So this is the most important paper in machine learning, and some people say it’s just mimicking previous ideas about learning. I disagree.
JEPA has a rigorous implementation of learning which integrates perception and consciousness into the physical architecture to create explainable superintelligence, without brute-forcing the manifold hypothesis.
Crucially, this means that models trained on LeCun’s architecture will retain their semantic trees as more data is added. Using LeCun’s architecture, new training data can be added without overwriting existing data!