Appendix C — Glossary and paper timeline

Give each term a one-line meaning, a concrete analogy, and a common mistake, then anchor the JEPA lineage in primary sources.

Course progress Course outline 34 of 34 lessons available

Part 0 — Get the map

Part 1 — Why a small objective can learn to see

Part 2 — Send a video through one encoder

Part 3 — Read the experiments, not just the headline

Part 4 — From the official repository to your own experiment

Part 5 — Put the representation back on the world-model road

28 Chapter 27 — The important boundary: an encoder is not a planner available now
29 Chapter 28 — How LeVJEPA might feed a future world model available now
30 Chapter 29 — Ten projects, from first experiment to paper-sized question available now

Appendices — A backpack for the trail

Names are labels, not bags of spare parts

FamilyJEPA is a blueprint
MethodI, V, Le, and LeV differ
Versionnewer is not stronger evidence
Evidencepaper, code, weights, reproduction

A glossary stops one paper’s machinery from sneaking into another paper under a shared surname.

A hammer, screwdriver, and drill may all live in a carpenter’s toolbox. That does not put a battery inside the hammer. JEPA is similarly a research blueprint and family of methods. Members choose different encoders, target branches, predictors, regularizers, and data.

The most common reading error is not a bad calculation. It is quietly installing I-JEPA’s EMA target encoder, V-JEPA’s predictor, and V-JEPA 2-AC’s action model inside LeVJEPA. This glossary gives every term a small “do not smuggle this in” label.

Start with the names themselves

The official spelling is LeVJEPA, not LEVJEPA, Le-VJEPA, or V-JEPA 3.
The LeJEPA paper expands its name as Latent-Euclidean JEPA. It contributes the general objective and SIGReg.
The LeVJEPA paper does not supply a formal long name to expand. The safe description is “a video encoder trained with the LeJEPA objective.”
All these papers are collaborations. Saying LeVJEPA follows Yann LeCun’s research road describes an intellectual lineage; it does not erase the work of the other authors.

The working glossary

Term	One-line meaning	Small picture	Common mix-up
JEPA	An architectural idea that models compatibility or prediction in representation space	guess meaning on a map, not every brick	It does not prescribe one loss or collapse remedy.
I-JEPA	Predicts representations of masked target regions from image context	infer the middle of a puzzle	It is a still-image method, not a planner.
V-JEPA	Extends latent feature prediction to video space-time blocks	infer a hidden stretch of a film	It uses a predictor and EMA-target route; it is not LeVJEPA.
V-JEPA 2	A large-scale video JEPA representation model	an observer trained on much video	Action-free pretraining does not itself control a robot.
V-JEPA 2-AC	Trains an action-conditioned predictor above a frozen encoder	add an “if I move” sandbox	It is a V-JEPA 2 post-training stage, not a LeVJEPA component.
LeJEPA	Joins a JEPA pairing objective with SIGReg in a lean general recipe	match same-answer cards without allowing blanks	Its theory has assumptions and is not video-specific.
LeVJEPA	Uses the LeJEPA collapse-control objective to train a video encoder	describe one film through large and small windows	It is encoder pretraining, not an action world model.
LeWorldModel	Trains next-embedding prediction, SIGReg, and action-conditioned dynamics for control	a sandbox that can try actions	It is a neighbouring branch, not another name for LeVJEPA.
self-supervised	Builds a training signal from the data itself	the video writes its own matching question	Humans still choose data, transformations, and objective.
representation / embedding	A numeric vector produced for an input	a numeric address card	Individual coordinates usually have no hand-written meaning.
encoder	Turns pixels into tokens and representations	a reader writing notes	LeVJEPA’s encoder has no built-in class answer.
global view	A broad spatial crop of one clip	a wide window	It is still an augmented crop, not the entire original recording.
local view	A smaller spatial crop from the same temporal window	a cardboard viewing tube	Moving it to another time changes the pairing contract.
augmentation	Changes a view while intending to preserve its relevant identity	crop a corner or alter brightness	An aggressive transform can erase information the task needs.
invariance	Views of the same clip should receive nearby representations	recognize the same act from two windows	Alone, it permits one constant answer.
complete collapse	Nearly every input maps to the same vector	the whole class hands in one blank card	A low invariance loss may hide it.
dimensional collapse	Embeddings occupy only a low-dimensional subspace	256 streets squeezed into one alley	Samples can differ and still collapse this way.
SIGReg	Constrains embedding distributions through Gaussianity of random projections	inspect a city from many shadows	It does not test physical meaning or controllability.
isotropic Gaussian	A zero-centered Gaussian reference with equal scale in every direction	a cloud with no favored compass point	It is neither a uniform sphere nor the topology of the real world.
random projection	Maps a high-dimensional point to one scalar with a random unit vector	shine a flashlight from one direction	Finitely many directions are an approximation.
characteristic function	The complex-exponential expectation that uniquely describes a distribution	a frequency fingerprint	The code compares only finitely many frequency nodes.
Epps–Pulley statistic	A smooth discrepancy between empirical and Gaussian characteristic functions	compare two fingerprints	A training loss is not a formal pass/fail normality certificate.
patch	A `16×16` pixel region in one frame	a mosaic tile	It becomes a transformer token only after embedding.
token	The vector made from an embedded patch	a block stamped with coordinates	Training randomly drops 95% of patch tokens.
tubelet	A spatial patch extended across several frames	stack matching tiles into a tube	LeVJEPA defaults to `tau=1`, so it does no temporal stacking here.
`[CLS]`	A learned token that summarizes the clip	a summary card	Patch tokens cannot read it, preventing future information from flowing back.
projector	Maps `[CLS]` into the 256-D space constrained by SIGReg	a temporary training translation desk	It is discarded after pretraining and is not a classifier.
LayerNorm	Normalizes features within each token	press each card onto a constrained surface	Its output geometry is one reason SIGReg follows a projector.
RoPE	Encodes relative time and space positions through rotations in attention	stamp time, row, and column on a block	It does not teach causal intervention by itself.
block-causal attention	Bidirectional within a frame, current-and-past only across frames	classmates share a page but cannot open tomorrow’s page	`[CLS]` may see the clip; patches cannot read `[CLS]`.
target encoder	A separate branch that supplies target features during training	a second teacher writing answers	LeVJEPA’s objective has none.
stop-gradient	Blocks backpropagation through a branch	a one-way valve	Both LeVJEPA view branches receive gradients.
predictor	Estimates a target representation from context	the part that guesses	LeVJEPA aligns a shared encoder’s view summaries directly and has no predictor.
Polyak/EMA weights	A moving average copy of training weights used for evaluation	a smoothed copy of the scorebook	The official copy does not make targets or enter the loss.
frozen probe	Trains a small readout while leaving the encoder fixed	test the notes without rewriting them	Decodability does not prove the model actively uses the information.
attentive probe	Pools tokens with a learned query before classifying	a marker who learns where to look	It is stronger than plain linear mean pooling; scores are not interchangeable.
epoch	One pass over a defined dataset	finish one lap of a workbook	The work per sample differs between methods.
FLOPs	An estimate of floating-point operations	accounting units for computation	It is not wall time or energy consumption.
world model	Predicts how a world state may evolve	an internal sandbox	A temporally causal encoder is only a possible foundation.
MPC	Plans a horizon, acts briefly, observes, and replans	take one step and reopen the map	LeVJEPA itself contains no MPC.

