I-JEPA
PulseAugur coverage of I-JEPA — every cluster mentioning I-JEPA across labs, papers, and developer communities, ranked by signal.
3 day(s) with sentiment data
-
JEPA models in medical imaging show vulnerability to attacks, research finds
A new research paper titled "Cracks in Embeddings" investigates the robustness and stability of Joint Embedding Predictive Architecture (JEPA) models when applied to medical imaging, specifically chest X-rays. The study…
-
Anatomy-guided I-JEPA improves retinal OCT representation learning
Researchers have explored an anatomy-guided approach for the I-JEPA self-supervised learning architecture when applied to retinal OCT scans. The standard I-JEPA method places targets uniformly, which is suboptimal for r…
-
AI planning enhanced with adapter on frozen visual embeddings · 1 source tracked
Researchers have developed a method to improve visual world-model planning by fitting a residual adapter to frozen I-JEPA embeddings. This adapter learns from successful and failed trajectories to better score predicted…
-
New S-JEPA model introduces stochasticity for diverse image completions
Researchers have introduced S-JEPA, a novel stochastic generalization of Joint-Embedding Predictive Architectures (JEPAs) for masked image prediction. Unlike previous deterministic approaches that estimate a single comp…
-
Predictor width significantly impacts I-JEPA's feature learning, study finds
Researchers have investigated the effectiveness of masked latent prediction methods like I-JEPA in learning predictable features. By using a simulator where all image factors are known, they measured the readability of …
-
JEPA-Anything framework unifies world model creation across 7 domains
Researchers have introduced JEPA-Anything, a novel domain-agnostic framework designed to build world models. This framework extends joint-embedding predictive architectures (JEPAs) by incorporating Orthogonal Predictive…
-
New theory explains and mitigates representation collapse in AI architectures
Researchers have developed a new theoretical framework to understand and mitigate representation collapse in Joint-Embedding Predictive Architectures (JEPAs). By analyzing the gradient flow during early training, they i…
-
New JEDI framework distills large vision models for efficient satellite image segmentation
Researchers have developed JEDI (JEPA-to-Edge Distillation), a novel two-stage framework designed to efficiently transfer knowledge from large vision models to smaller, more deployable ones for satellite imagery segment…
-
PredErase method enables training-free object removal from images
Researchers have developed PredErase, a novel training-free method for removing objects and their associated effects like shadows from images. This technique utilizes a frozen image generation model, FLUX.2, and a predi…
-
UniJEPA unifies image and video visual world modeling
Researchers have introduced UniJEPA, a novel unified architecture for self-supervised visual world modeling. This new framework integrates both image-level photometric prediction and video-level temporal prediction into…
-
DSeq-JEPA architecture enhances visual representation learning with sequential prediction
Researchers have introduced DSeq-JEPA, a novel architecture for self-supervised visual representation learning. This model builds upon the Image-based Joint-Embedding Predictive Architecture (I-JEPA) by incorporating a …
-
New method aligns AI self-supervised learning with scientific imaging physics
Researchers have developed a new method for designing data augmentations in self-supervised learning (SSL) specifically for scientific imaging. This approach, termed physics-aligned augmentation, considers the unique sy…
-
JEPADepth framework enhances self-supervised monocular depth estimation
Researchers have developed JEPADepth, a novel self-supervised framework for monocular depth estimation that integrates a masked predictive representation learning objective inspired by Image Joint-Embedding Predictive A…
-
New Bootleg method enhances self-supervised learning for AI models
Researchers have developed a new self-supervised learning method called Bootleg, which aims to combine the stability of generative approaches with the efficiency of predictive methods. Bootleg trains a model to predict …
-
JEPA models face challenges with language's conditional structure
A new paper explores the challenges of applying Joint-Embedding Predictive Architectures (JEPAs) to language processing, contrasting their effectiveness in image and audio domains with their limitations in text. The res…
-
JEPA predictors prove portable for occluded feature completion
Researchers have demonstrated that the predictor component of Joint-Embedding Predictive Architectures (JEPAs), typically discarded after training, can be repurposed as a transferable operator for occluded feature compl…
-
University of Michigan unveils NeuroVFM for neuroimaging analysis
Researchers at the University of Michigan have developed NeuroVFM, a novel foundation model for neuroimaging. Trained using the Vol-JEPA approach on over 5.24 million clinical MRI and CT scans, NeuroVFM learns from uncu…
-
New JEPA-style model learns useful network fingerprint embeddings
Researchers have developed JA4-JEPA, a Transformer-based model that applies JEPA-style predictive learning to network fingerprints. This approach, which learns by matching latent predictions rather than regenerating inp…
-
Face-trace pipeline enables open-set attribution of synthetic face generators
Researchers have developed a new pipeline called Face-trace for open-set synthetic face source attribution. This method can identify the generator of a synthetic face image, even if the generator was not known during tr…
-
SiamJEPA uses Siamese encoders for improved self-supervised learning
Researchers have introduced SiamJEPA, a novel approach to self-supervised representation learning that utilizes Siamese student encoders within Joint Embedding Predictive Architectures (JEPAs). Unlike previous JEPA mode…