Core concepts
Learn the ideas that determine whether an experiment is trustworthy. Each lesson uses a small notebook to make an abstract failure mode visible.
9 concepts
Train, Validation, and Test Sets
Training data fits the model, validation data chooses the design, and test data estimates the performance of the finished design.
Generalization
Generalization is the ability to perform well on new examples from the conditions where the model will actually be used.
Bias and Variance
Bias is systematic error from an overly restrictive model; variance is instability caused by sensitivity to the particular training sample.
Metrics
A metric compresses model behavior into a number, so choosing one means choosing which mistakes matter.
Data Leakage
Leakage lets training or evaluation use information that will not exist when a real prediction is made.
Regularization
Regularization makes some fitted solutions preferable to others so the learned behavior is more likely to survive new data.
Loss and Optimization
The loss defines what counts as a training mistake; the optimizer searches for parameters that reduce that loss.
The Perceptron
A perceptron adds weighted inputs and turns the result into one of two decisions.
Backpropagation
Backpropagation sends the prediction error backward through a network to measure how much each parameter contributed to it.