LightMySky

Machine Learning

All 34 topics on the way, in the order they build on each other.

  1. Joint, Marginal and Conditional Probabilityages 18 to 19 · on the way
  2. Bayes' Rule and Updating a Beliefages 18 to 19 · on the way
  3. Likelihood and the Maximum Likelihood Estimateages 18 to 20 · on the way
  4. Data as a Matrix: Rows, Features and the Targetages 18 to 20 · on the way
  5. Linear Regression by Least Squaresages 19 to 20 · on the way
  6. Gradient Descent: Following the Slope Downhillages 19 to 20 · on the way
  7. Logistic Regression and the Decision Boundaryages 19 to 21 · on the way
  8. Overfitting and the Bias-Variance Trade-offages 19 to 21 · on the way
  9. Regularisation: Charging for Complexityages 20 to 21 · on the way
  10. Cross-Validation and Honest Model Selectionages 20 to 21 · on the way
  11. Beyond Accuracy: Precision, Recall and the Cost of an Errorages 20 to 22 · on the way
  12. Decision Trees and Ensemblesages 20 to 22 · on the way
  13. Clustering: Structure Without Labelsages 20 to 22 · on the way
  14. Principal Components and Dimensionality Reductionages 20 to 22 · on the way
  15. From Perceptron to Multilayer Networkages 21 to 22 · on the way
  16. Backpropagation: Assigning Blame for an Errorages 21 to 22 · on the way
  17. Deep Learning: Convolution, Sequence and Scaleages 21 to 22 · on the way
  18. Attention as a Learned Lookupages 22 to 23 · on the way
  19. The Transformer Block: Heads, Residuals and Normalisationages 22 to 23 · on the way
  20. Tokenisation and What a Model Actually Readsages 22 to 23 · on the way
  21. Pretraining by Predicting the Next Tokenages 22 to 23 · on the way
  22. Scaling Laws and the Compute Budgetages 22 to 23 · on the way
  23. Adapting a Pretrained Model: Fine-Tuning and Low-Rank Updatesages 22 to 24 · on the way
  24. Generative Models: Learning a Distribution You Can Sampleages 22 to 24 · on the way
  25. Diffusion Models: Learning to Undo Noiseages 22 to 24 · on the way
  26. Markov Decision Processes: States, Actions and Returnages 22 to 24 · on the way
  27. Temporal-Difference Learning and Q-Learningages 23 to 24 · on the way
  28. Policy Gradients and the Actor-Critic Splitages 23 to 24 · on the way
  29. Learning from Human Preferences: Reward Models and Policy Optimisationages 23 to 24 · on the way
  30. Training Across Many Devices: Data, Model and Pipeline Parallelismages 23 to 24 · on the way
  31. Serving a Model: Batching, the Key-Value Cache and Quantisationages 23 to 24 · on the way
  32. Benchmarks and the Traps in Themages 22 to 24 · on the way
  33. Fairness, Accountability and the Limits of a Modelages 21 to 22 · on the way
  34. Interpretability: Probes, Features and Circuitsages 23 to 24 · on the way
Machine Learning · Computing · LightMySky