Skip to content

Task Statement 1.1: Explain Basic AI Concepts and Terminologies

AWS AIF-C01 Exam Focus: Task 1.1 makes up a primary portion of Domain 1 (Fundamentals of AI and ML - 20% of exam). It tests core vocabulary, paradigm boundaries, inferencing delivery mechanisms, data types, and machine learning methods.

1. Core AI/ML Definitions & Glossary

Foundational Concepts

  • Artificial Intelligence (AI): The broad field of computer science focused on building hardware and software systems capable of executing tasks that typically require human cognition (e.g., visual perception, speech comprehension, reasoning, and decision-making).
  • Machine Learning (ML): A sub-field of AI where algorithms analyze historical data, learn underlying mathematical patterns, and improve prediction performance over time without explicit step-by-step programming.
  • Deep Learning (DL): A specialized subset of ML based on Artificial Neural Networks (ANNs) with multiple hidden layers. It excels at automatically extracting features from high-dimensional unstructured data (images, audio, video, raw text).
  • Neural Networks (ANNs): Computational models inspired by biological neural structures, organized into interconnected layers of nodes (input, hidden, and output) with learnable weights, biases, and activation functions.
  • Algorithm: The mathematical formula, loss function, or optimization procedure used to process training data and update model weights (e.g., Linear Regression, XGBoost, K-Means, Adam Optimizer).
  • Model: The final mathematical artifact produced after an algorithm completes training on a dataset. It consists of learned weights and parameters used to evaluate new, unseen inputs.
  • Training: The compute-intensive development phase where an algorithm iteratively processes training data, measures prediction errors against ground truth (loss), and adjusts internal weights to minimize error.
  • Inferencing: The operational production phase where a fully trained model processes new input data to make predictions, output classification scores, or generate synthetic content.

Specialized Sub-Fields

  • Computer Vision (CV): Enabling software to extract semantic meaning, spatial boundaries, and visual context from digital images, video frames, and visual sensor streams (e.g., object detection, facial recognition, OCR).
  • Natural Language Processing (NLP): Enabling software to parse, comprehend, translate, summarize, and generate human written text or spoken language.

Model Governance & Fitting Terms

  • Bias (Algorithmic / Data): Systematic error or skewed predictions produced by a model, typically originating from unrepresentative, incomplete, or historically prejudiced training data.
  • Fairness: The operational and ethical framework ensuring AI systems treat demographic subgroups equitably and without discriminatory bias.
  • Model Fit:
  • Underfitting: Occurs when a model is too simple to capture the underlying relationships in the data (high bias). Performs poorly on both training and validation datasets.
  • Overfitting: Occurs when a model memorizes training data, including noise and random outliers (high variance). Achieves near-perfect training accuracy but fails to generalize to new data.
  • Large Language Model (LLM): A specialized deep learning transformer model containing billions or trillions of parameters, trained on broad text datasets to execute natural language tasks.
  • Generative AI (GenAI): A branch of AI powered by foundation models (FMs) capable of generating new synthetic content (text, images, audio, video, code) based on user prompts.
  • Agentic AI: Advanced GenAI systems operating with multi-step autonomy. They maintain state memory, break goals into sub-tasks, execute external tool calls (APIs, web searches, SQL queries), and dynamically adapt their plan based on intermediate observations.

2. Comparative Analysis of AI Paradigms

Dimension Traditional Machine Learning Deep Learning Generative AI Agentic AI
Primary Input Data Structured / Tabular (CSV, SQL) Unstructured (Images, Audio, Text) Prompts + Unstructured Multi-modal inputs Goal prompts + System tool schemas / API definitions
Output Type Discrete classifications, probabilities, forecasts Feature extractions, object bounding boxes, classifications New synthetic artifacts (text, image, audio, code) Autonomous multi-step actions, tool calls, workflow execution
Feature Engineering Manual (requires human domain experts) Automatic (learned across multiple neural layers) Automatic (learned across transformer attention layers) Automatic (contextual reasoning + tool selection)
Autonomy Level Low (Executes single mathematical function) Low-to-Moderate (Executes specific task) Moderate (Generates completion for a prompt) High (Self-directed planning, tool calling, loop execution)
Primary AWS Services Amazon SageMaker AI, SageMaker Canvas SageMaker AI (PyTorch / TensorFlow) Amazon Bedrock, SageMaker JumpStart Agents for Amazon Bedrock, Bedrock AgentCore

