Methods.md

---
tags: [meta]
---

# Methods

One note per named method in `methods/`. The `family` label is coarse
grouping only — the information lives in each note's **aspects** (what it
stores, its update rule, its declared assumptions, its costs) and in
**documented behavior** (who observed it working/breaking, where). Edges:
`variant-of`, `refines`, `combines` for lineage; `uses:: [[trick]] (how)` for
reusable components — the parenthesis is mandatory, a bare trick link is just
a label.

For the **lineage view** — method family trees walked from the `variant-of` /
`refines` / `combines` edges — see [[_methods-index]] (generated). It is a
genealogy, not a taxonomy: the `family` label stays coarse and the substance
lives in the aspects below.

## Assumptions & costs at a glance

```dataview
TABLE WITHOUT ID file.link AS Method, family AS Family, stores AS Stores, requires-boundaries AS "Boundaries?", requires-task-id AS "Task id?", memory-growth AS "Memory growth"
FROM "methods"
SORT family ASC, file.name ASC
```

This table is the method-level application of
[[settings-should-declare-drift-assumptions]]: each row is a method's
declared assumptions, so mismatches between a method and a setting become
visible before any benchmark is run.

## By family

```dataview
TABLE WITHOUT ID file.link AS Method, introduced-in AS "Introduced in"
FROM "methods"
GROUP BY family
```

## Lineage at a glance

- Replay line: [[experience-replay]] → variants [[generative-replay]] (no
  storage), [[gdumb]] (no continual training), [[der]] (logits instead of
  labels) → refinement [[er-ace]] (asymmetric updates).
- Regularization line: [[ewc]] → [[si]] (online importance); functional
  cousin [[lwf]].
- Architecture line: [[progressive-networks]] → [[mntdp]] (sub-linear growth
  via module reuse).
- Hybrid: [[icarl]] (replay + distillation + prototype classifier).