Fetching the paper…

Backprop with Approximate Activations for Memory-efficient Network Training · Around