Teach
AI models differ by what they do and when they run.
Model type overview
- Supervised models learn from labeled data (regression, classification).
- Unsupervised models find patterns in unlabeled data (clustering, dimensionality reduction).
- Deep learning models use neural networks for language, vision, and generative tasks.
- Foundation / large models are pre-trained broadly and adapted with prompting, retrieval, or fine-tuning.
Training vs inference
- Training is the learning phase: model parameters are optimized using historical data.
- Inference is the serving phase: trained model makes predictions on new inputs.
- Training is compute-heavy and periodic; inference is latency-sensitive and ongoing.
- Monitoring is essential in inference to detect drift, quality drop, and safety issues.
Practical Azure perspective
- Use prebuilt or hosted models when speed matters.
- Use custom training when domain-specific performance is required.
- Separate training and inference pipelines for reliability and cost control.
For exam and architecture decisions, always ask: What model type? Where does training happen? Where does inference happen?
Practice
Practice 1
A model trained with labeled examples belongs to:
Practice 2
Customer segmentation without labels is commonly:
Practice 3
When a deployed model predicts on new user input, that phase is:
Practice 4
Which phase is usually more latency-sensitive in production?
Practice 5
Foundation models are primarily known for: