ML Fundamentals

For ownersClassic prediction, such as scoring leads or forecasting a busy week. Often simpler and cheaper than a chat model.

Classical machine learning is still the backbone of most production AI.

21 lessons

Start with lesson 1

Lesson 1 of 21, ~45 minutesWhat Is Machine Learning

The lessons

  1. 01What Is Machine LearningMachine learning means a computer learns patterns from examples instead of following rules a person wrote. Show it past data with answers, and it learns to guess answers for new data.Learn · ~45 minutes
  2. 02Linear RegressionLinear regression fits the best straight line through data to predict a number, such as price from size. It is simple, fast and easy to explain.Build · ~90 minutes
  3. 03Logistic RegressionLogistic regression predicts the chance of a yes or no outcome, such as whether a customer will buy, by squeezing a straight-line score into a probability between 0 and 1.Build · ~90 minutes
  4. 04Decision Trees and Random ForestsA decision tree is a flowchart of yes or no questions learned from data. A random forest combines many trees trained on different samples and votes, which is far more accurate.Build · ~90 minutes
  5. 05Support Vector MachinesA support vector machine finds the widest possible gap between two groups of points. With kernels it can also draw curved boundaries for data that a straight line cannot split.Build · ~90 minutes
  6. 06K-Nearest Neighbors and DistancesK-nearest neighbors stores all past examples and labels a new one by looking at the k most similar stored examples. The choice of distance decides what similar means.Build · ~90 minutes
  7. 07Unsupervised LearningUnsupervised learning finds structure in data that has no labels. Clustering groups similar items together, and methods like k-means and DBSCAN do this automatically.Build · ~90 minutes
  8. 08Feature Engineering & SelectionFeature engineering turns raw data into inputs a model can use well, such as days since last order, average ticket size or text turned into word counts.Build · ~90 minutes
  9. 09Model EvaluationModel evaluation measures how well a model works on data it has not seen, using held-out sets, cross-validation and the right metrics such as accuracy, precision, recall or error size.Build · ~90 minutes
  10. 10Bias-Variance TradeoffModel errors come from bias (too simple to see the pattern), variance (so flexible it memorizes noise) and plain noise. Good models balance the first two.Learn · ~75 minutes
  11. 11Ensemble MethodsEnsembles combine many models so their mistakes partly cancel. Bagging averages models trained on random samples, boosting trains models in sequence to fix earlier errors, and stacking blends them.Build · ~120 minutes
  12. 12Hyperparameter TuningHyperparameters are settings chosen before training, like tree depth or learning rate. Tuning searches for good values using grid search, random search or smarter Bayesian methods.Build · ~90 minutes
  13. 13ML PipelinesAn ML pipeline chains every step from raw data to prediction: cleaning, feature building, training and scoring, so the exact same steps run in testing and in real use.Build · ~120 minutes
  14. 14Naive BayesNaive Bayes predicts a category by multiplying the odds of each feature, assuming features do not affect each other. The assumption is wrong, yet it works well for text like spam filtering.Build · ~75 minutes
  15. 15Time Series FundamentalsTime series data is measured in order over time, like daily sales. Forecasting it means handling trend, seasonality and stationarity, and never testing on dates before the training data.Build · ~90 minutes
  16. 16Anomaly DetectionAnomaly detection learns what normal looks like and flags anything that does not fit, using methods like z-scores, isolation forests or distance to typical points.Build · ~75 minutes
  17. 17Handling Imbalanced DataImbalanced data means one outcome is rare, like fraud among normal orders. A model can score high accuracy by always guessing the common case, so special metrics and fixes are needed.Build · ~90 minutes
  18. 18Feature SelectionFeature selection keeps only the inputs that help a model and drops the rest, using filter scores, wrapper searches, or models that rank feature importance.Build · ~75 minutes
  19. 19ML System Design — From Problem to Deployed ModelML system design plans the whole system around a model: what it predicts, who acts on it, how fast and cheap it must be, how it is checked against simple rules, and what happens when it fails.Build · ~110 minutes
  20. 20Monitoring Models in Production — Data Drift and Concept DriftMonitoring watches a live model for data drift (inputs change) and concept drift (what inputs mean changes), since models keep giving answers even after they stop being right.Build · ~110 minutes
  21. 21Feature Stores and Training-Serving SkewA feature store keeps one definition of each model input and serves both its history for training and its latest value for live predictions, so the model sees the same thing in both places.Build · ~100 minutes

Lessons from AI Engineering from Scratch by Rohit Ghumare and contributors, MIT license. Notes for owners by Lozo. License

Next step

Rather have it built for you?

Reading is optional. Tell me what's stuck and I'll build it into the business you already run.

Tell me what's stuck
  • A real conversation, not a pitch
  • You keep your accounts
  • Nothing sends without your yes