Fundamentals of AI and ML
What this domain tests
You must separate AI, machine learning, and deep learning; choose a learning paradigm from a scenario; and read evaluation metrics correctly. Service selection (custom SageMaker builds vs. prebuilt AI services) appears constantly.
Core rules
- AI ⊃ ML ⊃ deep learning. Deep learning uses multi-layer neural networks and suits unstructured data (images, audio, text); classic ML suits tabular data.
- Supervised learning needs labeled data: classification predicts a discrete label, regression a continuous value. Unsupervised finds structure without labels (clustering, anomaly detection). Reinforcement learning learns from reward signals by trial and error.
- Training fits parameters; inference serves predictions. Overfitting means strong training scores and weak generalization; counter it with more data, regularization, or a simpler model.
- SageMaker covers the build path: Canvas (no-code), Data Wrangler (prep), Autopilot (AutoML), Experiments, Model Registry, Pipelines. Prebuilt services (Rekognition, Comprehend, Transcribe, Forecast, Personalize, Fraud Detector) are the answer when no custom model is needed.
Worked example
A fraud detector scores 1,000 transactions; 50 are fraud. It flags 100, of which 40 are real fraud. Precision = 40/100 = 0.40. Recall = 40/50 = 0.80. F1 = 2 × (0.40 × 0.80)/(0.40 + 0.80) ≈ 0.53. Accuracy = 980/1,000 = 98%, which hides 10 missed frauds. Because a missed fraud costs more than a false alarm, lower the decision threshold to favor recall.
Traps
- Quoting accuracy on heavily imbalanced data.
- Assuming unsupervised learning needs labels.
- Choosing SageMaker for a task Comprehend already solves.
- Confusing model evaluation with production monitoring.