One road with two important forks

Public date	Primary work	What it advances	What it does not establish
2006	Energy-based learning tutorial	Low energy for compatible configurations, high energy for incompatible ones	A JEPA or video recipe
2022-06-27	A Path Towards Autonomous Machine Intelligence v0.9.2	A blueprint joining perception, configurable world model, cost, actor, memory, and hierarchical JEPA	A completed autonomous system
2023	I-JEPA	Representation prediction for large target blocks in still images	Actions or planning
2024	V-JEPA	Latent prediction for video space-time blocks	A control model
2025-06	V-JEPA 2 / 2-AC	Large-scale observation pretraining, then a separate action-conditioned stage with MPC	General long-horizon autonomy
2025-11	LeJEPA v3	An explicit, scalable SIGReg route around collapse	An assumption-free global optimum for every task
2026-03	V-JEPA 2.1	Stronger dense token representations on the original V-JEPA branch	A move to the LeJEPA/SIGReg branch
2026-03	LeWorldModel v3	Action-conditioned latent dynamics and planning on the LeJEPA branch	Broad real-world validation from small control studies
2026-08-27	LeVJEPA v1	Lean video-encoder pretraining with sparse observation and block-causal attention	An action predictor or planner

This is a tree, not a sequence of automatic upgrades. V-JEPA 2.1 continues the asymmetric feature-prediction branch. LeWorldModel and LeVJEPA both draw on LeJEPA/SIGReg but pursue action dynamics and general video representation, respectively.

Put an evidence label on every claim

Label	What it can support	What it cannot support automatically
method definition	A paper explicitly specifies a graph or objective	The implementation is bug-free or reproducible.
author-reported	Authors report a value under a named protocol	An independent team will obtain it.
author code	A concrete interface or default can be inspected	Every setting matches every paper experiment.
public weights	Feature extraction and downstream evaluation can be run	The training process has been reproduced independently.
independent reproduction	A third party repeats a stated scope	The claim holds for all models, data, and tasks.
assumption-bound theory	A conclusion holds under listed assumptions	An unconditional guarantee in the physical world.

At this course’s 2026-09-05 (Asia/Tokyo) evidence cutoff, LeVJEPA remains a newly released v1 preprint. Its results are author-reported. This review found no independent reproduction that would justify a stronger label.

Sources for the family tree

The lineage above was cross-checked against LeCun’s 2022 AMI paper, the I-JEPA CVPR record, V-JEPA at TMLR/OpenReview, the V-JEPA 2 publication page, LeJEPA v3, and LeVJEPA v1. Public dates establish order, not strength of evidence.

Keep the labels on the drawers

JEPA is a family blueprint; each member’s computation graph must be checked separately.
LeVJEPA has no official long name to invent. It is a video encoder trained with the LeJEPA objective.
A later date, public code, and independent reproduction are three different facts.

Name the right machine

Why can V-JEPA’s predictor not be drawn inside LeVJEPA?
Why may we say “no EMA target encoder” even though the official implementation keeps EMA weights?
What is the central task boundary between LeVJEPA and LeWorldModel?

Check the answers

They use different training graphs; LeVJEPA directly aligns global and local [CLS] outputs from a shared encoder and uses SIGReg against collapse.
The EMA copy is a smoothed evaluation checkpoint. It neither produces targets nor participates in the loss.
LeVJEPA learns general video representations; LeWorldModel explicitly learns action-conditioned next-latent states and uses them for planning.