Yan LeCun is working on a new type of model with just 15M parameters, trainable on a single GPU in a few hours. Two concepts : - JEPA : "JEPA is a framework for learning world models that predict the dynamic evolution of a system in a compact, low-dimensional latent space".- SIGReg : " the Sketched-Isotropic-Gaussian Regularizer (SIGReg) [25] due to its simplicity, scalability, and stability. SIGReg encourages the latent embeddings to match an isotropic Gaussian target distribution."#AI
Yan LeCun is working on a new type of model with just 15M parameters, trainable on a single GPU in a few hours. Two conc...
