arthurdouillard / continual-learning-terminology

continual-learning-terminology

Terminology of Continual Learning, a key word -> a definition. Ideally, each key word is linked to the first paper that introduced it.

Multiple keywords can correspond to actually the same definition. In that case, a link is made to centralize all of them to the most common one.

Keyword	Citation	Definition	Similar keyword
NC: New Classes	Lomonaco17	Each new task is made of new never-seen-before classes.	Class-Incremental Learning
NI: New Instances	Lomonaco17	Each new task is made of new samples, often from a different domain.	Domain Incremental Learning
NIC: New Instances and Classes	Lomonaco17	Each new task is made of existing classes from new domains and from new classes.
Domain Incremental Learning		Each new task is made of new samples coming from new domains
CIL: Class Incremental Learning		Each new task is made of new never-seen-before classes; no task id at test-time.	Single Head
Task Incremental Learning		Each new task is made of new never-seen-before classes; access task id at test-time.	Multi Heads
Rehearsal Learning		Re-use a bounded amount of previous data in a rehearsal memory.
Episodic Replay		Re-use a bounded amount of previous data in a rehearsal memory.	Rehearsal Learning
Pseudo-Rehearsal Learning		Generate previous data to be rehearsed.	Rehearsal Learning
Generative Replay		Generate previous data to be rehearsed.	Pseudo-Rehearsal Learning
Single Head	Chaudhry18	No access to task identifier at test-time.
Multi Heads	Chaudhry18	Access to task identifier at test-time.
Joint Model		Oracle model that learns jointly on all tasks