Catalogues
Families of methods compared side by side — architectures as a lineage, metrics and losses as a family. Each one is a table you can scan, then a page for every entry that has one.
- Activation functionsWithout a non-linearity between layers, a deep network collapses into a single linear one — these are the functions that stop that happening, and what each costs in return.Open catalogue →
- Classification metricsSix common classification scores are ratios of the same four confusion-matrix cells and two read the probabilities instead — which each one uses decides what it can and cannot see.Open catalogue →
- Clustering methodsRead the "Shape it can find" column first. Every method here partitions whatever it is given, so the choice is not which is best but which one's idea of a cluster matches yours.Open catalogue →
- Computer vision architecturesThe lineage of image-recognition networks from AlexNet to the Vision Transformer, and what each one changed about how images get processed.Open catalogue →
- Forecasting methodsEvery row is a different answer to what the future should look like. The first one has no parameters at all, and more methods fail to beat it than anybody reports.Open catalogue →
- Linear model familyEvery model here is the same weighted sum fitted by least squares — what separates them is the penalty added to the objective, and each penalty is a different claim about your features.Open catalogue →
- Missing-data methodsDropping rows and filling gaps are both decisions about the answer. Which one is safe depends on why the values are missing, and on which number you report.Open catalogue →
- Multiple-testing correctionsTest many things at the 5% level and something will pass by chance. Each correction makes a different promise about the mistakes you will make, and pays for it in findings.Open catalogue →
- NLP tasksNine tasks arranged by the one thing that separates them — how large the space of acceptable answers is. Read the “Scored by” column first.Open catalogue →
- OptimizersEvery optimiser here descends a gradient — they differ in how much data each step looks at, and how much they remember between steps. Those two axes explain the whole family.Open catalogue →
- Regression metricsSquared, absolute, relative or logarithmic — each regression metric encodes a different belief about what a bad prediction costs, and choosing one is making that claim.Open catalogue →
- Text metricsTen ways to score generated text, and what each one can and cannot see. The "Breaks when" column is the reason you should never report only one.Open catalogue →
- Transformer familyEvery model here is the same block from Attention Is All You Need, stacked. Read the first column — what each position is allowed to attend to decides the objective, the task, and everything after.Open catalogue →
- Tree ensemblesEvery method here is many decision trees combined into one prediction. They split into two families that want opposite things from their members, and that split explains most of the differences.Open catalogue →
- Validation strategiesEvery method here answers one question — which rows may a model be scored on. Read the "Breaks when" column first, because choosing wrongly does not produce an error, it produces a better number.Open catalogue →
- Vision tasksSeven tasks arranged by what the answer is attached to — the image, a box, a pixel, an instance, or nothing. Read the “Costs” column first.Open catalogue →