Backprop with Approximate Activations for Memory-efficient Network Training