Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture

By Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael G. Rabbat, Yann LeCun, Nicolas Ballas. Computer Vision and Pattern Recognition, 2023

1,029 citations recorded by the source corpus.

This paper demonstrates an approach for learning highly semantic image representations without relying on hand-crafted data-augmentations. We introduce the Image-based Joint-Embedding Predictive Architecture (I-JEPA), a non-generative approach for self-supervised learning from images. The idea behind I-JEPA is simple: from a single context block, predict the representations of various target blocks in the same image. A core design choice to guide I-JEPA towards producing semantic representations is the masking strategy; specifically, it is crucial to (a) sample target blocks with sufficiently large scale (semantic), and to (b) use a sufficiently informative (spatially distributed) context block. Empirically, when combined with Vision Transformers, we find I-JEPA to be highly scalable. For instance, we train a ViT-Huge/14 on ImageNet using 16 A100 GPUs in under 72 hours to achieve strong downstream performance across a wide range of tasks, from linear classification to object counting and depth prediction.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/ee57e4d7a125f4ca8916284a857c3760d7d378d3

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

Citations in the corpus

Revisiting Feature Prediction for Learning Visual Representations from Video (2024, methodology): “On tasks that require a fine-grained understanding of motion (Something-Something-v2), the V-JEPA models provide a major improvement (over +21 points) compared to large-scale image baselines, such as DINOv2, OpenCLIP, and I-JEPA.”
UniVLA: Learning to Act Anywhere with Task-centric Latent Actions (2025, positive): “Inspired by joint-embedding predictive architectures (JEPA) [5, 6, 96], we propose using DINOv2 [62] spatial patch features as semantically rich representations.”
A Survey on Vision--Language--Action Models for Embodied AI (2024, positive): “I-JEPA [52] is motivated by the joint-embedding predictive architectures proposed by [55].”
Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think (2024, positive): “• I-JEPA (Assran et al., 2023) predicts missing parts of an image by learning representations through joint-embedding, focusing on the context of the entire image without relying on pixel-level reconstruction.”
Aligning Cyber Space With Physical World: A Comprehensive Survey on Embodied AI (2024, neutral): “The world models [14], [15] exhibit remarkable simulation capabilities and promising comprehension of physical laws, which makes embodiedmodelscomprehensivelyunderstandboththephysical and real environments.”
Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs (2024, background): “This category includes contrastive SSL [5, 7, 16, 17] and mask-based SSL [4, 18, 70].”
DINOv2: Learning Robust Visual Features without Supervision (2023, methodology): “Our distillation method ends up close to the one described by Duval et al. (2023), except we do not modify the loss terms for distillation and evaluate the EMA of the student.”
V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning (2025, positive): “…vision encoders that perform well along a wide range of downstream tasks using self-supervised learning approaches from images (Grill et al., 2020; Assran et al., 2023; Oquab et al., 2023; Fan et al., 2025), videos (Bardes et al., 2024; Carreira et al., 2024; Wang et al., 2023; Rajasegaran et…”
Cambrian-1: A Fully Open, Vision-Centric Exploration of Multimodal LLMs (2024, methodology): “However, there is a much broader pool of visual models that learn representations using only visual signals—such as self-supervised models [8, 96], segmentation [61], depth-supervised [67], and diffusion models [104] (see Fig.”
Giving Sensors a Voice: Multimodal JEPA for Semantic Time-Series Embeddings (2026, unknown): “The context and target encoders are fully defined by our embedding model as where ] where m θ represents learnable placeholders that guide the predictor encoder to generate embeddings for masked positions, see (Assran et al., 2023) for more information.”
Eigenfunction Extraction for Ordered Representation Learning (2025, unknown): “Prominent examples include BYOL (Grill et al., 2020), SimSIAM (Chen and He, 2021), VICReg (Bardes et al., 2022), Barlow Twins (Zbontar et al., 2021), SWAV (Caron et al., 2020), DINO (Caron et al., 2021; Oquab et al., 2024; Siméoni et al., 2025), and JEPA (Assran et al., 2023, 2025).”
Time to Embed: Unlocking Foundation Models for Time Series with Channel Descriptions (2025, unknown): “More recently, foundational models have revolutionized representation learning across natural language processing, computer vision, and audio (Devlin et al., 2019; Nussbaum et al., 2025; Assran et al., 2023; Kirillov et al., 2023; Baevski et al., 2020; Brown et al., 2020; Radford et al., 2021).”
Fusing Multi‐ and Hyperspectral Satellite Data for Harmful Algal Bloom Monitoring With Self‐Supervised and Hierarchical Deep Learning (2025, unknown): “…Spectroradiometer (MODIS) instrument onboard the Aqua platform, and the Ocean and Land Color Instrument (OLCI) onboard the Sentinel-3A, and Sentinel-3B platforms were used in the time period of June 1, 2018 to December 31, 2019 to overlap with testing and analysis done in (Luis et al., 2023).”