Post

"Newton Losses: Using Curvature Information for Learning with Differentiable Algorithms". interesting approach using 2nd order info to avoid the exploding/vanishing gradients that come from non differentiable relaxations.. www.youtube.com/watch?v=7aFP...