Schmidhuber claims LeCun's JEPA is his 1992 invention

Schmidhuber continues his long-standing crusade for proper attribution in deep learning, this time targeting Meta's heavily promoted self-supervised architecture. He argues that the core concept of non-generative networks predicting latent representations was fully detailed in his 1992 work. He extends this claim to state that modern methods like BYOL and Barlow Twins are also just instantiations of PMAX. The dispute highlights the ongoing tension between early neural network pioneers and the modern architects of the current AI boom over foundational credit.