# Learn

Long-form writing on the engineering discipline underneath data, machine learning and AI — derived from first principles, not summarised from elsewhere.

## Deep Learning

- [What Are Neurons?](/learn/deep-learning/foundations/what-are-neurons.md): An artificial neuron is a weighted sum followed by a non-linearity, which makes it a logistic regression that has stopped claiming to model a probability. This post explains what the non-linearity is actually for, why stacking linear units gains nothing, and how a layer becomes a matrix multiply.

## Series

- [From Matrices to Neural Networks](/learn/series/from-matrices-to-neural-networks.md): Eight posts across three categories, read in dependency order. Linear algebra and calculus give the space and the gradient, linear and logistic regression introduce the vocabulary of fitting, and neurons and backpropagation generalise all of it into a network you can train.
- [The Loss Function Toolkit](/learn/series/the-loss-function-toolkit.md): Maximum likelihood estimation shows where a loss function comes from statistically; entropy and cross-entropy show where the same loss comes from informationally. This series puts both derivations side by side and ends at the specific case every classifier relies on — why cross-entropy is the natural loss for classification, not merely the conventional one.
- [From One Model to Many: The Ensemble Story](/learn/series/the-ensemble-story.md): The bias-variance tradeoff explains why a single model is stuck choosing what to get wrong. Decision trees give a model simple enough to combine cheaply, forests and boosting show the two structurally different ways of combining it, and cross-validation supplies the honest measurement that tells you whether the combination actually helped.
- [Anatomy of a Modern Neural Network](/learn/series/anatomy-of-a-modern-neural-network.md): An architecture on paper and a network that trains are separated by five practical decisions this series takes in turn — the non-linearity, the starting weights, the normalisation that keeps activations workable, the regularisation that fights memorisation, and the optimiser doing the updating. Each is a fix to a specific, nameable failure.
- [The Road to Transformers](/learn/series/the-road-to-transformers.md): Recurrent networks were the obvious way to handle sequences until their own architecture — weight sharing across time — turned into the vanishing gradient problem. This series follows that failure directly to its resolution, attention, and ends with the transformer built from attention block by block.