self-supervised-learning

Why LeCun treats energy as the common language of machine learning. This lesson shows how classification, generative models, self-supervised learning, JEPA, and diffusion all read as energy-based models, ties the contrastive-versus-regularized split back to self-supervised learning, and gives an honest account of where explicit EBMs help and where they do not.

How to prevent collapse without negative samples. This lesson covers distillation methods (BYOL, SimSiam) that rely on stop-gradients and asymmetry, regularization methods (Barlow Twins, VICReg) that constrain the embedding's statistics, and the masked-modeling family (MAE, BERT), then ties them to JEPA and the energy-based view.

The first family to make joint-embedding self-supervised learning work: pull together two views of the same image, push apart different images. This lesson covers the InfoNCE loss, SimCLR and its reliance on large batches, MoCo's momentum encoder and queue, why negatives block collapse, and the practical costs that motivated negative-free methods.

Self-supervised learning turns unlabeled data into its own teacher. This lesson covers why labels are the bottleneck, how a pretext task manufactures free supervision, the shift from predicting pixels to learning embeddings invariant to augmentation, and the collapse problem that every method after it must solve.

The JEPA idea becomes concrete in I-JEPA on images and V-JEPA 2 on video. This lesson covers the multi-block masking that makes I-JEPA work, how an action-conditioned video model becomes a world model, and how predicting representations enables planning by model-predictive control instead of a fixed policy.

Predicting a learned representation invites a trivial cheat: represent nothing. This lesson opens the JEPA machine, the context encoder, target encoder, and predictor, shows the training loop in code, explains why representation collapse is the central danger, and how asymmetry, stop-gradient, and an EMA target defeat it without negatives.

Self-supervised learning has three families: reconstruct the input, contrast augmented views, or predict in a learned embedding space. This lesson explains why the third, the joint-embedding predictive approach behind JEPA, sidesteps both wasted capacity on unpredictable pixels and hand-crafted augmentation biases.