Key Architectural Nuances

  • Hierarchical Nesting: Machine Learning $\subset$ AI; Deep Learning $\subset$ Machine Learning; Generative AI is built predominantly on Deep Learning (Transformer) architectures.
  • Discriminative vs. Generative:
  • Discriminative Models (Traditional ML / CV / NLP): Estimate $P(Y|X)$ — learn boundaries to classify or predict an existing label (e.g., "Is this email spam?").
  • Generative Models (GenAI): Estimate $P(X, Y)$ — learn the underlying probability distribution of data to produce entirely new instances (e.g., "Draft a personalized email response").
  • GenAI vs. Agentic AI: Standard GenAI is a reactive single-turn engine (Prompt $\rightarrow$ Completion). Agentic AI uses GenAI as an internal reasoning engine to execute iterative loops: Plan $\rightarrow$ Tool Call $\rightarrow$ Observe Result $\rightarrow$ Re-evaluate $\rightarrow$ Final Action.

3. Inferencing Delivery Types

1. Real-Time (Synchronous) Inferencing

  • Mechanism: Immediate request-response cycle over an open HTTP/HTTPS connection.
  • Latency: Low (milliseconds to seconds).
  • Best Use Cases: E-commerce fraud checks at payment checkout, interactive chatbots, live recommendation engines.
  • AWS Services: Amazon SageMaker Real-Time Endpoints, Amazon Bedrock On-Demand Endpoints.

2. Asynchronous Inferencing

  • Mechanism: Requests are placed in an internal managed queue (e.g., Amazon SQS). The client receives a job ID immediately and is notified via Amazon SNS or S3 upload when processing completes.
  • Latency: Medium to High (seconds to minutes).
  • Best Use Cases: Large payloads (up to 1 GB), long model processing times (up to 15 minutes), high traffic spikes requiring queue buffering. Supports auto-scaling instances down to zero when idle.
  • AWS Services: Amazon SageMaker Asynchronous Inference.

3. Batch Inferencing (Batch Transform)

  • Mechanism: Offline processing of large pre-stored datasets in Amazon S3. Compute instances spin up, process the dataset in batch mode, write results back to S3, and immediately terminate.
  • Latency: High (hours or scheduled execution).
  • Best Use Cases: Nightly risk score updates, weekly customer churn predictions, batch document OCR processing across historical archives.
  • AWS Services: Amazon SageMaker Batch Transform, Amazon Bedrock Batch Inference.

4. Serverless Inferencing

  • Mechanism: On-demand execution managed entirely by AWS. Infrastructure scales up instantly upon receiving requests and scales down to zero when idle.
  • Latency: Low-to-Medium (subject to minor cold starts when initializing from zero).
  • Best Use Cases: Workloads with intermittent, unpredictable, or infrequent traffic patterns (e.g., an internal HR portal tool used occasionally during business hours).
  • AWS Services: Amazon SageMaker Serverless Inference.

Inferencing Comparison Matrix

Inferencing Type Latency Target Max Payload Limit Always-On Instance? Scale to Zero? Billing Model
Real-Time Milliseconds 6 MB Yes No (requires min instance count) Hourly instance runtime
Asynchronous Seconds to Minutes 1 GB Optional Yes Instance uptime + Queue processing
Batch Hours / Scheduled Terabytes (S3) No (Ephemeral) Yes (Terminates on job end) Job duration & dataset volume
Serverless Milliseconds to Seconds 6 MB No Yes Compute duration (ms) & Memory allocated

4. Data Types & Classifications in AI Systems

Data Labeling Classification

  • Labeled Data: Datasets where each record includes feature inputs ($X$) and ground-truth target tags ($Y$). Required for Supervised Learning (e.g., Medical images tagged as "Malignant" or "Benign").
  • Unlabeled Data: Raw datasets containing only input features ($X$) without target output tags. Used for discovering hidden clusters, detecting anomalies, or pre-training foundation models.

