Khalid ShamsPrincipal Solutions Architect

Learn AI · Guided path 2 of 6

How a machine learns

When you finish: You can follow one model from raw text to a released version, and explain what training actually changes.

About 25 minutes · 9 stops, 8 of them ideas from the map · progress is kept on this device ·

Start at stop 1 All paths

The stops, in order

Each stop is one page. Read it, picture it, mark it, move on.

Ideas open on their own page with three depths. Atlas stops open the interactive lesson at the right step and count as done when you follow the run to its end. Sections count when you have read them.

  1. 1TrainingThe practice phase: the model makes guesses, gets told how wrong it was, and adjusts. Millions of times.idea
  2. 2ParametersThe adjustable numbers inside a model. Learning means tuning them.idea
  3. 3Artificial NeuronA tiny calculator: it takes numbers in, weighs them, and passes on a signal if the total is strong enough.idea
  4. 4Neural NetworkMany artificial neurons wired together in layers, passing signals forward.idea
  5. 5BackpropagationHow a network figures out which of its knobs to blame for a mistake, so it can fix them.idea
  6. 6Follow one model from text to releaseThe second atlas lesson. Fifteen steps; try the three ways it can be held back.atlas lesson
  7. 7Supervised LearningLearning from examples that come with the right answer attached.idea
  8. 8Fine-tuningTraining an already-trained model a little more on your own examples so it behaves the way you need.idea
  9. 9AlignmentMaking sure AI does what people actually intend, safely and honestly.idea

← Path 1 Path 3: What a language model does →