Machine Learning Mind Map

Every machine learning visualization on this site on one map. The branches are the main ideas. Each leaf is a page; the grey line under it is the one thing to remember about it. Click a branch to focus on it, click a page to open it.

Recall mode hides the page names. Read the hint, say the name, then click the "?" to check.

1. How the map is organized

Overview of the machine learning mind map: Learning setup, Linear models, Other classic models, Unsupervised, Training, Neural networks, Architectures, Language models, LLM speed and tuning, Retrieval and agents. — The map at a glance: 10 branches, 57 pages.

Ask three questions about a page and you know its branch:

Is it a model or a way to train one? Gradients, losses and optimizers are training. The model itself goes on a model branch.
Does it use neural layers? If not, it is a linear, classic or unsupervised model. If yes: plain layers are neural networks, special wiring (convolution, recurrence, attention) is architectures.
Is it about text models? How an LLM reads and writes is language models; making it faster or aligned is speed and tuning; feeding it documents and tools is retrieval and agents.

2. Model families side by side

Branch	Needs labels?	How it learns	Example
Linear models	yes	gradient descent or closed form	logistic regression
Other classic models	yes	counts, splits, or stores the data	decision tree
Unsupervised	no	groups or flattens the points	k-means
Neural networks	usually	backpropagation + gradient descent	digit network
Language models	no, text is its own label	next-token cross-entropy	GPT
Retrieval and agents	no training	adds data and tools at run time	RAG

3. Five facts to remember

Almost every modern model is trained the same way: compute a loss, get the gradient by backpropagation, take a small step.
A learning rate that is too large makes training diverge; too small makes it crawl.
Accuracy misleads on imbalanced data: check precision and recall.
Overfitting means low training loss but high test loss; regularization, dropout and early stopping fight it.
An LLM only predicts the next token; attention lets each token look at all earlier tokens.

See all machine learning pages in the index, or the other mind maps.