Data Structure Formats

  • Structured Data: Highly organized data conforming to a fixed tabular schema (e.g., relational database tables, CSVs). Processed efficiently by traditional algorithms (XGBoost, Random Forests).
  • Unstructured Data: Data lacking a pre-defined schema (e.g., freeform text, raw photos, audio files, MP4 video). Represents ~80% of enterprise data and is processed primarily by Deep Learning models and FMs.
  • Semi-Structured Data: Data containing organizational markers, key-value pairs, or XML/JSON tags without a rigid relational schema.

Specialized Modalities

  • Tabular Data: Structured numerical and categorical rows and columns. (AWS tools: SageMaker Canvas, SageMaker Autopilot).
  • Time-Series Data: Sequence of data points indexed in successive, uniform temporal order (e.g., hourly server CPU metrics, daily stock prices). (AWS tool: Amazon Forecast).
  • Text Data: Natural language sequences used for NLP, sentiment analysis, and LLM training.
  • Image / Video Data: Multi-dimensional pixel arrays processed via Computer Vision architectures.

5. AI/ML Learning Methods

1. Supervised Learning

  • Definition: The model is trained on labeled data ($X, Y$) to learn a mathematical function mapping inputs to known outputs.
  • Sub-Types:
  • Classification: Predicting a discrete categorical label (e.g., "Spam vs. Not Spam", "Loan Approved vs. Denied").
  • Regression: Predicting a continuous numerical value (e.g., "Predicting quarterly revenue", "Estimating house valuation").
  • Common Algorithms: Linear/Logistic Regression, Decision Trees, Random Forests, XGBoost, Support Vector Machines (SVM).

2. Unsupervised Learning

  • Definition: The model is trained on unlabeled data ($X$) to discover implicit patterns, groupings, or structural anomalies without human target labels.
  • Sub-Types:
  • Clustering: Partitioning data into groups with shared characteristics (e.g., K-Means clustering for customer demographic segmentation).
  • Anomaly Detection: Identifying rare data points that deviate significantly from standard baseline patterns (e.g., credit card fraud detection, network intrusion monitoring).
  • Dimensionality Reduction: Reducing the number of feature variables while retaining core variance (e.g., Principal Component Analysis [PCA]).

3. Reinforcement Learning (RL)

  • Definition: An autonomous Agent learns how to achieve a goal by executing Actions within an Environment, receiving feedback in the form of numerical Rewards or Penalties to learn an optimal decision strategy through trial-and-error.
  • Key Components: Agent, Environment, Action, Reward Signal, Policy.
  • AWS Exam Anchor: AWS DeepRacer (1/18th scale autonomous race car trained via Reinforcement Learning).

6. Exam Decision Rules & Common Pitfalls

High-Yield Decision Rules

  1. Select Generative AI / Foundation Models whenever a prompt requires producing new original synthetic content (text, images, audio, code).
  2. Select Agentic AI whenever a prompt requires autonomous multi-step execution, dynamic planning, or tool integration (e.g., checking inventory, calling an internal API, sending an email).
  3. Select Supervised Learning whenever historical data contains known, verified ground-truth labels or historical target outcomes.
  4. Select Unsupervised Learning (Clustering) whenever the business wants to group un-categorized customers or discover hidden patterns in raw data.
  5. Select Serverless Inferencing when workload traffic is intermittent, unpredictable, and drops to zero for extended periods.
  6. Select Batch Inferencing when processing millions of records offline in S3 on a scheduled basis where immediate real-time response is not required.
  7. Select Asynchronous Inferencing when processing large payload files (up to 1 GB) or long-running tasks (up to 15 mins) that would cause a real-time HTTP endpoint to time out.

Common Exam Traps

  • Trap 1: Overfitting vs. Underfitting Fixes
  • Overfitting = High training accuracy, low validation accuracy (memorized noise). Remedies: Gather more training data, apply regularization, or simplify model.
  • Underfitting = Low training accuracy, low validation accuracy (model too simple). Remedies: Add more features, train longer, or use a more complex algorithm.
  • Trap 2: Real-Time vs. Batch Misallocation
  • Do NOT choose Real-Time endpoints for offline scheduled reporting jobs — choose Batch Transform to minimize costs.
  • Trap 3: Asynchronous vs. Real-Time Payload Limits
  • Real-Time endpoints cap payloads at 6 MB. If a scenario mentions analyzing 500 MB video files near real-time, select Asynchronous Inference.