JEPA · 2026-10-21
LocaLeJEPA: Maximum Entropy Representations under Geometric Priors
Kartik Tandon, Alessio Borgi, Pavel Snopov, In\'es Garc\'ia-Redondo, Davide D'Ascenzo, Sebastiano Cultrera di Montesano, Fabrizio Silvestri, Pietro Li\`o, Francesca Dominici, Randall Balestriero, Claudio Battiloro
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
LocaLeJEPA extends LeJEPA’s entropy-maximizing JEPA objective to non-Euclidean embedding geometries by using a locally Euclidean distribution-matching regularizer (LocaLeReg). It partitions each batch of embeddings into K latent-space neighborhoods (K-means), then matches residuals within each neighborhood to a uniform distribution on a Euclidean ball via an Epps-Pulley-based characteristic-function test. On ImageNet-1K pretraining, it outperforms LeJEPA across linear and k-NN and on multiple downstream datasets, with code released.
Why it matters
The authors argue LeJEPA’s isotropic Gaussian target relies on Euclidean assumptions, which may fail for exotic embedding spaces (e.g., hyperbolic ViTs, equivariant backbones). LocaLeJEPA provides a stable and scalable alternative that preserves manifold structure by making the entropy-maximization objective local, while matching different embedding geometries through a neighborhood-based target.
Method
- Introduce LocaLeJEPA: a manifold-theoretic JEPA objective coupled to the geometry of the data manifold and an embedding space, aiming for local entropy maximization rather than a single global Gaussian target.
- Use LocaLeReg: form K latent-space neighborhoods via cosine-similarity K-means on view means (with gradients not passing through assignment), then center residuals in each neighborhood before matching their distribution.
- Within each neighborhood, drive residuals toward the uniform distribution on a Euclidean ball using an Epps-Pulley distribution-matching objective over univariate projections (random directions) and aggregate over neighborhoods and augmented views.
Limitation
The construction requires each neighborhood to retain enough residuals for its statistic to be informative, since centering leaves only |Cq|−1 free directions and neighborhoods are formed within the batch.
Abstract (from arXiv)
Joint-Embedding Predictive Architectures (JEPA) offer an appealing approach to self-supervised learning, but are typically prone to representation collapse. LeJEPA resolves this by introducing a theoretically grounded distribution-matching objective that drives the embedding towards an isotropic Gaussian. However, this approach is fundamentally tied to Euclidean assumptions, and its global target does not reflect domain-specific geometric priors, such as the use of non-Euclidean embedding geometries, e.g., equivariant CNNs or hyperbolic ViTs. We thus introduce LocaLeJEPA, a framework motivated by a manifold-theoretic reformulation of LeJEPA's entropy-maximization property. The accompanying LocaLeReg, a novel locally Euclidean distribution-matching regularizer, enables a stable and scalable implementation of the LocaLeJEPA framework. We then perform extensive experiments in the vision setting on models trained on ImageNet-1K. LocaLeJEPA significantly outperforms LeJEPA when paired with hyperbolic ViT-S/8 and equivariant ResNet-50 backbones, and also with ordinary Euclidean ViT-S/8 and ViT-L/14 backbones. This holds both in terms of linear and k-NN classification performance on ImageNet and on a series of downstream tasks tested. We view the LocaLeJEPA formalism as a principled approach to the JEPA paradigm that better respects intrinsic data geometry while also accommodating generalizations to exotic embedding geometries. Code available at https://github.com/alessioborgi/LocaLeJEPA.
Related papers
- LEF-JEPA: Denoising-Trained Latent Energy Fields for Spatial Representation Learning
- A Stochastic Latent Variable Generalization of Joint-Embedding Predictive Architectures
- Drive vs. Decay: On the Training Dynamics of Joint-Embedding Predictive Architectures
- CI-JEPA: A Counterfactual Analysis of Latent Representations in Joint-Embedding Predictive Architectures for Self-Supervised Learning
- Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-20 · Indexed by Pub-AI: 2026-10-21 UTC
LocaLeJEPA: Maximum Entropy Representations under Geometric Priors
Preprint posted: 2026-10-20 · Original record ↗
- arxiv
- 2610.28458
Explicit source versions (1)
- v1 · 2026-10-20
- Preprint available
Source date: 2026-10-20 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Joint-Embedding Predictive Architectures (JEPA) offer an appealing approach to self-supervised learning, but are typically prone to representation collapse. LeJEPA resolves this by introducing a theoretically grounded distribution-matching objective that drives the embedding towards an isotropic Gaussian. However, this approach is fundamentally tied to Euclidean assumptions, and its global target does not reflect domain-specific geometric priors, such as the use of non-Euclidean embedding geometries, e.g., equivariant CNNs or hyperbolic ViTs. We thus introduce LocaLeJEPA, a framework motivated by a manifold-theoretic reformulation of LeJEPA's entropy-maximization property. The accompanying LocaLeReg, a novel locally Euclidean distribution-matching regularizer, enables a stable and scalable implementation of the LocaLeJEPA framework. We then perform extensive experiments in the vision setting on models trained on ImageNet-1K. LocaLeJEPA significantly outperforms LeJEPA when paired with hyperbolic ViT-S/8 and equivariant ResNet-50 backbones, and also with ordinary Euclidean ViT-S/8 and ViT-L/14 backbones. This holds both in terms of linear and k-NN classification performance on ImageNet and on a series of downstream tasks tested. We view the LocaLeJEPA formalism as a principled approach to the JEPA paradigm that better respects intrinsic data geometry while also accommodating generalizations to exotic embedding geometries. Code available at https://github.com/alessioborgi/LocaLeJEPA.
