Dmitrii Krylov

Moonwalk method trains deeper neural networks with less memory

Researchers have developed a new method called Moonwalk that bypasses the need to store intermediate activations during the forward pass of neural network training. This technique, based on inverse-forward differentiati…