Backprop with Approximate Activations for Memory-efficient Network Training

Training convolutional neural network models is memory intensive since back-propagation requires storing activations of all intermediate layers.