JEPA (Joint-Embedding Predictive Architecture) captures the idea of predicting missing high-level representations rather than reconstructing pixels, aiming for robust, scalable self-supervised learning. A context encoder ingests visible regions and predicts target embeddings for masked regions produced by a separate target encoder, avoiding low-level reconstruction losses that can overfit to texture. This makes learning focus on semantics and structure, yielding features that transfer well with simple linear probes and minimal fine-tuning. The repository provides training recipes, data pipelines, and evaluation utilities for image JEPA variants and often includes ablations that illuminate which masking and architectural choices matter. Because the objective is non-autoregressive and operates in embedding space, JEPA tends to be compute-efficient and stable at scale. The approach has become a strong alternative to contrastive or pixel-reconstruction methods for representation learning.
Features
- Predictive learning in embedding space instead of pixel reconstruction
- Separate context and target encoders with masked-region objectives
- Strong linear-probe and low-shot transfer performance
- Stable, efficient training without heavy negative sampling
- Clear recipes and ablations for masking and architecture choices
- Modular code for extending to new modalities or datasets
Project Samples
