LEARN. PRACTICE. GROW.Concept and mechanism
Start with the problem: predicting batch duration is regression; assigning a ticket category is classification; discovering groups without predefined labels is clustering. In supervised learning, examples include known outcomes. Unsupervised learning seeks structure without those labels; reinforcement learning learns through rewards associated with actions. Training adjusts a model using data; inference applies the model to new requests. AI is the broader field, ML learns patterns, and deep learning uses neural networks with multiple layers. A foundation model can support several tasks but still needs suitable data and evaluation. Deterministic rules remain useful when criteria are exact and stable. Define a simple baseline before comparing AI benefits.
Guided application
In an operations exercise, only 20 of 1,000 tickets need urgent escalation. A classifier that never flags urgency is correct 980 times but misses every important case. Accuracy alone hides this failure. Precision compares correct alerts with all emitted alerts; recall compares detected urgent cases with all genuinely urgent cases. Choose a trade-off considering false positives, missed cases, and human capacity. Separate training and evaluation data and avoid information available only after the outcome. After deployment, monitor data changes and performance. An overnight report can tolerate batch inference; an in-session decision needs suitable latency. SageMaker AI supports the ML lifecycle; Bedrock provides capabilities for applications using foundation models. Choose according to the task and required control, rather than the newest product name.
980 correct classifications out of 1,000 tickets can coexist with zero urgent cases detected.
Common pitfalls
Accuracy without class analysis; post-outcome data; complex models without a baseline; training confused with inference.
Related topics: Generative AI, context, and agents · Foundation models: RAG, prompts, and evaluation · Responsible AI and explainability
Connect the task, data, and metrics to the operational impact of errors.
Reference: AIF-C01 domain1: Fundamentals of AI and ML · AIF-C01