Overview
Explore a 40-minute technical video that breaks down the I-JEPA (Image Joint Embedding Predictive Architecture) paper, a collaborative research effort by Meta AI, McGill, Mila, and NYU focusing on non-generative self-supervised learning from images. Learn about semantic image representations, latent space concepts, and the fundamentals of invariance-based pre-training versus generative pre-training approaches. Understand the core mechanics of I-JEPA, its comparison with previous methodologies, and its implementation using Vision Transformer (ViT) architecture. Dive deep into technical aspects including context and target sampling, prediction and loss functions, latent space manipulation, and attention head mechanisms. Examine practical applications through image classification evaluation results, supported by references to related works like Masked Auto Encoder and comprehensive latent space diagrams. Access additional resources including the original paper, community discussions, and dataset implementations through provided links to Oxen.ai platform.
Syllabus
Intro to I-JEPA
Semantic Image Representations
Latent Representation
Invariance Based Pre-Training
Generative Pre-Training
What is I-JEPA
I-JEPA vs. Previous Approaches
ViT Method
Sampling Context and Targets
Prediction and Loss
Latent Space
Attention Head
Evaluation on Image Classification
Conclusion and Conversation
Taught by
Oxen