I-JEPA: Image Joint-Embedding Predictive Architecture

A PyTorch implementation of I-JEPA (Image Joint-Embedding Predictive Architecture), inspired by the work of Yann LeCun and Meta AI.

---

Introduction

I-JEPA is a self-supervised learning framework introduced in the paper:

"Self-supervised learning from images with a joint-embedding predictive architecture"
Yann LeCun, Mathilde Caron, Piotr Bojanowski, Armand Joulin, Ishan Misra, et al.
arXiv:2301.08243

Unlike pixel-level reconstruction methods (e.g., MAE), I-JEPA encourages models to reason at a semantic level by predicting high-level representations of masked image regions. This results in more robust and scalable visual representations for downstream tasks.

---

Key Highlights

• Predicts latent feature embeddings, not raw pixels
• Uses block-based masking and Vision Transformers (ViT)
• Dual-network architecture: encoder & predictor
• Flexible mask collator with custom scale/aspect-ratio
• Simple and extensible codebase for research or experimentation

---

Architecture Overview

Reference

Citation:

@article{lecun2023ijepa,
  title={Self-supervised learning from images with a joint-embedding predictive architecture},
  author={LeCun, Yann and Caron, Mathilde and Bojanowski, Piotr and Joulin, Armand and Misra, Ishan and Synnaeve, Gabriel and Zhai, Xiaohua},
  journal={arXiv preprint arXiv:2301.08243},
  year={2023}
}

• Paper on arXiv

---

Acknowledgements

• Core concept and methodology by Meta AI Research.
• Masking and collator logic inspired by the official I-JEPA and DINO repositories.

---

Contact

• GitHub: @aymen-000
• Email: aymne011@gmail.com

---

⭐️ Star this repo if you find it useful!