1) Given an accelerator with no memory to store the previous layer activations, Can you reuse the convolution output of a particular patch of the image for the next convolution operation? If so how, if not why? 2) Is it possible to reduce the multiple convolutions into a single convolution operation to speed up the execution?