddings.
RAG (Retrieval Augmented Generation): Give the model access to external information to look up facts before answering, reducing hallucinations.
pgvector: Add-on for PostgreSQL databases that lets you store and search through embeddings.
AWS AI and ML Services
Amazon Bedrock: A fully managed AWS service that provides access to a variety of high-performing Foundation Models from leading AI companies via a single API, simplifying the development of generative AI applications.
Bedrock Agents: Smart assistants in Bedrock that can figure out complex tasks and execute them automatically.
Guardrails for Bedrock: A safety feature that allows you to implement policies to control the types of content your generative AI application will generate, by defining denied topics and content filters.
Amazon SageMaker: Complete toolkit for building, training, and running your own custom machine learning models.
SageMaker Studio: All-in-one workspace where you do everything related to machine learning.
SageMaker Canvas: Build ML models without writing any code using a visual drag-and-drop interface.
SageMaker JumpStart: A feature of SageMaker that provides access to a wide range of publicly available, open-source Foundation Models, offering one-click deployment for fine-tuning and inference.
SageMaker Pipelines: A service to build, automate, and manage end-to-end ML workflows (CI/CD for ML).
SageMaker Model Monitor: A feature of SageMaker that automatically detects data drift and model quality degradation in deployed models.
SageMaker Clarify: A feature of SageMaker that helps improve ML models by detecting potential bias in data and explaining how models make predictions.
SageMaker Ground Truth: Get humans to label your data or review model outputs to make them better.
SageMaker Model Cards: A feature of SageMaker that provides a standardized way to document critical information about an ML model, acting as a “nutrition label” for transparency and governance.
Security, Governance and Pricing
IAM (Identity and Access Management): The foundational AWS service for securely controlling access to AWS services and resources by managing users, groups, and permissions.
AWS CloudTrail: Records every action taken in your AWS account so you can see who did what and when.
AWS Config: Keeps track of how your AWS resources are set up and checks if they meet your rules.
AWS Artifact: A service that provides on-demand access to AWS’s security and compliance reports, such as ISO certifications and SOC reports.
AWS Audit Manager: A service that helps you continuously audit your AWS usage to simplify how you assess risk and compliance with regulations and industry standards.
AWS PrivateLink: A networking service that provides private, secure connectivity between your VPCs and AWS services, without exposing your traffic to the public internet.
Shared Responsibility Model: The AWS security framework where AWS is responsible for the security OF the cloud (infrastructure, managed services), and the customer is responsible for security IN the cloud (data, access configuration, responsible AI).
On-Demand Pricing: Pay-per-use pricing model.
Provisioned Throughput: A pricing model for AI services where you purchase dedicated capacity for a fixed price to ensure consistent performance and potentially lower costs for high-volume, stable workloads.
Batch Pricing: Lower-cost processing with delayed results.