AIF-C01 Domain 1 - Fundamentals of AI and ML
Domain 1 is 20% of the scored AWS Certified AI Practitioner (AIF-C01) exam. It tests vocabulary, use-case selection, the AI/ML lifecycle, and evaluation at a conceptual level.
AI family tree
- Artificial intelligence (AI): systems that perform tasks associated with human intelligence, including rules-based and learned systems.
- Machine Learning: a subset of AI that learns patterns from data to make predictions or decisions rather than relying only on explicit rules.
- Deep Learning ML based on multi-layer neural networks; particularly effective with large, unstructured datasets such as text, images, audio, and video.
- Generative AI: models that create new content. Modern generative AI usually uses deep-learning foundation models, but not every deep-learning model is generative.
- Agentic AI: AI systems that pursue a goal by reasoning, planning, calling tools, using memory, and coordinating multi-step workflows.
- Foundation model (FM): a large pretrained model that can be adapted to many downstream tasks.
- Large Language Model (LLM): a foundation model specialized in processing and generating language.
An algorithm is the learning procedure; a model is the learned artifact. Training estimates model parameters from data. Inference applies the trained model to new input.
Learning methods
| Method | Training signal | Typical outcome | Example |
|---|---|---|---|
| Supervised Machine Learning | Labeled examples | Predict a class or numeric value | spam classification, price regression |
| Unsupervised Machine Learning | Unlabeled data | Find structure or groups | customer clustering, anomaly discovery |
| Reinforcement Learning (RL) | Rewards and penalties from interaction | Learn a policy that maximizes reward | robotics, game play, preference optimization |
- Classification predicts a category.
- Regression predicts a continuous number.
- Clustering groups similar observations without predefined labels.
Data vocabulary
- Labeled data: input paired with the desired answer; used for supervised learning and many fine-tuning tasks.
- Unlabeled data: input without target answers; used for unsupervised learning and large-scale pretraining.
- Structured data: rows/columns with a defined schema, such as relational tables.
- Semi-structured data: flexible structure or tags, such as JSON and application logs.
- Unstructured data: text, images, audio, video, and documents.
- Tabular data: records expressed as features in columns.
- Time-series data: observations ordered by time.
- Training, validation, and test sets: fit weights, tune/select approaches, and estimate final generalization. Prevent leakage between them.
Inference patterns
- Real-time inference: synchronous, low-latency prediction for interactive or sustained traffic.
- Batch inference: process a large dataset together when immediate results are unnecessary.
- Asynchronous inference: queue long-running requests or large payloads and retrieve results later.
- Serverless inference: on-demand managed endpoints that scale for intermittent traffic, trading some latency/control for lower operational effort.
Choose the right kind of solution
- Use deterministic software or rules when the result must be exact, the logic is stable and fully known, or mistakes are unacceptable.
- Use traditional ML for structured predictive tasks where explainability, low latency, smaller datasets, or lower cost matter.
- Use an FM for language-rich, multimodal, content-generation, or broad cross-domain tasks.
- Use a managed AI service when the problem directly matches a ready-made capability and custom training is unnecessary.
- Confirm that expected value exceeds data, development, inference, monitoring, and compliance cost.
Managed AI use-case matching
| Need | Service |
|---|---|
| Speech to text | Amazon Transcribe |
| Text to speech | Amazon Polly |
| Language translation | Amazon Translate |
| Sentiment, entities, key phrases, PII in text | Amazon Comprehend |
| Conversational bot interface | Amazon Lex |
| OCR and structured document extraction | Amazon Textract |
| Image/video labels, faces, moderation | Amazon Rekognition |
| Recommendations | Amazon Personalize |
| Enterprise search/retrieval | Amazon Kendra |
| Custom model lifecycle | Amazon Sage Maker |
| Foundation-model application | AWS Bedrock |
AI/ML lifecycle
- Define the business problem, constraints, success metric, and whether AI is appropriate.
- Collect and govern representative data.
- Explore, clean, transform, and label data; prevent leakage.
- Engineer/select useful features for traditional ML.
- Select a source: managed AI API, pretrained/open model, or custom model.
- Train and tune, then evaluate on held-out data.
- Deploy through a managed endpoint/API, asynchronous/batch job, or self-hosted service.
- Monitor quality, latency, cost, bias, drift, and failures; retrain or roll back when needed.
Relevant tools include Amazon S3 for data/artifacts, AWS Glue and AWS Glue DataBrew for preparation, SageMaker Data Wrangler for preparation, SageMaker Feature Store for reusable features, SageMaker Model Monitor for production monitoring, and Amazon CloudWatch for operational observability.
MLOps
MLOps applies repeatable, versioned, automated operational practices to ML. Key ideas are experiment tracking, reproducible pipelines, scalable infrastructure, model/data versioning, approval and deployment controls, monitoring, rollback, and retraining. It reduces one-off work and technical debt; it does not eliminate the need for human ownership.
Evaluation metrics
For a binary classifier:
- Accuracy:
(TP + TN) / all; useful when classes and error costs are balanced. - Precision:
TP / (TP + FP); prioritize when false positives are costly. - Recall:
TP / (TP + FN); prioritize when false negatives are costly. - F1: harmonic mean of precision and recall; useful when both matter, especially with class imbalance.
Also evaluate business results: ROI, cost per user/interaction, development cost, task completion, customer feedback, latency, and whether the model improves the target outcome. A technically strong metric is not enough if business value is absent.
Common traps
- AI is broader than ML; ML is broader than deep learning.
- Inference is model use, not model training.
- A prediction is probabilistic; use rules when an exact guaranteed outcome is required.
- High accuracy can hide poor minority-class performance.
- Human review during inference is Amazon Augmented AI (A2I); human labeling of training data is SageMaker Ground Truth.