AI LEARNING / 02 / FOUNDATION TO INTERMEDIATE
Machine learning
& data
Understand the data, assumptions and evaluation behind a useful prediction.
THE IDEA TO KEEP IN VIEW
A model’s performance is meaningful only in relation to the data, validation and decision behind it.
SUGGESTED READING ORDER
Build one idea
at a time.
Start with the meaning of the inputs. Then examine how models estimate relationships and how their errors should be judged in the task where they will be used.
AFTER READING
Choose a prediction task and explain why its evaluation metric fits the consequence of a wrong result.
01 / DATA & CONTEXT
Give raw data its operating context
See why timestamps, process history and product or machine context change the meaning of a measurement.
02 / AN INTERPRETABLE MODEL
Start with linear regression
Interpret coefficients, residuals and validation while distinguishing observed association from a causal claim.
03 / EVALUATION
Choose metrics that match the decision
Compare model fit and error measures, then connect the choice of metric with operating tolerances and the cost of mistakes.
Check your understanding
Would the same average error be acceptable for every production decision? Consider operating tolerance, rare large errors, data shifts and the difference between evaluation data and future use.
FURTHER EXPLANATIONS
Deepen a specific idea.
Continue through the wider collection after the essential readings.
How a Prediction Is Decomposed Into Baseline and Feature Contributions
Understand SHAP values through an industrial risk example: baseline output, feature contributions, background data, output scale, and limits on causal interpretation.
How Industrial AI Preserves Entity Type, Relationship Meaning, and Event Time
Represent machines, orders, materials, and events with heterogeneous temporal graphs that preserve relationship meaning, event time, and prediction-time boundaries.
How AI Learns From Connected Machines, Orders, Materials, and Events
Learn how graph neural networks use message passing to connect manufacturing context, with practical checks for graph schemas, data leakage, and provenance.
Why the Objective Determines the Behavior an Agent Learns
Design industrial AI reward functions with explicit throughput, delay, energy, and quality trade-offs, then test weights, time horizons, shaping, and constraints.
How Agents Learn Alternatives Without Experimenting Unsafely
Understand exploration vs. exploitation for industrial AI, including epsilon-greedy policies, simulations, offline evidence, action limits, and controlled pilots.
How AI Agents Improve Action Values From Experience
Follow a Q-learning update step by step, then examine reward design, experience coverage, action constraints, and controlled deployment for manufacturing agents.