AI.info
How machines learn
Twenty-six lessons on what actually changes with experience — from a real problem to a learning task you can defend.
- Learning in AI: What Changes With Experience?
- From a Real Problem to a Learning Task
- Examples, Observations, and the Unit of Experience
- Labels, Targets, and the Limits of Ground Truth
- Features, Representations, and What the Model Can See
- Parameters, Hyperparameters, and Model Capacity
- Building a Dataset That Represents the Job
- Training, Validation, and Test: Three Different Jobs
- Data Leakage: When the Answer Sneaks Into the Question
- The Training Loop: Predict, Measure, Update
- Loss Functions: Designing the Feedback Signal
- Gradients, Batches, Epochs, and Learning Rate
- Baselines and Sanity Checks
- Underfitting: When the Model Cannot Learn Enough
- Overfitting: When the Model Learns the Dataset Too Closely
- Generalization and Inductive Bias
- Regularization, Early Stopping, and Capacity Control
- Metrics, Thresholds, and the Cost of Different Errors
- Error Analysis: Learn From the Examples the Model Misses
- Data Quality, Class Imbalance, and Label Noise
- Learning Curves and When More Data Helps
- Distribution Shift: When the World Stops Matching the Dataset
- Feedback Loops and Selective Labels
- Experiments, Randomness, and Reproducibility
- A Diagnosis-First Strategy for Improving Models
- Capstone: Design an Honest Learning Experiment