Jin
Research on learning without labels
A joint-embedding predictive architecture does not reconstruct pixels. It predicts the embedding of a piece of an image from the embedding of another piece — so the model spends its capacity on what the image means rather than on what it looks like. Jin is our work on that idea, in the open, with the code and the papers to read.
What is in the repository
Research code, not a product. It trains, it has no tests, and nothing in production depends on it. The repository is archived — read it, fork it, take the parts you want.
I-JEPA
A clean implementation of the image variant: a context encoder, a target encoder that is its moving average, and a predictor that has to guess the target embeddings. Both a ViT and an energy-transformer backbone.
Saccade JEPA
Ours. Eyes jitter constantly and the brain predicts what the jitter will do to the scene, so the model shifts an image, predicts how its embedding moves, and shifts back to check it lands where it started.
Three losses at once
Huber loss on the prediction, a cycle-consistency term that punishes drift on the round trip, and the invariance and covariance parts of VICReg to stop the representation collapsing.
Masked autoencoding
An MAE with self-distillation sits alongside, for the comparison that matters: is a decoder worth its parameters when the goal is a good representation rather than a good reconstruction?
Ways to see what it learned
Self-supervised training has no validation loss to watch, so there are linear probes, KNN, attention-map visualisation in a Dash dashboard, correlation dimension, and UMAP coloured by true class.
The papers
Grant proposals on hierarchical JEPA for edge AI, where the interesting constraint is learning useful representations on hardware that cannot phone home.